Skip to main content

4 posts tagged with "天池超节点"

百度天池超节点解决方案

View all tags

阿里磐久 AL128 超节点:平头哥真武 M890 撑起的 128 卡统一算力域

· 3 min read
Industry Research Team

中信建投 10 月 21 日研报把阿里磐久 AL128/AL64/AL144 列入国内外超节点第一梯队——与英伟达 NVL 系、华为 Atlas 950/960 并列。阿里的自研算力版图,终于从芯片延伸到了整机柜。

一、产品矩阵​

型号规模要点
磐久 AL128单柜 128 卡真武 M890 + ICN Switch 互联芯片
磐久 AL144单柜 144 卡对标业界最高密度档位
磐久 AL64单柜 64 卡中小规模训推

超节点的核心卖点与大同行一致:扩大 ScaleUp 互联域、统一内存编址、提升柜内互联效率——把高频 All-to-All 通信留在柜内完成,突破传统八卡服务器跨节点通信的瓶颈。

二、自研闭环的完成​

磐久超节点的意义在于阿里算力栈的完整度:

  • 芯片层:平头哥真武(Zhenwu)M890 计算芯片
  • 互联层:自研 ICN Switch 交换芯片
  • 整机层:磐久系列整机柜(液冷/供电一体化)
  • 模型层:Qwen 系列自用需求托底
  • 云层:阿里云对外分发

这五层与谷歌「TPU + 云」的纵向一体化同构——庞大自用需求托底,是自研算力能持续迭代的前提。

三、2026:国产超节点元年​

研报判断:随着头部云厂商与行业客户开始部署,2026 年是国产超节点放量元年,渗透率进入快速提升通道。受益链条依次为:服务器整机、交换芯片、交换机、连接器,随后是 PCB、液冷、电源、CPU 与国产算力芯片。

各家单柜规模的一览(详见工博会超节点观察):

  • 华为 Atlas 950/960(NPO 光互联、4096 卡扩展)
  • 阿里磐久 AL128/AL144
  • 天数智芯 IXL144(144 GPU/柜,10.1TB 显存)
  • 沐曦曦景 S600(64 卡,三零线缆)
  • 昆仑芯 32/64 卡(率先量产交付)
  • 燧原 ESL64-O(正交架构,与中兴合作)

四、竞争维度的迁移​

国产算力竞争已经从「单卡跑分」全面转向「整机、集群、软硬件一体化的系统能力竞争」。单卡 FP16 数字在这个维度上只是入场券——柜内互联、内存语义统一、机房友好度、软件栈完整度才是决胜项。

延伸阅读:超节点时代的选型框架 · 超节点放量元年解读 · 完整对比表

昆仑芯 M100 部署进展与百度自研芯片路线

· 7 min read
Industry Research Team

2026 年 6 月 12 日,在智源大会上昆仑芯研发副总裁漆维透露,面向推理市场的 M100 芯片目前已交付大客户进行测试,并已向客户小规模出货。与寒武纪 MLU690 迟迟无法交付形成对照,昆仑芯的节奏明显更靠前——尽管它面临的最大问题是需求侧而非供给侧。

M100 的定位:只做推理​

昆仑芯 M100 是昆仑芯科技 2026 年推出的第四代 AI 芯片产品,延续自研 XPU 架构理念,主要针对大规模推理场景(尤其是 MoE 模型)进行优化设计,旨在提升此类模型的推理性能并提供较高的性价比。

这颗芯片的发布时间可以追溯到 2025 年 11 月 13 日的百度世界大会。百度集团执行副总裁沈抖透露将发布新一代昆仑芯,其中包括面向大规模推理的 M100 和面向超大规模训练与推理的 M300,计划分别于 2026 年年初和 2027 年年初上市。2026 年 1 月 27 日,百度 CEO 李彦宏在接受《时代》杂志专访时评论芯片行业,表示百度刚刚发布了新款 M100 芯片,并正在开发新的芯片集群。

在百度 2026 年的业绩会上,李彦宏进一步披露昆仑芯已经完成三代 AI 芯片的研发与商业化,并公布了 M100 与 M300 的产品路线图。

维度昆仑芯 M100昆仑芯 M300
上市时间2026 年年初2027 年年初
目标场景大规模推理,尤其是 MoE 模型超大规模多模态模型训练与推理
部署状态2026 年 6 月已交付大客户测试并小规模出货尚未发布
配套超节点天池 256 / 天池 512规划中

值得注意的是,2026 年 7 月百度昆仑芯科技自主研发新品 M100 芯片实物首次在央视报道中公开亮相——这类"实物亮相"的报道时点,通常意味着产品已进入可交付状态。

天池超节点:从千柜级部署到开放规范​

M100 的商业化依托的是天池超节点。据公开报道,天池 256 与天池 512 超节点在 2025 年百度世界大会发布:最高支持 256 及 512 卡极速互联,卡间互联总带宽分别提升 4 倍及 1 倍,其中天池 256 主流大模型推理任务单卡 tokens 吞吐提升 3.5 倍,天池 512 靠单节点即可完成万亿参数模型训练。

这两款超节点计划分别于 2026 年上半年及下半年推出。天池超节点覆盖 32 至 512 卡全场景,采用标准机柜形态、三总线盲插与模块化设计,兼容液冷与风冷 IDC。

更重要的进展发生在 2026 年 9 月 16 日。据报道,百度智能云在 2026 智能经济论坛上宣布天池超节点已实现全栈国产,并已在多地智算中心进行千柜级规模部署,支撑万亿参数大模型训练;同时宣布全面开放天池超节点的架构、供电、散热、节点、互联和管理六大核心模块技术规范,助力国产芯片更快完成从单卡验证到超节点级系统适配。

这份规范开放的意味值得深挖:它与昇腾推动 NPO 成为 OIF 标准、摩尔线程联合中国移动研究院与之江实验室发布《OISA 高密超节点参考设计技术规范》指向同一件事——超节点的技术规范正在成为国产厂商争夺的新战场。

百度的芯片路线:五年五芯与集群梯度​

百度在芯片路线上的公开表述相当清晰。第一是产品化节奏,即"五年五芯"计划,推动每年推出新版芯片,把芯片开发模式从项目制转向产品化。第二是场景化细分策略:M100 专注提升大规模推理任务的性价比,M300 针对超大规模多模态模型优化。

第三是集群规模的梯度规划。据公开报道,百度智能云计划将昆仑芯单一集群规模从三万卡扩展至百万卡级别;另有报道称百度正在研发显卡级超级节点,预计 2027 年下半年开始交付千卡至四千卡规模的计算集群,并已有四千卡超节点的规划。

现有的部署规模方面,据公开报道,2026 年上半年百度点亮了昆仑芯 P800 的 3 万卡机,这是国产算力发展的重要里程碑;昆仑芯累计已完成数万卡部署。外部落地方面,昆仑芯已进入招商银行大模型算力项目,并在国家电网、中钢研完成部署,且已构建覆盖互联网、能源、金融、制造、汽车、通信、教育等行业的全栈昆仑芯集群。

在安全认证上,据公开报道,昆仑芯 M100 在最新一轮安全可靠测评中获得 Level I 评级,同期华为昇腾 950 获得 Level II 评级,龙芯与海思鲲鹏的通用处理器则获得最高等级的 Level III 评级。该三级体系正在塑造政府采购的实质性导向。需要注意,测评结果不等于性能结论——官方报道也指出 M100 较低的评级意味着它在满足该框架最严苛基准上仍有距离。

研判​

昆仑芯当前的处境与华为、寒武纪、摩尔线程都不同:它不缺乏产能,也不缺乏标杆模型,它缺的是足够大的外部需求。据公开报道的行业调研,昆仑芯自用需求 2026 年无明显提升,P800 和 M100 需求趋于饱和,更多是借 AI 软件影响力做生态推广;对外主要靠推理业务与百度软件绑定,面向政府、教育、政务等传统政企行业,侧重点在生态渗透而非自身业务放量。

这解释了百度为何要开放天池超节点的六大核心模块技术规范——当自家消纳能力见顶,向外输出系统规范与工具链就成了打开市场的方式。而安全可靠测评中 M100 仅获 Level I,也说明政企市场的主流通道目前更偏向评级更高的昇腾。

昆仑芯的战略价值因此不在短期营收,而在于它把国产自研芯片的"第三条路"走通了:不与华为拼超节点规模,也不与寒武纪拼产能,而是绑定云平台做算力输出。如果 M100 在推理侧把性价比做实,并借天池规范形成生态飞轮,它将在国产算力市场占据一个与华为错位的位置——用推理性价比与云服务捆绑,而不是用训练规模与系统架构正面对抗。M300 在 2027 年的表现,才是这条路线的最终答卷。

昇腾960 超节点:全球首个 NPO 超节点的 4096 卡野心

· 6 min read
Industry Research Team

2026 年 9 月 17 日,华为全联接大会(HC2026)上海站,华为轮值董事长汪涛官宣昇腾960 超节点——全球首个采用 NPO(近封装光学)技术的超节点。4096 卡、8 EFLOPS FP8、1PB HBM,加上提前三个季度就绪的 960DT,华为把"系统级算力"的赌注又加大了一档。

一、HC2026 发布:单节点 4096 卡意味着什么​

先看官方口径的核心数字:

指标昇腾 950 超节点昇腾 960 超节点
单节点卡数1024 卡(WAIC 2026 展出)4096 卡
FP8 算力8 EFLOPS(8192 卡满配)8 EFLOPS(单节点)
FP4 算力16 EFLOPS(8192 卡满配)16 EFLOPS(单节点)
HBM 总容量1152 TB(8192 卡满配)1 PB(4096 卡)
互联 RTT3 μs2 μs(全域 D2D)
光互联灵衢 2.0 总线NPO 光引擎 Hi-ONE

一个直观的对比:950 超节点需要 8192 卡满配才能达到的 8 EFLOPS FP8,960 超节点在单节点 4096 卡内即已实现——这背后是单卡规格翻倍与光互联密度提升的叠加。芯片侧,昇腾960DT 升级至 Da Vinci v6 架构:FP8 2 PFLOPS、FP4 4 PFLOPS、288GB 自研 HBM、9.6 TB/s 带宽,各项指标较 950 系列全部翻倍,华为称这一路径为"韬定律"。双版本节奏同样清晰:960DT 配风冷超节点 Atlas 860,960PR 配液冷超节点 Atlas 960,分别于 2027 年 Q2、Q3 上市。

二、NPO 光互连:一场功耗与密度的革命​

超节点规模越大,互联越是命门。传统可插拔光模块的功耗与密度,正是把超节点从"千卡"推向"万卡"的最大障碍。昇腾960 超节点的答案是 NPO(Near-Package Optics,近封装光学):

  • 5500 个 Hi-ONE 光引擎替代 48000 颗 800G 光模块,单引擎传输容量 7.2T,是业界首个量产、当前传输能力最大、唯一内置光源的 NPO 产品;
  • 系统功耗降低超 550 kW——相当于省出一个中型数据机房的总功率;
  • 无故障运行时间提升一倍,整体可用度达 99.8%。

华为选择 NPO 而非更激进的 CPO,理由是工程务实:NPO 在保留光引擎独立性的同时实现光电近距离握手,规避了 CPO 在可靠性、可制造性、可维护性上的难题,并延续现有光模块产业生态。

对照海外路线,这条路径的差异非常明显:NVIDIA 的 NVL72 系列在机柜内坚持铜互连(NVLink 铜背板),AMD 亦以铜互连为主、借助 UALink 开放生态扩展机柜间网络。铜互连在 72–144 卡尺度上功耗与成本占优,但要再往上走一个数量级,光进铜退几乎是必然。昇腾960 用 NPO 直接把光互联压进超节点内部,是在用互联架构换规模上限——这也是 4096 卡得以成立的前提。华为马尔科夫实验室的仿真给出了一组佐证:4K 超节点组成的 10 万卡集群,相比 8 卡服务器组网,MFU 提升达 2.75 倍——传统架构下,集群内通信要吃掉超过 40% 的训练时间。

三、4096 卡 vs NVL72:两种规模哲学​

把昇腾960 超节点与 NVIDIA NVL72 放在一起,看到的不是同一张设计图的两个版本,而是两种哲学:

维度NVIDIA NVL72昇腾960 超节点
单节点规模72 卡4096 卡
互联介质铜背板(NVLink)NPO 光引擎(Hi-ONE)
设计逻辑小域高带宽、机柜间以太网扩展大域统一内存、集群级扩展

NVL72 的思路是把最紧密的互联限制在 72 卡的小域内,用成熟的铜互连控制功耗,再靠网络横向扩展;昇腾960 则把"超节点"的边界直接推到 4096 卡,用统一内存编址和 2 μs 全域 RTT 让整个节点在逻辑上"像一台机器"。单卡绝对性能仍落后 NVIDIA 旗舰一代的背景下,华为选择用更大的 Scale-Up 域弥补代差——评估国产算力,不能只比单芯片 PFLOPS。

四、提前三个季度:执行信号与一年一代路线图​

比规格更值得注意的是节奏:昇腾960DT 较原计划提前三个季度就绪,960PR 提前一个季度。在先进制程受限的环境里,提前就绪意味着供应链、封装、整机调测的每一个环节都被压缩过——这是执行力的信号,也是对客户交付承诺的背书。

路线图同步浮出水面:2028 年昇腾970、2029 年昇腾980,"一年一代"的迭代节奏与算力规格持续翻倍的韬定律并列官宣。而 960 的部署底盘已经就位——昇腾910C 超节点已部署超 1000 套,950 超节点规模商用,CANN 外部开发者占比升至 61%、月活开发者突破 5200 人。

配套系统同步升级:鲲鹏 TaiShan 950 超节点支持全光组网下 4096 节点、256TB 统一内存池;OceanStor M900 上下文内存存储基于灵衢总线一跳直连 PB 级 KV Cache,把推理时代的上下文瓶颈纳入系统级解法。集群扩展路径上,多个 960 超节点经灵衢网络或 RoCE 互联可达 51.2 万卡,叠加多轨道拓扑后支持 100 万卡集群。

小结​

昇腾960 超节点的故事有三层:工程层面,NPO 光互连用 5500 个光引擎换下 4.8 万颗光模块、省电 550 kW,把光互联提前压进超节点内部,走出了与 NVIDIA 铜互连不同的路线;架构层面,4096 卡对 72 卡,是以更大的 Scale-Up 域对冲单卡代差的规模哲学;执行层面,960DT 提前三个季度就绪,让"一年一代"从路线图变成了可以兑现的承诺。2027 年 Q2 上市在即,真正的考验将是:这 4096 卡的野心,能否按期变成机房里实际运转的算力。

Kunlun P800 Deep Dive: Performance Data, Architectural Innovation, and SuperNode Deployment

· 10 min read
Industry Research Team

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​

PrecisionComputeReference
FP16345 TFLOPS2.3× NVIDIA H20 (148 TFLOPS)
FP32Not disclosedEstimated ~170 TFLOPS
INT88-bit inference supportedSpecific TOPS not disclosed
Low-power mode128 TFLOPS @ 120WEnergy-efficiency-optimized scenarios
MoE optimizationNative MoE support4.3× 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​

ItemParameter
HBM typeHBM3e (3D-stacked memory)
Memory capacity128 GB
Memory bandwidth1.5 TB/s
ECC protectionEnd-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​

ItemParameter
TDP400 W
Low-power mode128 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​

ItemParameter
Process7nm
TransistorsOver 50 billion
ArchitectureIn-house XPU-P
Form factorOAM module
VirtualizationHardware 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:

ItemConfiguration
P800 chips per node8
Inter-chip bandwidth1.2 TB/s (40% over previous gen)
Max model parameters500 billion
Typical power12 kW
InterconnectHardware 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:

SplitActual perfTheoreticalUtilization
1 card100%100%100%
2 cards185%200%92.5%
4 cards340%400%85%

3.2 Tianchi 512-node​

System specs:

ItemConfiguration
P800 chips per node16
Inter-chip bandwidth2.4 TB/s
Max model parameters1.2 trillion
Typical power24 kW
Recovery speedTraining 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​

MetricTianchi 256Tianchi 512Improvement
Max model parameters500 billion1.2 trillion2.4×
Inter-chip bandwidth1.2 TB/s2.4 TB/s2×
Typical power12 kW24 kW2×
Recovery time<5 min<5 minFlat
Latency reduction40%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):

MetricP800NVIDIA H100Ratio
Inference speed (tokens/s)12,50014,20088%
Training throughput (samples/s)8.510.283%
First-token latency (ms)9585112%
Memory usage (GB)11872164%

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​

ModelDeploymentNotes
ERNIE seriesBaidu Cloud nativeBaidu Smart Cloud main deployment
LLaMA seriesSupportedIncludes MoE-distilled versions
Qwen seriesSupportedAlibaba Cloud model adaptation
ChatGLM seriesSupportedZhipu AI model adaptation
Baichuan seriesSupportedBaichuan 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​

ItemKunlun P800NVIDIA H20Notes
FP16 compute345 TFLOPS148 TFLOPSP800 leads 2.3×
HBM capacity128 GB64 GBP800 +100%
HBM bandwidth1.5 TB/s4.0 TB/sH20 clear bandwidth lead
TDP400 W400 WFlat
Process7nm4nm (TSMC)H20 more advanced
Software ecosystemXPU-P (CUDA-compatible)CUDAH20 more mature
SupplyChina autonomousExport-controlledP800 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​

ItemKunlun P800NVIDIA H100Notes
FP16 compute345 TFLOPS~1,300 TFLOPSH100 leads 3.77×
HBM capacity128 GB80 GBP800 +60%
HBM bandwidth1.5 TB/s3.35 TB/sH100 leads 2.23×
TDP400 W700 WP800 only 57% of H100 power
Process7nm4nm (TSMC)H100 more advanced
DeepSeek inference speed12,500 tokens/s14,200 tokens/sP800 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​

ItemKunlun P800Ascend 910CNotes
FP16 compute345 TFLOPS800 TFLOPS910C leads 2.32×
HBM capacity128 GB128 GBFlat
HBM bandwidth1.5 TB/s784 GB/sP800 leads 91%
TDP400 W310 W910C lower power
Process7nm7nm (SMIC N+2)Same
Software ecosystemXPU-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​

AdvantageDescription
Compute leadershipFP16 345 TFLOPS, 2.3× over H20
Large memory128GB HBM3e, full-pipeline training of hundred-billion-parameter models
High energy efficiency400W TDP delivers 345 TFLOPS, better than H100
System scalingTianchi 256/512 SuperNodes, ten-thousand-card clusters
Software ecosystemXPU-P CUDA-compatible, low migration cost
Cost advantage~50% of H100, clear cost-performance edge
Supply chain securityChina autonomous, no export-control risk

6.3 Weaknesses and improvement directions​

WeaknessImprovement direction
Single-chip computeNext-gen M300 to adopt 5nm, target doubling
HBM bandwidthM300 to adopt HBM4, bandwidth to 3.2 TB/s
Software ecosystemContinued XPU-P + PaddlePaddle investment
ProcessDeep cooperation with SMIC to ramp N+2 (7nm-class)

7. 2026 Shipment Plan and Market Forecast​

7.1 Shipment plan​

PeriodShipmentsCumulativeKey customers
2024 Q1-Q450k50kBaidu Smart Cloud
2025 Q1-Q4150k200kChina Mobile, China Telecom
2026 Q1-Q2100k300kChina Unicom, iFlytek
2026 Q3-Q4100k400kGovernment projects, large-model startups
2027500k900kGlobal 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