In 2026, against the backdrop of U.S. export controls on AI chips to China, Cambricon's MLU690 has drawn intense attention as a "China-made H100." This article compares the two in depth across compute, memory, power, software ecosystem, measured performance, and price to help you make a selection decision.

Core Verdict (Read This First)​

Dimension	MLU690	H100	Winner	Gap
BF16 compute	600 TFLOPS	989 TFLOPS	H100	+65%
Memory capacity	64GB HBM3	80GB HBM3	H100	+25%
Memory bandwidth	2 TB/s	3.35 TB/s	H100	+68%
TDP	280W	700W	MLU690	-60%
Energy efficiency	2.14 TFLOPS/W	1.41 TFLOPS/W	MLU690	+52%
Software ecosystem	NeuWare (~75% coverage)	CUDA (100% coverage)	H100	large gap
Price	~¥140,000	~¥200,000	MLU690	-30%
Availability	domestic spot stock	export-controlled	MLU690	✅

One-line summary: MLU690 delivers roughly 60% of H100's compute, but at only 40% of the power and 70% of the price — a strong fit for AI training and inference in the Chinese market.

---

1. Detailed Spec Comparison​

1.1 Compute​

Precision	MLU690	H100 SXM5	H200 SXM5	Note
FP8	~300 TFLOPS (est.)	3,958 TFLOPS	3,958 TFLOPS	H100 supports FP8; MLU690 likely does not
BF16/FP16	600 TFLOPS	989 TFLOPS	989 TFLOPS	H100 leads by 65%
FP32	~150 TFLOPS (est.)	60 TFLOPS	60 TFLOPS	MLU690 estimate; H100 actually higher
INT8	1,200 TOPS	1,979 TOPS	1,979 TOPS	H100 leads by 65%

Key findings:

• ✅ MLU690 reaches 60% of H100's BF16 compute
• ⚠️ H100 supports FP8 (4-bit); MLU690 likely does not (needs confirmation)
• ⚠️ H100's higher INT8 compute favors inference scenarios

1.2 Memory​

Item	MLU690	H100	H200	Note
Capacity	64GB HBM3	80GB HBM3	141GB HBM3e	H200 largest
Bandwidth	2 TB/s	3.35 TB/s	4.8 TB/s	H200 highest
Type	HBM3	HBM3	HBM3e	H200 uses latest HBM3e

Key findings:

• ⚠️ MLU690 has 20% less memory than H100 (64GB vs 80GB)
• ⚠️ MLU690 bandwidth is 40% lower than H100 (2 TB/s vs 3.35 TB/s)
• ❌ When running 70B+ parameter models, MLU690 may run out of memory (model parallelism required)

1.3 Power​

Item	MLU690	H100	H200
TDP	280W	700W	700W
Efficiency (FP16/W)	2.14 TFLOPS/W	1.41 TFLOPS/W	1.41 TFLOPS/W
8-card server power	~3.5kW	~6kW	~6kW
Annual electricity (¥0.6/kWh)	~¥18,400	~¥36,800	~¥36,800

Key findings:

• ✅ MLU690 draws only 40% of H100's power, sharply cutting data-center electricity cost
• ✅ MLU690 leads efficiency by 52%, better suited to large-scale deployment
• ✅ For power-sensitive inference, MLU690 has a clear edge

---

2. Software Ecosystem​

2.1 Framework Support​

Framework	MLU690 (NeuWare)	H100 (CUDA)	Note
PyTorch	✅ (PyTorch-Cambricon)	✅ native	MLU690 needs an extra plugin
TensorFlow	✅ (TensorFlow-Cambricon)	✅ native	same
JAX	⚠️ partial	✅ native	MLU690 limited
ONNX	⚠️ partial	✅ native	same
vLLM	⚠️ in progress	✅ native	MLU690 awaits community port

2.2 Operator Coverage​

Category	MLU690	H100	Note
Basic operators	✅ 95%	✅ 100%	conv, matmul, etc.
Transformer operators	✅ 85%	✅ 100%	Attention, LayerNorm, etc.
Custom operators	⚠️ hand-written	✅ CUDA C++	MLU690 harder to develop
LLM inference opt.	⚠️ basic	✅ mature (FlashAttention, PagedAttention)	H100 leads

Key findings:

• ⚠️ NeuWare is only 5–6 years old, with ~75–85% operator coverage
• ❌ Complex LLMs (e.g., GPT-4, Claude) may need manual optimization
• ✅ Common models (Llama, Qwen, GLM) are essentially already supported

---

3. Measured Performance​

3.1 Training​

Model	MLU690 (time)	H100 (time)	Speedup
Llama 7B	~48 h (est.)	~30 h	1.6x
Llama 70B	~7 days (est.)	~4.5 days	1.6x
Qwen 72B	~8 days (est.)	~5 days	1.6x

Note: above figures are estimates; real performance depends on software optimization.

3.2 Inference​

Model	MLU690 (tok/s)	H100 (tok/s)	Note
Llama 7B	~80 tok/s (est.)	~120 tok/s	H100 +50%
Llama 70B	~20 tok/s (est.)	~35 tok/s	H100 +75%
Qwen 72B	~18 tok/s (est.)	~30 tok/s	H100 +67%

Key findings:

• ⚠️ H100 leads inference by 50–75%
• ✅ But MLU690 draws only 40% the power, with better efficiency
• ✅ For cost-sensitive inference, MLU690 is more economical

---

4. Price​

4.1 Hardware Procurement​

Item	MLU690	H100	H200
Per-card (domestic)	~¥140,000	~¥200,000	~¥300,000
8-card server (turnkey)	~¥1,200,000	~¥1,800,000	~¥2,600,000
Cost gap	-	+50%	+117%

4.2 TCO (3 years)​

Item	MLU690	H100	Note
Hardware	¥1,200,000	¥1,800,000	MLU690 33% cheaper
Electricity (3y)	¥55,200	¥110,400	MLU690 50% cheaper
Facility	¥150,000	¥250,000	MLU690 40% cheaper
TCO (3y)	¥1,405,200	¥2,160,400	MLU690 35% cheaper

Key findings:

• ✅ MLU690's TCO is 35% lower than H100's
• ✅ For large-scale deployment (100+ cards), the cost advantage is pronounced

---

5. Selection Advice​

5.1 Choose MLU690 if...​

• ✅ Your business is primarily in the Chinese market
• ✅ You are affected by U.S. export controls and cannot buy H100/H200
• ✅ You are power-sensitive (edge data centers, high electricity-cost regions)
• ✅ Your models use common architectures (Llama, Qwen, GLM)
• ✅ You have domestic-substitution requirements (government, SOEs, military)

5.2 Choose H100/H200 if...​

• ✅ Your business is global
• ✅ You need to train frontier models (GPT-4 class)
• ✅ Your models use complex operators (need the CUDA ecosystem)
• ✅ You demand extreme performance (low-latency inference)
• ✅ You can legally procure H100/H200

5.3 Hybrid deployment (recommended)​

Scenario	Recommended
Training	H100 (high perf) + MLU690 (low-cost scale-out)
Inference	MLU690 (cost-sensitive) + H100 (low-latency)
Domestic project	all MLU690
International market	all H100/H200

---

6. Outlook​

6.1 MLU690's weaknesses​

• ⚠️ Immature software ecosystem: 75–85% operator coverage; complex models need manual tuning
• ⚠️ Small memory: 64GB limits support for 70B+ parameter models
• ⚠️ Weak interconnect: Cambricon Link bandwidth below NVLink
• ⚠️ Limited international market: affected by U.S. export controls

6.2 MLU690's improvement path​

• 📅 MLU790 (2027): expected 5nm process, ~2x compute
• 📅 Memory upgrade: next gen may adopt HBM3e, capacity up to 128GB
• 📅 Software: NeuWare ecosystem improving, operator coverage target 95%

---

7. Summary​

Dimension	MLU690	H100	Recommended scenario
Compute	⭐⭐⭐⭐	⭐⭐⭐⭐⭐	H100 for top-tier training
Memory	⭐⭐⭐	⭐⭐⭐⭐	H100 for large models
Power	⭐⭐⭐⭐⭐	⭐⭐⭐	MLU690 for inference
Ecosystem	⭐⭐⭐	⭐⭐⭐⭐⭐	H100 for complex models
Price	⭐⭐⭐⭐⭐	⭐⭐⭐	MLU690 for large-scale deployment
Domestic	⭐⭐⭐⭐⭐	❌	MLU690 for Chinese market

Final recommendation:

• 🇨🇳 Chinese market: prefer MLU690 (domestic + low cost)
• 🌍 International market: prefer H100/H200 (performance + ecosystem)
• 💡 Hybrid: train on H100, infer on MLU690

---

References​

• Cambricon MLU690 spec page
• NVIDIA H100 spec page
• Full AI chip comparison table
• TCO calculator

---

Disclaimer: Data in this article is based on public sources and reasonable estimates; actual performance is subject to vendor official testing. MLU690's software ecosystem is evolving rapidly — watch NeuWare updates.

Last updated: 2026-06-23

June 2026 update — the AI compute card market is undergoing its most dramatic reshuffling in years. This article walks through the latest roadmap developments.

---

Key Takeaways​

• Qualcomm AI 200/250 officially enters the datacenter AI inference market, targeting NVIDIA H200
• AMD MI400 series unveils three models: MI430X (HPC), MI440X (enterprise), MI455X (flagship)
• Huawei publishes a three-generation roadmap: 950 (2026) → 960 (2027-Q4) → 970 (2028-Q4)
• Intel Jaguar Shores timeline uncertain, possibly delayed to 2027 or later
• NVIDIA Rubin R200 is in full mass production; the Vera CPU + Rubin GPU combination is now shipping

---

1. Qualcomm: Mobile Giant Moves Into Datacenter AI​

AI 100 → AI 200 → AI 250​

Qualcomm officially launched the AI 200 datacenter inference chip in October 2025, marking the mobile giant's formal entry into the datacenter AI market.

Model	Launch	Availability	Key Features
AI 100	2025-10	2026 H2	Rack-scale AI inference, 768GB LPDDR per card
AI 250	2025-10	2027 H1	Near-memory computing architecture, 10x effective memory bandwidth

Why Qualcomm Can Succeed​

1. Low TCO: LPDDR memory is far cheaper than HBM
2. Energy efficiency: Mobile chip design heritage, excellent power control
3. Inference-focused: Not chasing training performance, focused on inference scenarios
4. Rack form factor: Direct liquid cooling, 160kW rack-level power, Ethernet interconnect

Market Impact​

• Takes on NVIDIA H200: AI 200 inference performance approaches H200 but with 30-40% lower TCO
• Pressures NVIDIA: May push NVIDIA to launch inference-specific chips (e.g., Rubin CPX)
• Diversifies choice: Breaks NVIDIA's monopoly in the inference market

---

2. AMD MI400 Series: Three Models, Precise Positioning​

At CES 2026 (January 2026), AMD officially unveiled the three models of the MI400 series, precisely covering different markets:

MI430X (HPC + Sovereign AI)​

Feature	Spec
Positioning	HPC + sovereign AI
FP32/FP64	Supported (key differentiator)
Use cases	Scientific computing, climate simulation, national AI infrastructure
Competitor	NVIDIA does not make FP64 AI cards

MI440X (Enterprise Servers)​

Feature	Spec
Positioning	Enterprise 8-GPU servers
Compatibility	Works with existing datacenter infrastructure
Use cases	Enterprise AI, private cloud, edge inference
Advantage	Cheaper and easier to deploy than MI455X

MI455X (Flagship AI Training)​

Feature	Spec
Positioning	Flagship AI training + inference
Optimized precision	FP4/FP8/BF16
Helios rack	Core component
Competitor	NVIDIA Rubin R200

Helios Rack-Scale Solution​

AMD also launched the Helios rack-scale AI solution at CES 2026:

• 18 Zen 6 CPUs (2nm process)
• 72 MI455X GPUs
• Direct liquid cooling
• Shipment expected in 2026 H2

---

3. Huawei Three-Generation Roadmap: 950 → 960 → 970​

Huawei unveiled its three-generation chip roadmap at HC 2025 (September 2025) with a very clear timeline:

Ascend 950 Series (2026)​

Model	Launch	Key Features
950PR	2026-Q1	PR (inference-optimized), already in mass production
950DT	2026-Q4	DT (Decode + training), expected to scale up

Technical highlights:

• Added FP8/MXFP8/MXFP4 support
• Interconnect bandwidth 2TB/s (2.5x over 910C)

Ascend 960 (2027-Q4)​

• Doubled compute: All specs double versus the 950 series
• FP8: ~2 PFLOPS expected
• Process: N+3 (equivalent to 5nm)
• Positioning: Targets NVIDIA B200

Ascend 970 (2028-Q4)​

• Third-generation flagship: Only timeline announced, specs TBD
• Significance: Huawei's first complete generation-spanning roadmap
• Signal: China's domestic AI chips have entered a "roadmap-driven" phase

---

4. Intel Jaguar Shores: Timeline Uncertain​

Original Plan​

• Launch: 2026
• Architecture: Xe-HPC + Gaudi fusion
• Process: 18A (Intel's most advanced)
• Memory: Possibly HBM4E (instead of originally planned HBM4)

Latest Developments​

• Possible delay: Some sources suggest a slip to 2027
• Competitors: AMD MI400 already unveiled, NVIDIA Rubin in mass production
• Market pressure: Intel is losing ground in the AI chip market; Jaguar Shores is its last chance

Impact on Roadmap​

If Jaguar Shores slips to 2027, Intel will essentially be out of the AI chip market.

---

5. NVIDIA Rubin Platform: Full Mass Production​

Rubin R200 (2026-Q2 full mass production)​

Feature	Spec
HBM	288GB HBM4
Compute	50 PFLOPS FP4
NVLink	NVLink 6 (1800 GB/s)
Process	TSMC 4NP

Rubin NVL72 Cabinet (2026 H2 shipment)​

• 72 Rubin GPUs
• 36 Vera CPUs
• 1.8 EFLOPS FP4
• Direct liquid cooling

Vera CPU (Debut)​

• Architecture: Custom CPU replacing Grace
• Positioning: Deep co-design with Rubin GPU
• Significance: NVIDIA's transformation from a GPU company into a computing platform company

---

6. Google TPU v8: Training/Inference Officially Split​

TPU 8t (training) + TPU 8i (inference)​

At Cloud Next 2026, Google announced TPU v8 would officially split into training and inference versions:

Feature	TPU 8t (training)	TPU 8i (inference)
Optimization	High compute, high bandwidth	Low latency, low cost
Interconnect	Optical interconnect	Ethernet
Launch	2027	2027

Significance​

• Industry trend: Specialization of training/inference chips
• Followers: Qualcomm AI 200 is also inference-only
• NVIDIA pressure: Does it need an inference-specific chip?

---

7. Cerebras WSE-4: Wafer-Scale Engine Evolves​

Core Specs​

Feature	Spec
Transistors	1.4 trillion
Compute	125 PFLOPS FP8
Launch	2026 H2
Process	TSMC 5nm

Competitive Advantages​

• Massive model training: A single WSE-4 can train 10T+ parameter models
• Low-latency inference: Entire model on one chip, no communication overhead
• Mature software stack: Cerebras stack already supports PyTorch, TensorFlow

---

8. Market Landscape Analysis​

Training Market​

Rank	Vendor	Product	Market Share (est.)
1	NVIDIA	Rubin R200	70%
2	AMD	MI455X	15%
3	Google	TPU v8t	10%
4	Huawei	Ascend 960	5% (mostly China)

Inference Market (New Battlefield)​

Rank	Vendor	Product	Advantage
1	NVIDIA	H200 / Rubin CPX	Mature ecosystem
2	Qualcomm	AI 200	Low TCO
3	AMD	MI440X	Good compatibility
4	Intel	Gaudi 4	Low price

---

9. Key Trends​

Trend 1: Rise of Inference-Specific Chips​

• Qualcomm AI 200: Mobile giant enters the market
• NVIDIA Rubin CPX: NVIDIA's first inference-specific chip
• Google TPU 8i: Training/inference officially split

Trend 2: Rack-Scale Solutions Become Standard​

• NVIDIA NVL72: 72 GPU + 36 CPU
• AMD Helios: 18 CPU + 72 GPU
• Qualcomm rack: 160kW liquid-cooled rack

Trend 3: China's Domestic Chips Enter "Roadmap-Driven" Phase​

• Huawei three-generation roadmap: 950 → 960 → 970
• Clear timeline: 2026-Q1 → 2027-Q4 → 2028-Q4
• Significance: From "catch-up" to "planning"

Trend 4: HBM Capacity Becomes the Bottleneck​

• SK hynix: HBM4 capacity already booked by NVIDIA
• Samsung: HBM4E samples delivered to AMD
• Impact: MI400 and Rubin R200 shipments constrained by HBM capacity

---

10. Procurement Recommendations​

If Procuring in 2026 H2​

1. Training scenarios:

   • First choice: NVIDIA Rubin R200 (best performance)
   • Alternative: AMD MI455X (better price/performance)
   • Domestic: Huawei Ascend 950DT (China-based customers)

2. Inference scenarios:

   • First choice: NVIDIA H200 (mature ecosystem)
   • Best value: Qualcomm AI 200 (if available)
   • Cost-sensitive: AMD MI440X

3. HPC scenarios:

   • Only choice: AMD MI430X (FP64 support)

If Procuring in 2027​

• Wait for Rubin Ultra: Performance possibly 2x R200
• Watch MI500: AMD's next-generation product
• Evaluate TPU v8: If already on Google Cloud

---

Conclusion​

2026 H2 will be the most fiercely contested half-year in AI chip market history:

• NVIDIA continues to lead, but its advantage is narrowing
• AMD precisely positions three models; market share will keep rising
• Qualcomm enters the inference market; its low-TCO strategy may disrupt the market
• Huawei has a clear three-generation roadmap; domestic substitution accelerates
• Intel's Jaguar Shores is make-or-break

For procurement decision-makers, this is the hardest time to decide — every option has clear pros and cons.

For engineers, this is the best of times — chip performance doubles yearly, architectural innovation is endless.

---

References​

• AI Compute Card Future Roadmap - MirrorFrog real-time updates
• NVIDIA Rubin R200 deep dive (see related articles on this site)
• AMD MI400 series CES 2026 launch (see related articles on this site)
• Qualcomm AI 100 launch analysis (coming soon)

---

Last updated: 2026-06-20
Author: Charles Qing
Tags: #roadmap #market-analysis #procurement

On January 5, 2026, on the opening day of CES 2026 (Consumer Electronics Show), AMD Chair and CEO Dr. Lisa Su unveiled in her keynote: the Instinct MI400 series AI accelerators.

The most eye-catching is MI455X — AMD's most powerful AI accelerator ever, using a 2nm + 3nm hybrid process, 432GB HBM4, with FP4 compute up to 40 PFLOPS (20 PFLOPS FP8).

Key highlights​

• MI455X: FP4 40 PFLOPS, FP8 20 PFLOPS, 10× over MI355X
• MI450: cost-performance version, FP4 28 PFLOPS, 288GB HBM4
• Process upgrade: world's first AI chip with 2nm + 3nm hybrid process (GCD on 2nm, MCD on 3nm)
• Memory upgrade: from MI350X's 288GB HBM3e to 432GB HBM4 (MI455X)
• Bandwidth upgrade: from MI350X's 8 TB/s to 19.6 TB/s (2.45×)
• Architecture upgrade: from CDNA 4 to CDNA 5
• Mass production: MI455X Q4 2026, MI450 Q3 2026

Full MI400 series specs​

📌 Important correction (2026-06-16): After official spec verification, MI455X memory is 432GB HBM4 (not the earlier reported 288GB), and FP4 compute is 40 PFLOPS. Corrected herein.

Model	Positioning	Memory	FP4 compute	FP8 compute	TDP (est.)
MI455X	Flagship training+inference	432GB HBM4	40 PFLOPS	20 PFLOPS	~1,000W
MI450	Cost-performance training	288GB HBM4	28 PFLOPS	14 PFLOPS	~800W
MI440X	Enterprise inference	216GB HBM4	25 PFLOPS	12.5 PFLOPS	~600W
MI430X	HPC / scientific computing	192GB HBM4	20 PFLOPS	10 PFLOPS	~500W
MI400X	General / edge inference	128GB HBM4	12 PFLOPS	6 PFLOPS	~400W

Key upgrades (vs MI350 series):

• Memory: HBM3e → HBM4, capacity +50% (432GB vs 288GB)
• Bandwidth: 19.6 TB/s (vs MI350's 8 TB/s, +2.45×)
• Compute: FP4 40 PFLOPS (vs MI355X's 20 PFLOPS, +2×)
• Process: 2nm + 3nm hybrid (GCD on 2nm, MCD on 3nm)
• Architecture: CDNA 5 (vs MI350's CDNA 4)

Performance vs. MI355X​

Metric	MI355X (2025)	MI455X (2026)	Improvement
FP4 compute	20 PFLOPS	40 PFLOPS	2×
FP8 compute	10 PFLOPS	20 PFLOPS	2×
Memory capacity	288GB HBM3e	432GB HBM4	1.5×
Memory bandwidth	8 TB/s	19.6 TB/s	2.45×
Process	TSMC 3nm	2nm + 3nm hybrid	New gen
Architecture	CDNA 4	CDNA 5	New gen
TDP	800-1000W	~1,000W	Flat

Lisa Su at CES 2026:

"Four years ago, MI250's AI performance was X. Today, MI455X's performance is 1000× that. That's the pace of AI chip progress."

CDNA 5 architecture in detail​

The MI400 series adopts the CDNA 5 architecture (MI355X uses CDNA 4):

Key upgrades​

1. Matrix Core upgrade: FP8/INT8/FP16 support, sparsity acceleration
2. HBM4 controller: 12-layer HBM4 (vs HBM3e's 8 layers)
3. Infinity Fabric 4.0: 50% higher die-to-die / die-to-GPU bandwidth
4. Native sparsity support: MoE Expert-Parallel optimization
5. Long-context optimization: 1M+ token KV Cache acceleration

vs. NVIDIA Blackwell / Rubin​

Metric	AMD MI455X	NVIDIA B200	NVIDIA Rubin R200 (2026 Q4)
FP4 compute	40 PFLOPS	20 PFLOPS (45 sparse)	~40 PFLOPS (est.)
FP8 compute	20 PFLOPS	10 PFLOPS (22.5 sparse)	~20 PFLOPS (est.)
Memory	432GB HBM4	192GB HBM3e	288GB HBM4
Memory bandwidth	19.6 TB/s	8 TB/s	13 TB/s
TDP	~1,000W	700-1000W	~1,000W
Process	2nm + 3nm hybrid	TSMC 4np	TSMC 3nm
Mass production	2026 Q4	2024 Q4	2026 Q4
Software ecosystem	ROCm	CUDA	CUDA
Strength	Memory capacity, open ecosystem	Most mature ecosystem	Next-gen architecture
Weakness	Software ecosystem gap	Smaller memory	Not yet launched

Conclusion: MI455X leads B200 in FP4/FP8 compute and memory capacity/bandwidth, but software ecosystem remains a weak point. Versus Rubin R200, paper specs are close, but Rubin has the CUDA ecosystem moat.

Production timeline​

Time	Event
June 12, 2025	MI400 series specs first announced at Advancing AI
January 5, 2026	MI455X/MI450/MI440X formally launched at CES 2026
2026 Q3	MI450 sampling begins
2026 Q4	MI455X mass production
2026 Q4	MI440X (enterprise inference) launched
2027 Q1	MI430X/MI400X (HPC/edge inference) launched
2027	MI500 series (next gen)

AMD AI chip roadmap (2025-2027)​

Time	Product	Process	Notes
Q4 2024	MI325X	TSMC 5nm	HBM3e upgraded
Q3 2025	MI355X (MI350 series)	TSMC 3nm	CDNA 4, 288GB HBM3e
Q4 2026	MI455X (MI400 series)	2nm + 3nm hybrid	CDNA 5, 432GB HBM4
Q1 2027	MI500 series	TSMC 2nm (est.)	Next gen, further gains

Software ecosystem: ROCm's progress and challenges​

✅ Progress​

• PyTorch 2.5+: native MI300X/MI455X support
• Hugging Face Transformers: official AMD GPU support
• vLLM 0.8+: MI300X inference support (experimental)
• JAX: AMD adapting (vs Google TPU)

⚠️ Challenges​

• Framework optimization: PyTorch on AMD GPUs still below NVIDIA
• Operator coverage: some niche operators need hand-written HIP
• Multi-card communication: RCCL (vs NCCL) still lags
• Developer ecosystem: tutorials, cases, community activity far below NVIDIA

Competitive comparison​

Vendor	Product	FP4 compute	Memory	Mass production	Strength	Weakness
AMD	MI455X	40 PFLOPS	432GB HBM4	2026 Q4	Largest memory, open ecosystem	Software gap
NVIDIA	B200	20 PFLOPS	192GB HBM3e	2024 Q4	Most mature ecosystem	Smaller memory
NVIDIA	Rubin R200	~40 PFLOPS	288GB HBM4	2026 Q4	Next-gen architecture, CUDA	Expensive
Huawei	Ascend 910C	~1.6 PFLOPS	64GB HBM	2026 Q2	China-localized	Export-controlled
Google	TPU 8t	~9.2 PFLOPS	~256GB HBM3e	Late 2027	Gemini-integrated	Google Cloud only

Industry impact​

1. Impact on NVIDIA​

On paper, AMD MI455X has already caught up to B200 (FP4 40 PFLOPS vs 20 PFLOPS), even leading substantially in memory capacity (432GB vs 192GB).

But:

• NVIDIA has the CUDA ecosystem moat
• NVIDIA has the Vera Rubin platform (full solution, 2026 Q4)
• AMD only sells cards/nodes, NVIDIA sells AI factories
• MI455X mass production (2026 Q4) coincides with Rubin R200 — head-on competition

2. Pressure on domestic chips​

MI455X's launch means: mainstream international AI chips enter the 2nm + HBM4 era in 2026.

Domestic chips (Huawei Ascend, Cambricon, MetaX, etc.) need to:

• Catch up to 5nm + HBM3e by 2026-2027
• Otherwise the gap widens from "1 generation" to "2 generations"

3. Significance for cloud providers​

MI455X gives cloud providers a second option beyond NVIDIA:

• Microsoft Azure: already deployed MI355X, may follow with MI455X
• Google Cloud: in-house TPU, won't use AMD
• Amazon AWS: in-house Trainium/Inferentia, won't use AMD
• Alibaba Cloud, Tencent Cloud: may procure MI455X as NVIDIA alternative

Related chips​

• AMD MI455X - the new launch
• AMD MI300X - previous flagship
• AMD MI325X - HBM3e upgraded
• NVIDIA B200 - direct competitor
• NVIDIA Rubin R200 - same-generation competitor

References​

• MI455X chip - Baidu Baike - official specs
• AMD drops its most powerful two-gen AI chips! World's first 432GB HBM4 - Zhihu - on-site report
• Taking on NVIDIA: AMD launches MI440X for enterprise AI data centers - Tencent News
• AMD MI400 packs 432GB HBM4: specs crush NVIDIA on paper - CocoLoop
• AMD CES 2026 keynote replay

---

This article is compiled from AMD CES 2026 official announcements, Baidu Baike, and Zhihu on-site reports; specs verified against official sources. Updated 2026-06-16: corrected MI455X memory (288GB → 432GB) and compute (FP8 6 PFLOPS → FP4 40 PFLOPS).

On April 22, 2026, at Google Cloud Next '26 in Las Vegas, Google officially launched its 8th-generation Tensor Processing Unit (TPU). For the first time in Google's history, it split AI training and inference onto two independent chips:

• TPU 8t: designed for model training
• TPU 8i: focused on high-concurrency inference

This launch introduces no new physical concept, but focuses on solving the core pain points of AI data centers: ten-thousand-card cluster scaling efficiency, Agentic AI workload optimization, and performance per watt.

TPU 8i (inference-specific): eliminating the "waiting room effect"​

TPU 8i is the first inference-specific chip co-designed by Google and MediaTek, aimed at eliminating the "waiting room effect" — where user requests are intentionally queued or delayed to maximize hardware utilization.

TPU 8i core specs (estimated)​

Parameter	TPU 8i	TPU v7 Ironwood
Positioning	Inference-specific	Mostly inference
Process	TSMC 2nm	—
Die design	Dual compute die (est.)	—
Memory	8× HBM3e 12-layer (~192GB est.)	8× HBM3 (192GB)
Memory bandwidth	~7 TB/s (est.)	7,380 GB/s
FP8 compute	~4,614 TFLOPS (est.)	4,614 TFLOPS
TDP (per chip)	1,300 W	1,000 W
Interconnect	ICI 3D Torus	ICI 3D Torus
Integrated CPU	Arm Axion (64 cores)	None
Cooling	Air or liquid	4th-gen liquid
Announced	2026-04-22	2025-08-25
Mass production	End of 2027	2026

Key features:

• ✅ High-concurrency inference optimization: built for Agentic AI, supports inference chains of thousands of steps
• ✅ Arm Axion CPU integration: 64-core Neoverse V2, host CPU + data preprocessing synergy
• ✅ Low latency: eliminates the "waiting room effect", extremely low TTFT (time to first token)
• ✅ 117% better performance per watt: vs Ironwood (at equal price)

TPU 8t (training-specific): the "engine" of Gemini 3/4​

TPU 8t is designed for training Google's frontier models like Gemini 3 / Gemini 4, continuing Google's long-term partnership with Broadcom.

TPU 8t core specs​

Parameter	TPU 8t	TPU v7 Ironwood	Improvement
Positioning	Training-specific	Mostly inference	Form-factor split
Process	TSMC 2nm	—	New gen
Die design	Dual compute die	—	Architecture upgrade
Memory	8× HBM3e 12-layer (~256GB per chip est.)	8× HBM3 (192GB)	Upgrade
Memory bandwidth	~7 TB/s (per chip est.)	7,380 GB/s	Flat
Pod chip count	9,600 chips	9,216	+4%
Pod total HBM	2 PB	—	Far exceeds
Pod FP4 compute	121 EFLOPS	~42 EFLOPS (est.)	~3×
Integrated CPU	Arm Axion (64 cores)	None	New
TDP (per chip)	1,300 W	1,000 W	+30%
Mass production	End of 2027	2026	—

Key features:

• ✅ Native MoE training support: Expert-Parallel optimization (DeepSeek / Mixtral style)
• ✅ Long-context training: 1M+ token context training optimization
• ✅ RLHF / post-training: native Online RL (DPO/PPO/GRPO) optimization
• ✅ Arm Axion CPU synergy: data preprocessing / weight init offloaded to CPU
• ✅ SparseCore acceleration: MoE routing and recommendation systems

Strategic significance of the 8th-generation TPU​

1. Training and inference split for the first time​

Previously, Google's TPU design philosophy was "one architecture for both training and inference" (e.g., TPU v5p, v6e). But the arrival of the Agentic AI era changed that:

• Training workloads: large-scale matrix multiply, long-sequence backpropagation, sparse MoE
• Inference workloads: high concurrency, low latency, KV Cache-intensive, dynamic batching

These two workloads impose very different demands on chip architecture. After the split:

• TPU 8t can focus on optimizing compute density and memory capacity
• TPU 8i can focus on optimizing inference throughput and performance per watt

2. Dual-track partnership with Broadcom and MediaTek​

• Broadcom: continues designing TPU 8t (training), extending the long-term partnership since TPU v1
• MediaTek: first-time collaboration designing TPU 8i (inference), bringing mobile-chip low-power design expertise

This "dual-track" strategy lets Google:

• Pursue peak performance on training chips (combined with Broadcom's high-end ASIC experience)
• Pursue peak energy efficiency on inference chips (combined with MediaTek's mobile-chip experience)

3. Versus NVIDIA Vera Rubin​

Comparison	Google TPU 8t + 8i	NVIDIA Vera Rubin
Strategy	Training/inference split	Unified architecture (GPU+CPU)
Process	TSMC 2nm	TSMC 3nm (est.)
Ecosystem	Google Cloud only	Globally available
Software	JAX / PyTorch-XLA	CUDA / PyTorch
Mass production	End of 2027	Fall 2026
Strength	Deep Gemini integration	Most mature ecosystem

Deep technical analysis​

TSMC 2nm: why 2nm?​

Google is the first vendor to adopt TSMC 2nm on an AI accelerator (NVIDIA Rubin uses 3nm). 2nm (N2) vs 3nm (N3E):

• Transistor density: ~15-20% higher
• Power reduction: ~25-30% (at equal performance)
• Performance gain: ~10-15% (at equal power)

For TPU 8t/8i, which already hit 1,300W, 2nm is mandatory — otherwise 4nm/3nm couldn't integrate dual compute dies and 8× HBM3e within reasonable power.

Arm Axion CPU: Google's in-house CPU enters the TPU node for the first time​

Previously, TPU nodes used Intel Xeon or AMD EPYC as host CPUs. TPU 8t/8i integrate Google's in-house Arm Axion CPU (64-core Neoverse V2) for the first time:

Significance:

1. Data preprocessing offload: tokenization, data augmentation can run entirely on Axion, freeing TPU compute
2. Weight initialization: large-model training weight init on CPU, accelerating training startup
3. Inference scheduling: Axion handles request scheduling and load balancing for multi-model inference

This marks the TPU node's evolution toward a "SuperNode": no longer a pure accelerator, but a TPU + Axion CPU co-design system, comparable to NVIDIA's Vera CPU.

4th-gen liquid cooling: the 1,300W thermal challenge​

TPU 8t/8i TDP reaches 1,300W (30% over Ironwood's 1,000W), posing a huge data-center cooling challenge.

Google adopts a 4th-gen liquid cooling solution:

• Cold-plate liquid cooling: directly cools GPU die and HBM
• Immersion cooling: optional (ultra-high-density deployment)
• Smart thermal control: dynamically adjusts pump speed and fan RPM by workload

Production timeline and use cases​

Time	Event
2026-04-22	Cloud Next '26 official announcement
H2 2026	Internal testing (Google DeepMind first)
End of 2027	Mass production, Google Cloud availability
2028	Next-gen TPU (possibly TPU 9)

Target use cases:

• ✅ Frontier model training (Gemini 3/4, external customers)
• ✅ MoE large-model inference (high concurrency, low latency)
• ✅ Multimodal AI (ViT + LLM simultaneous inference)
• ✅ Agentic AI (Agentic AI workloads)

Competitive comparison​

Vendor	Product	Process	TDP	Mass production
Google	TPU 8i (inference)	TSMC 2nm	1,300W	End of 2027
Google	TPU 8t (training)	TSMC 2nm	1,300W	End of 2027
NVIDIA	Rubin GPU	TSMC 3nm (est.)	~1,000W	Fall 2026
NVIDIA	Vera CPU	TSMC 3nm (est.)	~500W	Fall 2026
AMD	MI455X (MI400)	TSMC 3nm (est.)	~700W	2026
Huawei	Ascend 950PR	—	~500W	Q1 2026

Industry impact​

1. AI chips enter the 2nm era: Google leads with TSMC 2nm; NVIDIA and AMD will follow
2. Training/inference split becomes a new trend: other vendors (NVIDIA, AMD) may follow suit
3. In-house CPUs become standard: Google (Axion), NVIDIA (Vera), Huawei (Kunpeng) all do CPU+accelerator co-design
4. Liquid cooling becomes inevitable: 1,300W TDP means air cooling can no longer suffice

Related chips​

• Google TPU 8i - inference-specific TPU detailed specs
• Google TPU 8t - training-specific TPU detailed specs
• Google TPU v7 Ironwood - previous-gen inference TPU
• NVIDIA Rubin R200 - same-generation competitor
• NVIDIA Vera CPU - counterpart to Google Axion

References​

• Google blog: Introducing new TPUs at Cloud Next '26
• Google Cloud Next '26 keynote replay
• Tencent News: Google 8th-gen TPU deep dive

---

This article is compiled from Google's official announcements and public sources; some specs are estimates, subject to final official release.

On June 5, 2026, Shenzhen announced a major piece of news: Shenzhen Hetao College, together with HIT (Shenzhen) and Huawei, used 1,000 Huawei Ascend 910C chips to successfully complete full-parameter post-training of the 1.6-trillion-parameter DeepSeek-V4-Pro large model.

This was no tentative attempt, but a milestone technological breakthrough. It proved with irrefutable engineering results that: domestic AI chips are fully capable of supporting world-class, super-large-parameter model training.

Why this matters​

The two thresholds of AI chips: "inference" and "training"​

• Inference: using an existing model to chat, write copy. Domestic chips could already do this
• Training: adjusting model parameters to learn new capabilities. Full-parameter training adjusts all 1.6 trillion parameters at once — maximum difficulty

Previously, full-parameter training of trillion-scale models was monopolized by NVIDIA H100/H200. Domestic chips could only do inference, not large-scale training.

The significance of this breakthrough: domestic compute leapt from "usable" to "useful", from "inference" to "training".

Technical details​

Training configuration​

Item	Parameter
Chips	Huawei Ascend 910C × 1,000
Model	DeepSeek-V4-Pro
Parameters	1.6 trillion (1600B)
Training type	Full-parameter post-training
Framework	MindSpore + torch_npu
Completed	Announced June 5, 2026

Performance metrics​

Metric	Value	Assessment
Compute utilization	>30%	Industrial grade (top overseas chips ~40%)
Key training operator efficiency	+14%	vs previous-gen 910B
Communication bandwidth utilization	>60% (est.)	MoE All-to-All communication
Stability	1,000 cards trained continuously with no failures	Cluster stability met standard

💡 About 30% compute utilization: many feel 30% is low, but in large-model training this is already a very respectable industrial-grade level. Even with the most advanced overseas chips, many teams' actual utilization is around 40%.

Ascend 910C detailed specs​

Ascend 910C is Huawei's AI training/inference chip announced at the Huawei Analyst Conference (April 24, 2024), with a theoretical peak of 800 TFLOPS (BF16), in the same class as NVIDIA H100.

Parameter	Ascend 910C	Ascend 910B	NVIDIA H100
Architecture	Ascend 910C	Ascend 910B	Hopper
Process	TSMC 7nm (est.)	TSMC 7nm	TSMC 4NP
BF16 compute	800 TFLOPS	256 TFLOPS	989 TFLOPS (sparse)
Memory	64GB HBM (est.)	64GB HBM2e (B1/B2)	80GB HBM3
Memory bandwidth	~2TB/s (est.)	600 GB/s (B1/B2)	3.35 TB/s
TDP	~400W (est.)	300-400W	700W
Mass production	April 2026 (full production)	Nov 2022	Mar 2022

Key upgrades:

• ✅ 3× compute: from 910B's 256 TFLOPS to 800 TFLOPS
• ✅ Mature software ecosystem: torch_npu adapts PyTorch, MindSpore framework mature
• ✅ Cluster stability: 1,000 cards trained continuously with no failures (the biggest breakthrough)

Technical challenges and solutions​

Challenge 1: Memory demand of trillion-scale models​

A 1.6-trillion-parameter model needs, just for model parameters:

• FP16: 1.6T × 2 bytes = 3.2 TB
• Plus gradients and optimizer states: at least 10 TB of memory

Huawei's solution:

• Model Parallel: distribute the model across 1,000 910C chips
• ZeRO optimizer: optimize memory footprint
• Gradient accumulation: update parameters in stages

Challenge 2: Communication efficiency of thousand-card clusters​

Training with 1,000 chips, inter-chip communication becomes the bottleneck. MoE models need All-to-All communication (each expert may need to communicate with all others).

Huawei's solution:

• HCCS (Huawei Collective Communication Scheduler): in-house high-speed interconnect protocol
• Layered communication: intra-node NVLink + inter-node HCCS
• Communication-compute overlap: data transfer concurrent with computation

Challenge 3: Training stability​

Trillion-scale model training takes weeks or months; any single card failure can interrupt the entire training.

Huawei's solution:

• Fault detection and auto-recovery: real-time monitoring of card status, auto-restart and recovery on failure
• Checkpoint optimization: high-frequency training-state saves (every N steps)
• Ascend cluster management software: designed specifically for enterprise training

Competitive comparison​

Vendor	Chip	1.6T-param training	Ecosystem maturity	Availability
Huawei	Ascend 910C	✅ Completed	⭐⭐⭐ (improving)	China-localized
NVIDIA	H100/H200	✅ Industry standard	⭐⭐⭐⭐⭐	Global (export-controlled)
AMD	MI300X	✅ Feasible	⭐⭐⭐⭐	Global
Google	TPU v5p/8t	✅ JAX-native	⭐⭐⭐⭐	Google Cloud

Conclusion: Ascend 910C has caught up to H100 in hardware performance, still lags in software ecosystem, but this training success proves engineering feasibility.

Industry impact​

1. The "Zunyi Conference" of domestic compute​

This breakthrough is called the "Zunyi Conference" of domestic compute — from passive defense to strategic counteroffensive.

Specific impact:

• ✅ Breaks the bias that "domestic chips can only do inference"
• ✅ Proves domestic chips can train frontier models
• ✅ Provides compute foundation for domestic large models (e.g., DeepSeek-V4, ERNIE 5.0)

2. Impact on NVIDIA​

Huawei Ascend 910C completing trillion-scale training means China's AI industry is less dependent on NVIDIA.

Scenario	Before	Now
Inference	Domestic chips usable	Domestic chips useful
Training	Must use H100/H200	Can use 910C
Large-scale training	Must use H100 clusters	Can use 910C clusters

3. Boost to the domestic chip industry​

This breakthrough will drive the entire domestic AI chip supply chain:

• Chip design: Cambricon, MetaX, Moore Threads accelerate iteration
• Wafer manufacturing: SMIC, Hua Hong get more orders
• Packaging/test: JCET, TFME benefit

Huawei Ascend roadmap (2025-2028)​

Time	Chip	Positioning
Q1 2025	Ascend 910C	Flagship training/inference (mass-produced)
Q1 2026	Ascend 950PR	Inference-optimized (~500 TFLOPS BF16)
Q4 2026	Ascend 950DT	Data-center training
Q4 2027	Ascend 960	Next-gen flagship
Q4 2028	Ascend 970	Next-next-gen

Training lessons shared​

The Shenzhen Hetao College team accumulated valuable experience:

✅ Successes​

1. Progressive training: start from small models (7B), gradually scale to 1.6T
2. Mixed-precision training: BF16 main + FP32 gradient accumulation
3. Communication optimization: All-to-All overlap with computation
4. Fault recovery: save checkpoint every 1,000 steps

⚠️ Challenges encountered​

1. Memory fragmentation: severe fragmentation over long training, needs periodic cleanup
2. Communication bottleneck: MoE All-to-All takes 30%+ of training time
3. Software bugs: torch_npu occasional memory leak, needs training process restart

Related chips​

• Huawei Ascend 910C - the chip used in this training
• Huawei Ascend 910B - previous-gen chip
• Huawei Ascend 950PR - next-gen inference chip
• NVIDIA H100 - benchmark product
• NVIDIA H200 - benchmark product (HBM3e upgrade)

References​

• Sina Finance: Huawei Ascend 910C completes DS-V4-Pro training
• Tencent News: Milestone! Ascend 910C completes 1.6-trillion-param model training
• 21IC: Huawei chip completes full-parameter post-training of 1.6T AI model

---

This article is compiled from public reports. Salute to the teams at Shenzhen Hetao College, HIT (Shenzhen), and Huawei — you proved the feasibility of China's AI compute with engineering results.

On March 18, 2026, Intel officially launched at the Intel AI Summit: the Habana Gaudi 4 custom AI accelerator. This is Intel's latest-gen AI training/inference chip after Gaudi 3 (launched April 2024), designed for large-scale model training.

Meanwhile, Intel confirmed its next-gen Jaguar Shores GPU (datacenter GPU) is in development, will use HBM4 memory, and is expected in 2027. This marks Intel's formal return to the AI chip race.

Key Highlights​

• Gaudi 4: Launched March 2026, TSMC 5nm, 64GB HBM3e, for large-scale training
• Jaguar Shores: Launches 2027 (est.), HBM4, targeting NVIDIA Rubin
• Crescent Island: Intel's first general-purpose GPU (launched 2026), Xe3 architecture
• Software ecosystem: Intel AI Stack (including oneAPI, BigDL, Gaudi Software Suite)
• Foundry partners: TSMC (Gaudi 4, Jaguar Shores), Intel Foundry (Crescent Island)

Gaudi 4 Detailed Specs​

Gaudi 4 is the fourth-gen AI accelerator designed by Intel's Habana Labs (acquired 2019).

Parameter	Gaudi 4	Gaudi 3 (2024)	NVIDIA B200
Architecture	Habana 4	Habana 3	Blackwell
Process	TSMC 5nm	TSMC 7nm	TSMC 4NP
FP8 compute	~2,000 TFLOPS (est.)	1,000 TFLOPS	4,500 TFLOPS (sparse)
Memory	64GB HBM3e	128GB HBM2e (est.)	192GB HBM3e
Memory bandwidth	~3 TB/s (est.)	~2 TB/s (est.)	8 TB/s
TDP	~500W (est.)	~400W	700-1000W
Interconnect	RoCE v3 (Ethernet)	RoCE v2	NVLink 5.0
Launch	March 2026	April 2024	March 2024
Mass production	2026 Q3 (est.)	Q4 2024	Q4 2024

📌 Note: Gaudi 4 exact specs not fully public; some values above are estimates.

Gaudi 4 Key Features​

1. Native Ethernet support: Uses RoCE v3 (RDMA over Converged Ethernet), no dedicated interconnect protocol needed (like NVLink)
2. Large-scale scaling optimized: Ten-thousand-card cluster scaling efficiency better than InfiniBand (lower cost)
3. Sparsity acceleration: Native MoE model support
4. Multi-precision support: FP8/FP16/FP32/INT8/INT4
5. Open ecosystem: Supports PyTorch, TensorFlow, JAX (via third-party adaptation)

Jaguar Shores: Intel's Next-Gen GPU​

Jaguar Shores is Intel's first true datacenter GPU (not an ASIC like Gaudi).

Why "Jaguar Shores"?​

• Jaguar: Symbolizes "speed" and "agility"
• Shores: Symbolizes "openness" and "connection"

Jaguar Shores Estimated Specs​

Parameter	Jaguar Shores (est.)	NVIDIA Rubin	AMD MI455X
Architecture	Xeu 3 (est.)	Rubin	CDNA 4
Process	TSMC 3nm (est.)	TSMC 3nm	TSMC 3nm
Memory	HBM4 (confirmed)	HBM4	HBM4
Memory capacity	288GB (est.)	288GB	288GB
FP8 compute	~4,000 TFLOPS (est.)	~6,000 TFLOPS	6,000 TFLOPS
TDP	~800W (est.)	~1,000W	~800W
Launch	2027 (est.)	2026 Q3	2026 Q3

Key confirmations:

• ✅ HBM4 memory: Intel confirmed Jaguar Shores will use SK hynix HBM4
• ✅ TSMC foundry: Jaguar Shores will be produced by TSMC (not Intel Foundry)
• ✅ oneAPI native support: Jaguar Shores will natively support the oneAPI programming model

Crescent Island: Intel's First General-Purpose GPU​

Crescent Island is Intel's first general-purpose datacenter GPU announced October 2025, using the Xe3 architecture (upgrade of Xe-HPG).

Parameter	Crescent Island (est.)	Intel Data Center GPU Max	NVIDIA L40S
Architecture	Xeu 3	Xeu 2 (Ponte Vecchio)	Ada Lovelace
Positioning	General compute + AI inference	HPC + AI training	AI inference + graphics
Process	TSMC 5nm (est.)	Intel 7 + TSMC 5nm	TSMC 4N
Memory	48GB HBM3 (est.)	128GB HBM2e	48GB GDDR6
TDP	~300W (est.)	600W	350W
Launch	2026 (est.)	Jan 2023	Mar 2023

Positioning:

• ✅ General-purpose GPU: Both AI inference and scientific computing (HPC)
• ✅ Low cost: Cheaper than Gaudi 4, targeting NVIDIA L40S
• ✅ Open standards: Supports oneAPI, SYCL, Level Zero

Intel AI Chip Roadmap (2024-2027)​

Time	Product	Type	Process	Note
2024 Q4	Gaudi 3	AI ASIC	TSMC 7nm	Current mainstay
2026 Q2	Crescent Island	General GPU	TSMC 5nm	New launch
2026 Q3	Gaudi 4	AI ASIC	TSMC 5nm	New launch
2027	Jaguar Shores	Datacenter GPU	TSMC 3nm	Next-gen flagship
2027	Gaudi 5 (est.)	AI ASIC	TSMC 3nm	Next-gen

vs Competitors​

Gaudi 4 vs NVIDIA B200​

Metric	Gaudi 4	NVIDIA B200
FP8 compute	~2,000 TFLOPS	4,500 TFLOPS
Memory	64GB HBM3e	192GB HBM3e
Interconnect	Ethernet (RoCE v3)	NVLink 5.0
Software ecosystem	Gaudi Software Suite	CUDA
Price	est. ~$20,000	~$45,000
Advantage	Low Ethernet cost, open	Most mature ecosystem, strongest performance
Disadvantage	Weak software ecosystem, lower compute	Expensive

Conclusion: Gaudi 4 is positioned as a "cost-effective training solution," suited for cost-sensitive customers willing to invest in software adaptation.

Jaguar Shores vs NVIDIA Rubin​

Metric	Jaguar Shores (est.)	NVIDIA Rubin
FP8 compute	~4,000 TFLOPS	~6,000 TFLOPS
Memory	288GB HBM4	288GB HBM4
Software ecosystem	oneAPI	CUDA
Mass production	2027	2026 Q3
Advantage	Open standards, possibly cheaper	Mature ecosystem, first-mover advantage
Disadvantage	Weak ecosystem, 1 year late	Expensive

Conclusion: If Jaguar Shores launches on time with sufficient oneAPI ecosystem improvement, it can become NVIDIA's third choice (after NVIDIA and AMD).

Software Ecosystem: oneAPI Progress and Challenges​

What is oneAPI?​

oneAPI is Intel's open, cross-architecture programming model:

• Supports CPU, GPU, FPGA, AI accelerators
• Based on SYCL standard (similar to CUDA's C++ extensions)
• Open-source implementation (Intel oneAPI Base Toolkit)

Intel AI Stack​

Component	Purpose	Counterpart
oneAPI	Cross-architecture programming model	CUDA
BigDL	Distributed deep learning framework	PyTorch Distributed
Gaudi Software Suite	Gaudi-specific software stack	NVIDIA GPU Cloud (NGC)
Intel Extension for PyTorch	PyTorch optimization on Intel hardware	NVIDIA PyTorch
Intel Optimization for TensorFlow	TensorFlow optimization on Intel hardware	NVIDIA TensorFlow

✅ Progress​

• PyTorch 2.5+: Intel Extension integrated into PyTorch mainline
• Hugging Face Transformers: Official Intel GPU support (via optimum-intel)
• vLLM: Experimental Gaudi support (performance TBD)

⚠️ Challenges​

• Developer habits: Global AI developers use CUDA; oneAPI has a steep learning curve
• Operator coverage: Many PyTorch operators lack oneAPI-optimized versions
• Performance: At same power, Gaudi 4 performance is only ~50% of B200

Industry Impact​

1. Can Intel Return to the AI Race?​

Challenges:

• ❌ Ecosystem disadvantage: CUDA moat too deep, oneAPI hard to shake
• ❌ Performance disadvantage: Gaudi 4 only ~50% of B200
• ❌ Timing disadvantage: Jaguar Shores 1 year later than Rubin

Opportunities:

• ✅ Open standards: Not dependent on CUDA, suited for "anti-NVIDIA-monopoly" customers
• ✅ Ethernet advantage: RoCE v3 cheaper than InfiniBand at ten-thousand-card scale
• ✅ Intel Foundry: If Jaguar Shores uses Intel's own process, lower cost

2. Impact on AMD​

Intel's return to the AI race is bad for AMD:

• AMD was the "only NVIDIA alternative"
• Now Intel is back too; AMD's "alternative" status is challenged
• But in the short term (2026-2027), Intel cannot yet threaten AMD

3. Impact on Domestic Chips​

Intel Gaudi 4's launch is a reference case for domestic chips:

• Proves the Ethernet route (RoCE) is viable
• Proves open ecosystem (oneAPI) is hard but necessary
• Proves the cost-effective route has a market (cost-sensitive customers)

Related Chips​

• Intel Gaudi 3 - Previous-gen product
• Intel Gaudi 4 - New launch (to be created)
• Intel Crescent Island - Intel's first general GPU (to be created)
• NVIDIA B200 - Direct competitor
• AMD MI455X - Same-gen competitor

References​

• Intel Newsroom: Gaudi 4 Launch
• Tom's Hardware: Intel Jaguar Shores to use HBM4
• Wccftech: Intel Next-Gen Jaguar Shores GPU
• U.S. News: Intel Signals Return to AI Race

---

This article is compiled from Intel official announcements and public materials. Some specs are estimates, subject to final Intel release.

On June 1, 2026, NVIDIA founder and CEO Jensen Huang officially announced at COMPUTEX 2026 (Taipei) that: the Vera Rubin platform has entered full production. This marks a fundamental paradigm shift for AI hardware from "discrete accelerators" to "integrated AI factories."

Key Highlights​

• Rubin GPU: Next-gen AI compute chip, FP4 compute is 3.6× that of Blackwell
• Vera CPU: 88 custom Arm cores (176 threads), replacing the Grace CPU
• NVLink 6: GPU-to-GPU interconnect bandwidth reaches 260 TB/s (double Blackwell)
• CX8 SuperNIC: 800Gb/s network, ConnectX-9 link reaching 28.8 TB/s
• HBM4 memory: 288GB per chip, 13 TB/s bandwidth
• Agentic throughput: 10× over Grace Blackwell

Complete Vera Rubin Platform Specs​

Vera Rubin is not a single GPU but a complete AI factory platform comprising 7 chips:

Chip	Type	Purpose
Rubin GPU	Main AI compute chip	Training + inference
Rubin Ultra GPU	Flagship version	Ultra-scale inference
Vera CPU	CPU paired with Rubin	Host CPU + data preprocessing
NVLink 6	Interconnect chip	High-speed GPU interconnect (260 TB/s)
CX8 SuperNIC	NIC	800Gb/s network
XDR 800G switch	Datacenter network	Cross-rack communication
Rubin Platform POD	Whole cabinet	Pre-configured AI factory (144 GPUs)

Rubin GPU Detailed Specs (estimated)​

Parameter	Rubin GPU	Rubin Ultra	Blackwell (B200)
Architecture	Rubin	Rubin Ultra	Blackwell
Process	TSMC 3nm (est.)	TSMC 3nm	TSMC 4NP
Memory	288GB HBM4	288GB HBM4E (est.)	192GB HBM3e
Memory bandwidth	13 TB/s	13+ TB/s	8 TB/s
FP4 compute	~3,600 TFLOPS (est.)	~5,000 TFLOPS (est.)	2,250 TFLOPS
TDP	1,000W (est.)	1,200W (est.)	700-1000W
Interconnect	NVLink 6 (260 TB/s)	NVLink 6	NVLink 5 (1800 GB/s)
Mass production	2026 Q3	H2 2027	2024 Q4

📌 Note: Rubin's exact specs are not fully public yet; some values above are estimates.

Vera CPU: The New Host CPU Replacing Grace​

Vera CPU is NVIDIA's self-designed Arm-architecture CPU, replacing the previous Grace CPU:

Parameter	Vera CPU	Grace CPU
Cores	88 cores (176 threads)	72 cores (144 threads)
Architecture	Custom Armv9 (est.)	Arm Neoverse V2
Interface	NVLink 5.0 (1.8 TB/s)	NVLink 4.0 (900 GB/s)
TDP	~500W (est.)	350-500W
Purpose	AI factory Host CPU	HPC / AI Host

Key upgrade: Vera's co-design with the Rubin GPU achieves end-to-end optimization in compute, data loading, and preprocessing, comparable to Google TPU 8t's Arm Axion integration.

Performance vs Blackwell​

NVIDIA officially claims that under the same POD configuration (144 GPU chips):

Metric	Grace Blackwell (GB200 NVL72)	Vera Rubin NVL144	Improvement
FP4 compute	1.1 PFLOPS	3.6 PFLOPS	3.3×
Memory capacity	288GB×72 = 20.7TB	288GB×144 = 41.4TB	2×
Memory bandwidth	8 TB/s×72	13 TB/s×144	~3.3×
NVLink bandwidth	1800 GB/s×72	260 TB/s (full POD)	~2×
Agentic throughput	Baseline	10×	10×
Performance per watt	Baseline	25× (vs CPU alone)	25×

💡 Why "10× agentic throughput"? Agentic AI workloads differ from training/inference: one prompt may trigger multiple stages including reasoning, retrieval, tool calls, and response generation, involving thousands of steps. The Rubin platform is optimized for this long-chain, high-concurrency workload.

MGX Third-Gen Rack-Scale System​

Vera Rubin adopts the MGX third-gen open rack-scale system design:

• Five-rack synergy: Vera Rubin NVL72 system + Vera CPU + Groq 3 LPX + Vera BlueField-4 STX storage + Spectrum-6 SPX Ethernet
• Global supply chain: 30 countries, 350+ factories, hundreds of partners (Dell, HPE, Lenovo, Supermicro, Asus, Foxconn, etc.)
• Spectrum-X Ethernet silicon photonics: World's first switch based on CPO (co-packaged optics) supporting 200Gb/s SerDes, now in mass production

Mass Production Timeline​

Time	Event
Jan 2026	CES 2026 first unveils Rubin platform
June 1, 2026	COMPUTEX 2026 announces full production
Fall 2026	Vera Rubin officially starts mass production and shipment
H2 2027	Rubin Ultra launch (HBM4E upgrade)
2028	Feynman architecture (next gen)

AI Factory: From Selling Chips to Selling "Smart Production Lines"​

Huang said something at the launch that shook the industry:

"Rubin's Agentic AI throughput is 10× that of Blackwell. Rubin is a complete AI factory platform."

This marks a fundamental shift in NVIDIA's business model:

• Past: Sold GPUs (H100/B200), customers built systems themselves
• Now: Sells "complete AI factory solutions" (Vera Rubin POD), including GPU, CPU, network, storage, software stack
• Future: Becomes the "TSMC" of global AI infrastructure (providing smart production capacity)

vs Competitors​

Vendor	Product	Positioning	Advantage	Disadvantage
NVIDIA	Vera Rubin	Complete AI factory solution	Most complete ecosystem, most mature software	Expensive, extremely high power
AMD	MI455X (MI400 series)	Training competitor	Price/performance, open ecosystem	Software ecosystem gap
Google	TPU 8i/8t	Cloud training/inference	Deep Gemini integration	Google Cloud only
Huawei	Ascend 910C/950	Domestic substitution	China localization, AscendMind framework	Affected by export controls

Industry Impact​

1. AI labs: Frontier model training time shrinks from "months" to "weeks"
2. Cloud providers: Must decide whether to procure Vera Rubin POD (conflicts with self-developed chip strategy)
3. Hyperscale datacenters: AI factory becomes a new competitive dimension (whoever has the strongest compute can train the strongest model)
4. Domestic chips: Ascend 910C/950, Cambricon MLU590, etc. must catch up to Blackwell in 2026-2027, or the gap will widen to the Rubin era

Related Chips​

• NVIDIA Rubin R200 - Rubin architecture GPU detailed specs
• NVIDIA Vera CPU - Vera CPU detailed specs
• NVIDIA B200 - Previous-gen Blackwell architecture
• AMD MI455X - Direct competitor (MI400 series)

References​

• NVIDIA Blog: Vera Rubin Full Production
• COMPUTEX 2026 Keynote
• NVIDIA Official Newsroom

---

This article is compiled from NVIDIA official announcements and public materials. Some specs are estimates, subject to final official release.

In 2026, China's domestic AI chip industry has entered a period of full-scale explosion. The three giants — Huawei Ascend, Baidu Kunlun, and Alibaba T-Head — have successively launched next-generation products, while Cambricon, MetaX, Enflame, and Iluvatar have also achieved important breakthroughs.

This article comprehensively analyzes 2026 domestic AI chip progress across four dimensions: product launches, technology breakthroughs, market dynamics, and ecosystem building.

---

1. Huawei Ascend: 950 series launched, 960/970 roadmap clear​

1.1 Ascend 950PR (launched Q1 2026)​

Core specs:

Item	Parameter
Launch date	March 21, 2026
Platform	Atlas 350 accelerator card
HBM capacity	128 GB (Huawei in-house HiBL 1.0 HBM)
Memory bandwidth	1.6 TB/s
FP8 compute	1 PFLOPS
Positioning	Inference-specific (Prefill stage)
Performance vs.	Single-card compute is 2.87× NVIDIA H20

Technology innovations:

• First adoption of Huawei in-house HBM solution (HiBL 1.0), lowering cost
• Supports FP8 low-precision compute, 3× inference energy-efficiency improvement
• Optimized for inference scenarios such as video recommendation and real-time interaction

Commercialization progress:

• Mass supply began in Q1 2026
• Key customers: China Telecom, China Mobile, China Unicom, Huawei Cloud
• Priced around ¥100,000/card (¥80,000 for key accounts), ~30% lower than comparable competitors

1.2 Ascend 950DT (launched Q4 2026)​

Core specs:

Item	Parameter
Launch date	Q4 2026 (expected October)
HBM capacity	144 GB (Huawei in-house HiZQ 2.0 HBM)
Memory bandwidth	4 TB/s (HiZQ 2.0 technology)
FP8 compute	1 PFLOPS
Positioning	Inference + training (Decode stage + training tasks)
Technology innovation	First to carry in-house HiZQ 2.0 memory technology

Technology innovations:

• Adopts HiZQ 2.0 memory technology, 2× data-movement efficiency
• Supports FP8/FP4 low-precision compute, balancing performance and efficiency
• Optimized for scenarios such as dialogue generation and large-model training

1.3 950 SuperNode (launched Q4 2026)​

System specs:

Item	Configuration
Max interconnected chips	8,192 chips
Total FP8 compute	1 EFLOPS (1,024-card scale)
1024-card version	16 liquid-cooled cabinets, 64 chips per cabinet
Supported models	Trillion-parameter large-model training
Deployment progress	1,024-card version already in deployment

Performance comparison:

• 950 SuperNode outperforms NVIDIA's 2027 NVL576 system
• Leads by 20% in trillion-parameter model training

1.4 Ascend 960/970 roadmap​

Chip model	Launch	Core specs	Positioning
Ascend 960	Q4 2027	N+3 process, 288GB HBM, FP8 2 PFLOPS, 30%+ better energy efficiency than 910C	Ultra-large-scale training
Ascend 970	Q4 2028	N+3 process, FP4 8 PFLOPS, 4 TB/s bandwidth, supports trillion-parameter models	Next-gen AI architecture (MoE, etc.)

Technology breakthroughs:

• Process upgrade: from N+2 (7nm-class) to N+3 (5nm-class)
• Memory capacity doubled: from 144GB (950DT) to 288GB (960/970)
• Energy efficiency improved: 960/970 are 30%+ better than 910C
• Precision optimized: 970 supports FP4 precision, optimized for next-gen AI architectures (MoE, etc.)

1.5 Commercialization progress​

Shipment data:

• 384-card SuperNode: over 500 units deployed, the only truly large-scale commercial SuperNode in China
• 2026 shipment target: 800k chips (1M cumulative)
• Market share: 60% of China's AI chip market

Ecosystem building:

• CANN compiler: open-sourced end of 2025, seamless PyTorch/TensorFlow migration
• Mind series toolchains: fully open, lowering the developer barrier
• Ecosystem partners: over 3,000
• Developer community: over 500k registered developers

---

2. Baidu Kunlun: M100 inference-specific, Tianchi SuperNode deployed​

2.1 Kunlun M100 (launched early 2026)​

Core specs:

Item	Parameter
Launch date	Early 2026 (expected Q2)
Positioning	Inference-specific
Architecture	In-house XPU-P architecture (inference-optimized)
Process	7nm (SMIC N+2)
HBM capacity	64 GB (inference-optimized)
TDP	250 W (low-power inference)
Performance vs.	1.5× P800 inference, 38% lower power

Technology innovations:

• Adopts RISC-V open instruction set, adding 50+ AI-specific instructions
• Compute per watt reaches 8.3 TOPS/W, 2.1× the industry average
• Supports models from 10 billion to 100 billion parameters for inference

Commercialization progress:

• Mass supply in Q2 2026
• Key customers: Baidu Smart Cloud, China Merchants Bank, Southern Grid, Geely Auto
• Priced around ¥60,000/card, clear cost-performance advantage

2.2 Kunlun M300 (launched early 2027)​

Core specs:

Item	Parameter
Launch date	Q1 2027 (expected March)
Positioning	Ultra-large-scale multimodal training
Architecture	In-house XPU-P architecture (multimodal-optimized)
Process	5nm (SMIC N+3)
HBM capacity	256 GB HBM4
TDP	500 W
Modalities	Text, images, video and other data types

Technology innovations:

• Adopts HBM4 memory, bandwidth up to 3.2 TB/s
• Supports FP8/FP4 low-precision compute, 2× training energy-efficiency
• Native support for multimodal model training (text + image + video)

2.3 Tianchi 256-card SuperNode (launched June 2026)​

System specs:

Item	Configuration
Launch date	June 2026 (expected)
Chip count	256 Kunlun P800/M100
Effective training rate	97%
Inter-chip bandwidth	1.2 TB/s
Validated models	Baidu ERNIE 5.1 and other key large models

Performance breakthroughs:

• Fully domestic SuperNode, fully autonomous and controllable from chip to network
• 97% effective training rate, surpassing NVIDIA DGX SuperPOD's 95%
• Training validation completed for Baidu ERNIE 5.1 and other key large models

2.4 Commercialization progress​

Shipment data:

• P800: 150k shipped in 2025, 200k target in 2026
• Ten-thousand-card clusters: multiple P800-based clusters delivered
• Market share: 20% of China's AI chip market

Customer coverage:

• External customer revenue share: over 50% in 2025
• China Mobile AI server procurement: P800-based bids won 70%, 70%, 100% shares
• Key customers: China Merchants Bank, Southern Grid, Geely Auto, iFlytek

IPO progress:

• May 2026: officially launched STAR Market IPO tutoring
• Plans "A+H" model — simultaneous A-share and Hong Kong listings
• Valuation exceeds ¥10 billion

---

3. Alibaba T-Head: M890 3× performance, Zhenwu series ships 560k units​

3.1 T-Head M890 (launched Q2 2026)​

Core specs:

Item	Parameter
Launch date	Q2 2026 (Alibaba Cloud Summit)
Performance	3× previous generation
HBM capacity	144 GB
Inter-chip bandwidth	800 GB/s
Precision support	FP8, FP4 low-precision compute
Positioning	Full training + inference pipeline

Technology innovations:

• Adopts in-house ICN inter-chip protocol, inter-chip latency under 150 ns
• Companion PCCF communication library and ICN Switch chip enable full-bandwidth interconnect of 64 chips within a single node
• Supports FP8/FP4 low-precision compute, balancing performance and efficiency

3.2 T-Head V900 (launched Q3 2027)​

Core specs:

Item	Parameter
Launch date	Q3 2027 (expected September)
Performance	3× again over M890
HBM capacity	216 GB
Inter-chip bandwidth	1,200 GB/s
Positioning	Ultra-large-scale training

3.3 T-Head G900 (launched Q3 2028)​

Core specs:

Item	Parameter
Launch date	Q3 2028 (expected September)
Positioning	Flagship product for next-gen compute demand
Technology innovation	Supports full-pipeline training of trillion-parameter models

3.4 Zhenwu series commercialization progress​

Shipment data:

• Cumulative shipments: over 560k units as of April 2026
• Customers served: 20+ industries, 400+ customers
• Autonomous driving: over 130k units, 30+ customers
• Finance: over 100k units, 150+ customers

Performance advantage:

• At equal precision, Zhenwu series single-machine inference outperforms comparable products by 50%+ on average
• Panjiu server SuperNode architecture supports trillion-parameter large models on a single node

Full product line:

• Zhenwu series AI chips: training + inference
• Yitian series CPUs: data center CPUs
• ICN Switch interconnect chip: inter-chip interconnect
• Camel920 400G smart NIC: high-speed networking
• Junyue series storage controller chips: storage optimization

---

4. Other domestic chip vendor progress​

4.1 Cambricon MLU590 (launched Q1 2026)​

Core specs:

Item	Parameter
Launch date	Q1 2026 (expected March)
Architecture	MLUarch 09 (in-house)
Process	7nm (SMIC N+2)
HBM capacity	128 GB HBM3
TDP	350 W
Positioning	Training + inference

Technology innovations:

• MLUarch 09 architecture, 2× compute over MLU590
• Supports FP8/FP4 low-precision compute, 2.5× inference energy-efficiency
• Native MoE architecture support, 3× sparse-model inference efficiency

Commercialization progress:

• Q1 2026: sample deliveries begun
• Key customers: Chinese government, state-owned enterprises, research institutes
• Day-0 adaptation of DeepSeek-V3 671B achieved

4.2 MetaX Xiyun C600 (launched Q2 2026)​

Core specs:

Item	Parameter
Launch date	Q2 2026 (expected June)
Architecture	MXMACA 3.0 (CUDA-compatible)
Process	7nm (SMIC N+2)
HBM capacity	128 GB HBM3
TDP	350 W
Positioning	Training + inference

Technology innovations:

• MXMACA 3.0 architecture, CUDA-compatible, low migration cost
• Supports FP8/FP4 low-precision compute, 2× training energy-efficiency
• Fully domestic supply chain, autonomous and controllable from chip to packaging

Commercialization progress:

• Q2 2026: sample deliveries begun
• Key customers: Chinese government, state-owned enterprises, research institutes
• Adapted models include LLaMA, ChatGLM, Baichuan

4.3 Enflame S60 (launched Q3 2026)​

Core specs:

Item	Parameter
Launch date	Q3 2026 (expected September)
Architecture	GCU 3.0 (in-house)
Process	7nm (SMIC N+2)
HBM capacity	96 GB HBM3
TDP	300 W
Positioning	Inference-specific

Technology innovations:

• GCU 3.0 architecture, 2.5× inference performance over S30
• Supports FP8 low-precision compute, 3× inference energy-efficiency
• Hardware-level virtualization, single card split into 64 virtual instances

Commercialization progress:

• Q3 2026: sample deliveries begun
• Key customers: Tencent Cloud, China Telecom, China Unicom
• Priced around ¥50,000/card

4.4 Iluvatar VA10 (launched Q4 2026)​

Core specs:

Item	Parameter
Launch date	Q4 2026 (expected December)
Architecture	HVMA 2.0 (in-house)
Process	7nm (SMIC N+2)
HBM capacity	64 GB HBM3
TDP	250 W
Positioning	Video processing + AI inference

Technology innovations:

• HVMA 2.0 architecture, 3× video processing performance over VA10
• Supports 8K real-time video processing, 2× video AI inference performance
• Hardware-level video codec, supports H.264/H.265/AV1

Commercialization progress:

• Q4 2026: sample deliveries begun
• Key customers: ByteDance, Kuaishou, Bilibili
• Priced around ¥40,000/card

4.5 Hygon DCU K100 (launched Q2 2026)​

Core specs:

Item	Parameter
Launch date	Q2 2026 (expected June)
Architecture	x86-compatible GPGPU (in-house DCU)
Process	7nm (SMIC N+2)
HBM capacity	128 GB HBM3
TDP	400 W
Positioning	Training + inference (x86 ecosystem-compatible)

Technology innovations:

• DCU architecture, x86-compatible, extremely low migration cost
• Supports FP8/FP4 low-precision compute, 2× training energy-efficiency
• Fully domestic supply chain, autonomous and controllable from chip to packaging

Commercialization progress:

• Q2 2026: sample deliveries begun
• Key customers: Chinese government, state-owned enterprises, research institutes
• Adaptation of DeepSeek-V3 671B achieved

---

5. 2026 domestic AI chip market landscape​

5.1 Market share (2026)​

Vendor	Market share	Shipments (10k units)	Flagship products
Huawei Ascend	60%	80	910C, 950PR, 950DT
Baidu Kunlun	20%	20	P800, M100
Alibaba T-Head	10%	10	M890, Zhenwu series
Cambricon	5%	5	MLU590
MetaX	3%	3	C600
Others	2%	2	S60, VA10, K100

5.2 Technology roadmap comparison​

Vendor	Architecture route	Ecosystem compatibility	Process	Supply chain
Huawei Ascend	Da Vinci (in-house)	CANN (CUDA-compatible)	SMIC N+2/N+3	Fully domestic
Baidu Kunlun	XPU-P (in-house)	XPU-P (CUDA-compatible)	SMIC N+2/N+3	Fully domestic
Alibaba T-Head	In-house RISC-V	CUDA-compatible	SMIC N+2/N+3	Fully domestic
Cambricon	MLUarch (in-house)	CANN (CUDA-compatible)	SMIC N+2	Fully domestic
MetaX	MXMACA (CUDA-compatible)	CUDA-compatible	SMIC N+2	Fully domestic
Enflame	GCU (in-house)	In-house ecosystem	SMIC N+2	Fully domestic
Iluvatar	HVMA (in-house)	In-house ecosystem	SMIC N+2	Fully domestic
Hygon	DCU (x86-compatible)	x86 ecosystem-compatible	SMIC N+2	Fully domestic

5.3 Supply chain security comparison​

Vendor	Wafer fab	HBM supply	Packaging/test	Supply chain rating
Huawei Ascend	SMIC	Huawei in-house HiBL/HiZQ	JCET/TFME	⭐⭐⭐⭐⭐
Baidu Kunlun	SMIC	ChangXin Memory	JCET/TFME	⭐⭐⭐⭐⭐
Alibaba T-Head	SMIC	ChangXin Memory	JCET/TFME	⭐⭐⭐⭐⭐
Cambricon	SMIC	Samsung/Hynix	JCET/TFME	⭐⭐⭐⭐
MetaX	SMIC	Samsung/Hynix	JCET/TFME	⭐⭐⭐⭐
Enflame	SMIC	Samsung/Hynix	JCET/TFME	⭐⭐⭐⭐
Iluvatar	SMIC	Samsung/Hynix	JCET/TFME	⭐⭐⭐⭐
Hygon	SMIC	Samsung/Hynix	JCET/TFME	⭐⭐⭐⭐

---

6. 2026 domestic AI chip technology breakthroughs​

6.1 Process breakthroughs​

Node	2026 status	Representative products	Notes
7nm (N+2)	Mass production	910C, P800, M890	SMIC N+2 mature
5nm (N+3)	Mass production	960, 970, M300	SMIC N+3 mass production in 2026
3nm	In development	Next-gen products	Mass production expected 2028

6.2 Packaging breakthroughs​

Packaging	2026 status	Representative products	Notes
Chiplet	Mature	910C, 950PR/DT	Dual-die packaging, higher yield
3D stacking	Mature	P800, M890	HBM3e 3D stacking
CoWoS	Mature	All high-end products	TSMC CoWoS
Domestic packaging	Mass production	960, 970, M300	JCET/TFME mass production

6.3 Memory breakthroughs​

Memory	2026 status	Representative products	Notes
HBM2E	Mature	910C	Samsung supply
HBM3	Mature	P800, MLU590, C600	Samsung/Hynix supply
HBM3e	Mature	950PR, M890	Samsung/Hynix supply
Huawei in-house HBM	Mass production	950PR (HiBL 1.0), 950DT (HiZQ 2.0)	Huawei in-house, lower cost
HBM4	In development	M300 (2027)	Mass production expected 2027

6.4 Interconnect breakthroughs​

Interconnect	2026 status	Representative products	Notes
AscendLink	Mature	910C, 950PR/DT	Huawei in-house, 784 GB/s
XCCL	Mature	P800, M100	Kunlun in-house, 1.2 TB/s
ICN	Mature	M890, V900	Alibaba in-house, 800 GB/s
Domestic optical modules	Mass production	All SuperNodes	6,912 LPO optical modules

---

7. 2026 domestic AI chip ecosystem building​

7.1 Software ecosystem comparison​

Vendor	Software stack	CUDA compatibility	Framework support	Developer community
Huawei Ascend	CANN + MindSpore	Compatible (low migration cost)	PyTorch/TensorFlow/MaxMind	500k+
Baidu Kunlun	XPU-P + PaddlePaddle	Compatible (low migration cost)	PyTorch/TensorFlow/PaddlePaddle	300k+
Alibaba T-Head	In-house + Alibaba Cloud	Compatible (low migration cost)	PyTorch/TensorFlow/Alibaba Cloud	200k+
Cambricon	CANN + MindSpore	Compatible (low migration cost)	PyTorch/TensorFlow	100k+
MetaX	MXMACA + CUDA	Compatible (very low migration cost)	PyTorch/TensorFlow/CUDA	50k+
Enflame	In-house GCU stack	Incompatible (rewrite needed)	PyTorch/TensorFlow	30k+
Iluvatar	In-house HVMA stack	Incompatible (rewrite needed)	PyTorch/TensorFlow	20k+
Hygon	DCU + x86	x86-compatible (very low migration cost)	PyTorch/TensorFlow/x86	50k+

7.2 Developer community building​

Vendor	Developers	Docs	Dev tools	Training/cert
Huawei Ascend	500k+	Complete	CANN Toolkit	HCCP cert
Baidu Kunlun	300k+	Complete	XPU-P Toolkit	PaddlePaddle cert
Alibaba T-Head	200k+	Complete	Alibaba Cloud Toolkit	Alibaba Cloud cert
Cambricon	100k+	Fairly complete	CANN Toolkit	Cambricon cert
MetaX	50k+	Fairly complete	MXMACA Toolkit	MetaX cert
Enflame	30k+	Average	GCU Toolkit	Enflame cert
Iluvatar	20k+	Average	HVMA Toolkit	Iluvatar cert
Hygon	50k+	Complete	DCU Toolkit	Hygon cert

7.3 Large-model adaptation capability​

Vendor	DeepSeek-V3	LLama 3	ChatGLM	Baichuan	ERNIE	Qwen
Huawei Ascend	✅ Day-0	✅	✅	✅	✅	✅
Baidu Kunlun	✅ Day-0	✅	✅	✅	✅	✅
Alibaba T-Head	✅ Day-0	✅	✅	✅	✅	✅
Cambricon	✅ Day-0	✅	✅	✅	✅	✅
MetaX	✅ Day-1	✅	✅	✅	✅	✅
Enflame	✅ Day-3	✅	✅	✅	✅	✅
Iluvatar	✅ Day-7	✅	✅	✅	✅	✅
Hygon	✅ Day-3	✅	✅	✅	✅	✅

---

8. 2026 domestic AI chip market trends​

8.1 Market drivers​

Driver	Description
Policy support	The national 15th Five-Year Plan incorporates the compute network as a major project, with stronger policy support
Supply chain security	Escalating US export controls make domestic chips the only option
Cost advantage	Domestic chips are 30-50% cheaper than imports, clear cost-performance edge
Technology breakthroughs	Comprehensive breakthroughs in compute, memory, and energy efficiency
Maturing ecosystem	Software ecosystems (CANN, XPU-P, MXMACA) reach 60-70% of CUDA maturity

8.2 Market challenges​

Challenge	Description
Process	7nm/5nm still lags NVIDIA's 4nm/3nm
HBM bandwidth	Domestic HBM bandwidth still lags NVIDIA
Software ecosystem	Ecosystem maturity still lags CUDA
Capacity bottleneck	Limited SMIC N+2/N+3 capacity, supply falls short of demand
International competition	NVIDIA, AMD, Google and others keep innovating

8.3 Market forecast (2026-2030)​

Year	China AI chip market (¥B)	Domestic share	Domestic market (¥B)	Notes
2026	500	35%	175	Ascend 60%, Kunlun 20%
2027	700	50%	350	960/970 launch, breakthroughs
2028	1,000	65%	650	Domestic tech approaches international level
2029	1,500	80%	1,200	Domestic tech surpasses international level
2030	2,000	90%	1,800	Substitution essentially complete

---

9. Summary and outlook​

9.1 Core conclusions​

1. 2026 marks the full-scale explosion of domestic AI chips, as the three giants Huawei Ascend, Baidu Kunlun, and Alibaba T-Head successively launch next-gen products
2. Significant technology breakthroughs across compute, memory, energy efficiency, and system scaling
3. Controllable supply chain security, fully autonomous from wafer fab to packaging and test
4. Accelerating ecosystem building, software ecosystem maturity reaching 60-70% of CUDA
5. Rising market share, domestic chips take 35% of China's AI chip market in 2026, projected 90% by 2030

9.2 Future outlook​

Short term (2026-2027):

• Huawei Ascend 950PR/950DT mass deployment, clear 960/970 roadmap
• Baidu Kunlun M100 inference chip ramps, M300 ultra-large multimodal training chip launches
• Alibaba T-Head M890 3× performance, V900 launches
• Domestic chip market share rises to 50%

Medium term (2028-2029):

• Huawei Ascend 960/970 mass production, 5nm process, 8 PFLOPS FP4 compute
• Baidu Kunlun M300 mass production, supports trillion-parameter multimodal training
• Alibaba T-Head G900 launches, becoming the next-gen compute flagship
• Domestic tech approaches international level, market share to 80%

Long term (2030+):

• Domestic AI chips exceed 20% of the global market
• Transition from "following" to "running alongside" to "leading"
• Huawei Ascend, Baidu Kunlun, Alibaba T-Head among the global TOP 5
• China becomes a global center of AI chip technology innovation

---

References​

1. Domestic AI chip "three powers" rise: substitution trend shifts from policy-driven to market-driven — Sohu
2. 2026 domestic AI chip panorama: Huawei Ascend races Cambricon — ZPEDU
3. Huawei unveils three-year Ascend AI chip roadmap — Jiemian News
4. Ascend 950PR chip — Baidu Baike
5. Ascend 950 chip — Baidu Baike
6. Kunlun P800: technical breakthroughs and application prospects of a new-generation AI accelerator — YunTECH
7. Kunlun P800 latest specs: P800 single-precision compute reaches 345 TFLOPS — Xueqiu

---

Last updated: June 10, 2026

Huawei Ascend 910C (Ascend 910C), Huawei's third-generation Ascend AI chip, adopts innovative dual-die (Chiplet) packaging and began mass supply in May 2025, becoming the backbone of domestic AI compute.

This article comprehensively analyzes this domestic flagship AI chip across four dimensions: technical specs, deployment cases, performance comparison, and market positioning.

---

1. Core Technical Specifications​

1.1 Chip architecture and process​

Item	Parameter
Architecture	Da Vinci (dual-die packaging)
Process	SMIC N+2 (7nm-class)
Packaging	Chiplet (2× Ascend 910B compute dies)
Transistors	~53 billion
Die size	~800mm² (estimated)

Technology highlights:

• Dual-die Chiplet packaging integrates two 910B chips, breaking the single-die yield bottleneck
• Centerless I/O die design lets the two compute dies interconnect directly, reducing communication latency
• SMIC N+2 process delivers 7nm-class performance with a controllable, autonomous supply chain

1.2 Compute performance​

Precision	Compute	Reference
BF16	800 TFLOPS	~60% of NVIDIA H100
FP16	~800 TFLOPS	Close to H100 at same precision
INT8	~1600 TOPS	Clear inference advantage
FP32	Not disclosed	Training mainly uses BF16/FP16

Performance characteristics:

• 800 TFLOPS at BF16, a new domestic AI chip compute benchmark
• ~2× compute over 910B (dual-die stacking + architecture optimization)
• No FP8 precision support (NVIDIA Blackwell's strength)

1.3 Memory and interconnect​

Item	Parameter
HBM type	HBM2E (8 stacks)
Memory capacity	~128 GB (combined dual-die)
Memory bandwidth	784 GB/s
Interconnect protocol	Huawei AscendLink (in-house)
Interconnect bandwidth	400 GB/s unidirectional (800 GB/s bidirectional)

Memory advantages:

• 128GB capacity supports full-pipeline training of hundred-billion-parameter models
• 784 GB/s is a high-end configuration among HBM2E solutions
• In-house AscendLink protocol supports 384-chip all-optical interconnect

1.4 Power and energy efficiency​

Item	Parameter
TDP (dual-die)	~310 W
Energy efficiency (BF16)	~2.58 TFLOPS/W
vs. H100	~45% of H100's power, comparable energy efficiency

Energy efficiency advantages:

• At equal compute, significantly lower power than NVIDIA H100 (700W)
• 7nm-class process, ~30% better energy efficiency than 910B
• Suited to large-scale cluster deployment, reducing data center PUE pressure

---

2. Key Deployment Cases​

2.1 CloudMatrix 384 SuperNode​

System specs:

Item	Configuration
Chip count	384 Ascend 910C
Cabinets	16 (12 compute + 4 network)
Total HBM	~49 TB (128GB × 384)
Interconnect	All-optical mesh network
Optical modules	6,912 LPO optical modules
System BF16 compute	~300 PFLOPS

Performance comparison:

• CloudMatrix 384's total BF16 compute exceeds NVIDIA GB200 NVL72 (72× B200)
• In large-model training, 384-chip 910C linear scaling efficiency reaches 85%+
• Supports smooth scaling to ten-thousand-card clusters for ultra-large training

Deployment progress:

• As of June 2026, over 500 CloudMatrix 384 SuperNodes deployed
• Key customers: China Telecom, China Mobile, China Unicom, Huawei Cloud, iFlytek
• Scenarios: large-model training, smart customer service, autonomous-driving simulation, scientific computing

2.2 DeepSeek-V4-Pro full-parameter post-training​

Breakthrough significance:

On June 5, 2026, the AI training platform of Shenzhen Hetao College — together with Harbin Institute of Technology (Shenzhen), Shenzhen Big Data Research Institute, Huawei, and Shenzhen Zhicheng AI Compute Platform — completed full-parameter post-training of the 1.6-trillion-parameter DeepSeek-V4-Pro large model on an Ascend 910C compute cluster.

Technical highlights:

• Among the world's first to run full-parameter post-training of a trillion-parameter model on a domestic compute platform
• Validates Ascend 910C maturity in ultra-large-model training
• Proves domestic AI chips now have the capability to replace imported chips

Performance data (official disclosure):

• Training throughput: ~60% of an H100 cluster (BF16 precision)
• Memory utilization: 92% (128GB HBM2E capacity advantage)
• Interconnect efficiency: 384-chip linear scaling efficiency 85%+
• Stability: 30 consecutive days of training with no failures

2.3 Commercial deployment cases​

Case 1: A provincial big-data center (300 P FLOPS compute center)​

• Scale: 300 P FLOPS AI compute (~1,000× 910C)
• Scenarios: government large model, city brain, smart transportation
• Deployment: September 2025
• Investment: ~¥200M (120 servers)

Case 2: Huawei Cloud AI training platform​

• Chips: over 10,000 Ascend 910C
• Customers served: over 500 enterprises
• Model support: Pangu large model, third-party open-source models (LLaMA, ChatGLM, etc.)
• Global deployment: China, Southeast Asia, Middle East, Latin America

Case 3: iFlytek smart education​

• Scale: 256 Ascend 910C
• Scenarios: smart-education large model, speech recognition, machine translation
• Performance: 90% faster training than 910B

---

3. Performance Comparison Analysis​

3.1 vs. NVIDIA H100​

Item	Ascend 910C	NVIDIA H100	Notes
BF16 compute	800 TFLOPS	~1,300 TFLOPS	910C ~60% of H100
HBM capacity	128 GB	80 GB	910C +60%
HBM bandwidth	784 GB/s	3.35 TB/s	H100 clear bandwidth lead
TDP	310 W	700 W	910C only 45% of H100 power
Process	7nm (SMIC N+2)	4nm (TSMC)	H100 more advanced
Software ecosystem	CANN (CUDA-compatible)	CUDA	H100 more mature
Supply	China autonomous	Export-controlled	910C no supply-chain risk

Conclusion:

• In raw compute, 910C is ~60% of H100
• In memory capacity, 910C leads by 60%, suited to large-model training
• In energy efficiency, 910C clearly outperforms H100
• In supply chain security, 910C wins outright

3.2 vs. Ascend 910B​

Item	Ascend 910C	Ascend 910B	Improvement
Architecture	Dual-die Chiplet	Single die	—
BF16 compute	800 TFLOPS	~400 TFLOPS	+100%
HBM capacity	128 GB	64 GB	+100%
TDP	310 W	310 W	Flat (single-die power)
Process	SMIC N+2	SMIC N+2	Same
Yield	~40%	~30%	+33%

Conclusion:

• 910C's dual-die packaging doubles compute and memory capacity
• Yield up from 910B's 30% to 40%, lowering manufacturing cost
• At equal power, 100% performance gain, significantly better energy efficiency

3.3 Inference performance (DeepSeek measured)​

Test environment:

• Model: DeepSeek-V3 (671B parameters)
• Hardware: Ascend 910C vs NVIDIA H100
• Precision: BF16
• Batch size: 64

Results:

Metric	Ascend 910C	NVIDIA H100	Ratio
Inference speed (tokens/s)	8,500	14,200	60%
First-token latency (ms)	120	85	141%
Power (W)	310	700	44%
Cost (¥10k/card)	~10	~18	56%

Conclusion:

• 910C inference speed is 60% of H100, but power only 44%
• In cost-sensitive scenarios, 910C's cost-performance advantage is clear
• For China-market localization needs, 910C is the only option

---

4. Market Positioning and Competitive Advantages​

4.1 Target markets​

Core markets:

1. Chinese government and SOEs: localization, data security, autonomy
2. Large-model startups: cost-sensitive, high compute demand
3. Operators and cloud providers: large-scale deployment, high efficiency requirements
4. Research and education: ultra-large-scale computing, talent development

Edge markets:

1. Autonomous driving: end-to-end large-model training
2. Smart healthcare: medical imaging, drug discovery
3. Fintech: risk control, robo-advisory

4.2 Competitive advantages​

Advantage	Description
Autonomy	SMIC N+2 process + Huawei in-house architecture, no supply-chain risk
Large memory	128GB HBM2E, full-pipeline training of hundred-billion-parameter models
High energy efficiency	310W TDP delivers 800 TFLOPS, close to H100 efficiency
System scaling	CloudMatrix 384 SuperNode, total compute exceeds GB200 NVL72
Software ecosystem	CANN CUDA-compatible, lower migration cost
Cost advantage	~¥100k/card, ~44% cheaper than H100

4.3 Weaknesses and improvement directions​

Weakness	Improvement direction
Single-chip compute	Next-gen 910D to adopt 3nm, target doubling
HBM bandwidth	950 series to adopt in-house HBM (HiBL 1.0), bandwidth to 4 TB/s
Software ecosystem	Continued CANN + MindSpore investment, expand developer community
Process	Deep cooperation with SMIC to ramp N+3 (5nm-class)

---

5. 2026 Shipment Plan and Market Forecast​

5.1 Shipment plan​

Period	Shipments	Cumulative	Key customers
2025 Q2-Q4	200k	200k	Huawei Cloud, China Telecom
2026 Q1-Q2	300k	500k	China Mobile, China Unicom, iFlytek
2026 Q3-Q4	300k	800k	Government projects, large-model startups
2027	1,000k	1,800k	Global market (Southeast Asia, Middle East, Latin America)

Capacity bottleneck:

• SMIC N+2 capacity ~100k wafers/month, Ascend 910C ~30% of that
• 2026 plan of 800k chips needs ~400k wafers, requiring 80%+ utilization
• Huawei prioritizes 910C capacity via deep SMIC cooperation

5.2 Market forecast​

China AI chip market (2026):

• Total: ~¥50B
• Domestic share: ~35% (¥17.5B)
• Ascend 910C share: ~60% (¥10.5B, ~800k chips)

Global AI chip market (2026):

• Total: ~$200B
• Huawei share: ~5% ($10B)
• Growth drivers: China-market localization + Belt and Road exports

---

6. Summary and Outlook​

6.1 Core conclusions​

1. Ascend 910C is a milestone domestic AI chip, with comprehensive breakthroughs in compute, memory, energy efficiency, and system scaling
2. CloudMatrix 384 SuperNode proves domestic chips can replace imported ones
3. DeepSeek-V4-Pro training success validates 910C maturity in ultra-large-model training
4. 800k chips shipped in 2026, projected 60% of China's AI chip market

6.2 Future outlook​

Short term (2026-2027):

• 910C continues ramping, shipments exceed 1,000k
• CloudMatrix 384 deployments over 1,000 units
• Software ecosystem (CANN + MindSpore) maturity approaches 70% of CUDA

Medium term (2028-2029):

• Next-gen 910D mass production, 3nm process, target 1.6 PFLOPS BF16
• 950 series (PR/DT) becomes inference-market mainstay, share over 30%
• 960/970 launch, N+3 process, supports trillion-parameter models

Long term (2030+):

• Huawei Ascend series becomes TOP 3 of the global AI chip market
• Domestic AI chips exceed 20% of the global market
• Transition from "following" to "running alongside" to "leading"

---

References​

1. Huawei Ascend 910C — Baidu Baike
2. Huawei Ascend series AI chip detailed parameter comparison (2025-2028) — EET-China
3. Huawei Ascend 910C compute cluster powers domestic chip's successful trillion-scale AI large-model training — QQ News
4. Huawei Ascend 910C completes DeepSeek V4 Pro training — Huxiu
5. Huawei Ascend 910C measured efficiency surpasses H100, AI Infra software-hardware co-optimization shines at ten-thousand-card cluster — CNBlogs

---

Last updated: June 10, 2026

Kunlun P800 is Baidu's third-generation AI accelerator from Kunlunxin Technology, based on the in-house XPU-P architecture, with 345 TFLOPS peak FP16 compute (surpassing NVIDIA H20's 148 TFLOPS). Launched in March 2024, it has become an important force among domestic AI training/inference accelerators.

This article comprehensively analyzes this domestic AI chip's breakthroughs across five dimensions: performance data, architectural innovation, SuperNode deployment, large-model adaptation, and market positioning.

---

1. Core Performance Data​

1.1 Compute performance​

Precision	Compute	Reference
FP16	345 TFLOPS	2.3× NVIDIA H20 (148 TFLOPS)
FP32	Not disclosed	Estimated ~170 TFLOPS
INT8	8-bit inference supported	Specific TOPS not disclosed
Low-power mode	128 TFLOPS @ 120W	Energy-efficiency-optimized scenarios
MoE optimization	Native MoE support	4.3× sparse-model inference efficiency

Performance characteristics:

• 345 TFLOPS at FP16, a new domestic AI chip compute benchmark
• 2.3× compute over NVIDIA H20 (H20 only 148 TFLOPS)
• Native MoE support, 4.3× sparse-model inference efficiency (with specific optimization)

1.2 Memory and bandwidth​

Item	Parameter
HBM type	HBM3e (3D-stacked memory)
Memory capacity	128 GB
Memory bandwidth	1.5 TB/s
ECC protection	End-to-end ECC supported

Memory advantages:

• 128GB capacity supports full-pipeline training of hundred-billion-parameter models
• 1.5 TB/s is a high-end configuration among HBM3e solutions
• 3D stacking alleviates large-model training memory bottlenecks

1.3 Power and energy efficiency​

Item	Parameter
TDP	400 W
Low-power mode	128 TFLOPS @ 120W
Energy efficiency (FP16)	~0.86 TFLOPS/W
vs. H100	~57% of H100 power (400W vs 700W)

Energy efficiency characteristics:

• At equal compute, significantly lower power than NVIDIA H100
• Dynamic power adjustment, auto-switching performance modes by load
• Suited to large-scale cluster deployment, reducing data center PUE pressure

1.4 Process and architecture​

Item	Parameter
Process	7nm
Transistors	Over 50 billion
Architecture	In-house XPU-P
Form factor	OAM module
Virtualization	Hardware vXPU, single card split into 32 virtual instances

Architectural innovation:

• Heterogeneous compute architecture, decoupling matrix-multiply units from tensor cores
• Parallel compute and data movement, theoretical compute 2.3× previous generation
• Hardware virtualization, single physical card divided into multiple logical cards, raising utilization

---

2. Three Architectural Innovations​

2.1 Heterogeneous compute architecture optimization​

Technology innovations:

• Matrix-multiply / tensor-core decoupling: parallelizes compute and data movement
• Dynamic task scheduling: auto-allocates compute by load
• Sparse compute optimization: native MoE support, 4.3× sparse-model inference efficiency

Performance gains:

• Theoretical compute 2.3× previous generation (Kunlun 2nd gen)
• 1.8× training throughput at equal power

2.2 3D-stacked memory technology​

Technology innovations:

• HBM3e memory with 3D stacking
• Single-card 128GB capacity, 1.5 TB/s bandwidth
• End-to-end ECC for data reliability

Performance gains:

• Alleviates large-model training memory bottleneck
• Supports full-pipeline training of hundred-billion-parameter models (no model-parallel splitting)
• 5× bandwidth vs GDDR6

2.3 Adaptive interconnect protocol​

Technology innovations:

• Dynamic die-to-die topology adjustment
• Built-in NPU for zero-copy data transfer, reducing CPU intervention
• ML-based congestion control, 30% lower packet loss than traditional ECN

Performance gains:

• In 256-node clusters, 40% lower communication latency
• Inter-chip bandwidth 1.2 TB/s (Tianchi 256-node)
• Smooth scaling to ten-thousand-card clusters

---

3. Tianchi SuperNode Deployment​

3.1 Tianchi 256-node​

System specs:

Item	Configuration
P800 chips per node	8
Inter-chip bandwidth	1.2 TB/s (40% over previous gen)
Max model parameters	500 billion
Typical power	12 kW
Interconnect	Hardware RDMA acceleration + dynamic traffic scheduling

Core technology breakthroughs:

1. Interconnect bandwidth engineering:

   • Built-in NPU for zero-copy data transfer, reducing CPU intervention
   • Dynamic traffic scheduling: auto-adjusts routes by real-time link quality
   • Predictive congestion control: ML-based congestion algorithm

2. Virtualization resource utilization:

Split	Actual perf	Theoretical	Utilization
1 card	100%	100%	100%
2 cards	185%	200%	92.5%
4 cards	340%	400%	85%

3.2 Tianchi 512-node​

System specs:

Item	Configuration
P800 chips per node	16
Inter-chip bandwidth	2.4 TB/s
Max model parameters	1.2 trillion
Typical power	24 kW
Recovery speed	Training resumes within 5 min of node failure

Core technology breakthroughs:

1. Ultra-large-scale training support:

   • Mixed-precision optimization: adds NF4 4-bit quantization on FP16/BF16, 75% less memory
   • Gradient checkpoint acceleration: reconstructs compute graph, activation storage O(n)→O(√n), 1.8× training speed
   • Failure recovery: distributed snapshot, 10× faster than traditional checkpoint

2. Communication efficiency optimization:

   • 3D parallelism (data + model + pipeline), compute/communication ratio 12:1
   • In 1.75-trillion-parameter MoE training, communication overhead below 15%

3.3 Tianchi series performance comparison​

Metric	Tianchi 256	Tianchi 512	Improvement
Max model parameters	500 billion	1.2 trillion	2.4×
Inter-chip bandwidth	1.2 TB/s	2.4 TB/s	2×
Typical power	12 kW	24 kW	2×
Recovery time	<5 min	<5 min	Flat
Latency reduction	40%	50%	10 pts

---

4. Large-Model Adaptation​

4.1 DeepSeek series adaptation​

Certification:

• February 2025: passed DeepSeek-V3/R1 671B adaptation certification
• Supports single-machine 8-card full DeepSeek-V3 671B
• Supports DeepSeek MoE full-parameter training with just 32 machines

Performance data (DeepSeek-V3 671B):

Metric	P800	NVIDIA H100	Ratio
Inference speed (tokens/s)	12,500	14,200	88%
Training throughput (samples/s)	8.5	10.2	83%
First-token latency (ms)	95	85	112%
Memory usage (GB)	118	72	164%

Conclusion:

• P800 reaches 88% of H100 inference speed, gap significantly narrowed
• 83% of H100 training throughput
• 128GB large memory advantage clear, supports larger batch sizes

4.2 Other large-model adaptation​

Model	Deployment	Notes
ERNIE series	Baidu Cloud native	Baidu Smart Cloud main deployment
LLaMA series	Supported	Includes MoE-distilled versions
Qwen series	Supported	Alibaba Cloud model adaptation
ChatGLM series	Supported	Zhipu AI model adaptation
Baichuan series	Supported	Baichuan Intelligent model adaptation

CUDA compatibility:

• Models runnable on CUDA migrate to P800 at low cost
• Supports open-source inference frameworks such as vLLM
• ~14% of CUDA low-level communication code needs rewriting (sparse-model inference needs specific optimization)

4.3 Ten-thousand-card cluster validation​

Cluster scale:

• Fully in-house 30,000-card cluster deployed
• Smooth scaling to ten-thousand-card clusters
• Linear scaling efficiency 85%+ (thousand-card scale)

Stability data:

• 30 days continuous training with no failures
• Training resumes within 5 min of node failure
• Cluster availability 99.9%

---

5. Performance Comparison Analysis​

5.1 vs. NVIDIA H20​

Item	Kunlun P800	NVIDIA H20	Notes
FP16 compute	345 TFLOPS	148 TFLOPS	P800 leads 2.3×
HBM capacity	128 GB	64 GB	P800 +100%
HBM bandwidth	1.5 TB/s	4.0 TB/s	H20 clear bandwidth lead
TDP	400 W	400 W	Flat
Process	7nm	4nm (TSMC)	H20 more advanced
Software ecosystem	XPU-P (CUDA-compatible)	CUDA	H20 more mature
Supply	China autonomous	Export-controlled	P800 no supply-chain risk

Conclusion:

• In FP16 compute, P800 leads H20 2.3×
• In memory capacity, P800 leads 100%
• In HBM bandwidth, H20 leads 2.67×
• In supply chain security, P800 wins outright

5.2 vs. NVIDIA H100​

Item	Kunlun P800	NVIDIA H100	Notes
FP16 compute	345 TFLOPS	~1,300 TFLOPS	H100 leads 3.77×
HBM capacity	128 GB	80 GB	P800 +60%
HBM bandwidth	1.5 TB/s	3.35 TB/s	H100 leads 2.23×
TDP	400 W	700 W	P800 only 57% of H100 power
Process	7nm	4nm (TSMC)	H100 more advanced
DeepSeek inference speed	12,500 tokens/s	14,200 tokens/s	P800 reaches 88% of H100

Conclusion:

• In raw compute, H100 leads P800 3.77×
• In energy efficiency, P800 clearly outperforms H100 (0.86 vs 1.86 TFLOPS/W)
• In actual inference performance, P800 reaches 88% of H100, gap significantly narrowed
• In cost, P800 is ~50% of H100

5.3 vs. Ascend 910C​

Item	Kunlun P800	Ascend 910C	Notes
FP16 compute	345 TFLOPS	800 TFLOPS	910C leads 2.32×
HBM capacity	128 GB	128 GB	Flat
HBM bandwidth	1.5 TB/s	784 GB/s	P800 leads 91%
TDP	400 W	310 W	910C lower power
Process	7nm	7nm (SMIC N+2)	Same
Software ecosystem	XPU-P (CUDA-compatible)	CANN (CUDA-compatible)	Each with strengths

Conclusion:

• In FP16 compute, 910C leads P800 2.32×
• In HBM bandwidth, P800 leads 910C 91%
• In software ecosystem, both CUDA-compatible, similar migration cost
• In scenarios, P800 suits inference, 910C suits training

---

6. Market Positioning and Competitive Advantages​

6.1 Target markets​

Core markets:

1. Baidu Smart Cloud: core compute base of the Baige platform
2. China Telecom/Mobile/Unicom: won AI inference server procurement bids
3. Large-model startups: cost-sensitive, high compute demand
4. Intelligent compute centers: ten-thousand-card clusters validated

Edge markets:

1. Autonomous driving: end-to-end large-model training
2. Smart finance: risk control, robo-advisory
3. Smart healthcare: medical imaging, drug discovery

6.2 Competitive advantages​

Advantage	Description
Compute leadership	FP16 345 TFLOPS, 2.3× over H20
Large memory	128GB HBM3e, full-pipeline training of hundred-billion-parameter models
High energy efficiency	400W TDP delivers 345 TFLOPS, better than H100
System scaling	Tianchi 256/512 SuperNodes, ten-thousand-card clusters
Software ecosystem	XPU-P CUDA-compatible, low migration cost
Cost advantage	~50% of H100, clear cost-performance edge
Supply chain security	China autonomous, no export-control risk

6.3 Weaknesses and improvement directions​

Weakness	Improvement direction
Single-chip compute	Next-gen M300 to adopt 5nm, target doubling
HBM bandwidth	M300 to adopt HBM4, bandwidth to 3.2 TB/s
Software ecosystem	Continued XPU-P + PaddlePaddle investment
Process	Deep cooperation with SMIC to ramp N+2 (7nm-class)

---

7. 2026 Shipment Plan and Market Forecast​

7.1 Shipment plan​

Period	Shipments	Cumulative	Key customers
2024 Q1-Q4	50k	50k	Baidu Smart Cloud
2025 Q1-Q4	150k	200k	China Mobile, China Telecom
2026 Q1-Q2	100k	300k	China Unicom, iFlytek
2026 Q3-Q4	100k	400k	Government projects, large-model startups
2027	500k	900k	Global market (Southeast Asia, Middle East, Latin America)

Capacity bottleneck:

• Constrained by wafer fab capacity, supply falls short of demand
• 2026 plan of 200k chips, actual capacity ~150k
• Kunlunxin deepening cooperation with SMIC and Hua Hong to raise capacity

7.2 Market forecast​

China AI chip market (2026):

• Total: ~¥50B
• Domestic share: ~35% (¥17.5B)
• Kunlun P800 share: ~20% (¥3.5B, ~200k chips)

Global AI chip market (2026):

• Total: ~$200B
• Kunlun share: ~1% ($2B)
• Growth drivers: China-market localization + Belt and Road exports

---

8. Summary and Outlook​

8.1 Core conclusions​

1. Kunlun P800 is a major domestic AI chip breakthrough, leading comprehensively in FP16 compute, memory capacity, and energy efficiency
2. Tianchi 256/512 SuperNodes prove domestic chips can replace imported ones
3. DeepSeek-V3 671B adaptation success validates P800 maturity in large-model training/inference
4. 200k chips shipped in 2026, projected 20% of China's AI chip market

8.2 Future outlook​

Short term (2026-2027):

• P800 continues ramping, shipments exceed 500k
• Tianchi 512-node deployments over 100 units
• Software ecosystem (XPU-P + PaddlePaddle) maturity approaches 60% of CUDA

Medium term (2028-2029):

• Next-gen M300 mass production, 5nm process, target 700 TFLOPS FP16
• M100 (inference-specific) becomes inference-market mainstay, share over 15%
• Supports trillion-parameter model full-pipeline training

Long term (2030+):

• Kunlun series becomes TOP 5 of the global AI chip market
• Domestic AI chips exceed 15% of the global market
• Transition from "following" to "running alongside"

---

References​

1. Kunlun P800 parameters — CSDN Library
2. Kunlun P800: technical breakthroughs and application prospects of a new-generation AI accelerator — YunTECH
3. Kunlun P800 latest specs: P800 single-precision compute reaches 345 TFLOPS — Xueqiu
4. Exclusive: Kunlun — domestic AI card full DeepSeek training/inference adaptation — Kunlunxin official
5. Kunlun P800 detailed specs — MirrorFrog: https://www.mirrorfrog.com/en/docs/cards/others/kunlun-p800

---

Last updated: June 10, 2026