Alibaba PPU (Incang 800 Next Generation)#
Product Overview#
Alibaba PPU is a next-generation AI inference chip developed by Pingtouge (T-Head), Alibaba's semiconductor subsidiary. It is the next-generation product of Incang 800. Tape-out completed at TSMC in 2023, originally planned for 2024 mass production, but affected by US sanctions, production was transferred to domestic enterprises. As of September 2025, the official name has not been announced ("PPU" is a temporary project-phase identifier).
PPU outperforms NVIDIA H20 in memory capacity (96GB HBM2e) and PCIe 5.0 interface. Single-card BOM cost is 40% lower than H20, which can drive Alibaba Cloud public cloud inference instance prices down by 50%.
⚠️ Naming Note: PPU is a temporary project-phase identifier. The official commercial model name has not been announced. Industry sources speculate it is the next-generation product of Incang 800.
Core Specifications (Public)#
| Item | Parameter |
|---|---|
| Architecture | Pingtouge self-developed NPU architecture |
| Process | 7nm (estimated, transferred to domestic production after sanctions) |
| HBM | 96 GB HBM2e |
| Memory Bandwidth | 700 GB/s |
| TDP | 400W |
| Interface | PCIe 5.0 ×16 |
| Inter-chip Interconnect | 700 GB/s (specific protocol not disclosed) |
| Mass Production | Originally 2024, delayed due to sanctions |
| Commercial Availability | Testing in 2025, not yet announced |
📝 Data Note: Core metrics such as FP32/FP16/INT8 compute and high-speed interconnect technology have not been publicly disclosed. Actual performance requires subsequent empirical verification.
PPU vs NVIDIA H20 Comparison#
| Metric | Alibaba PPU | NVIDIA H20 | Comparison |
|---|---|---|---|
| HBM Capacity | 96GB HBM2e | 96GB HBM3 | Capacity comparable |
| Memory Bandwidth | 700 GB/s | >1 TB/s | H20 bandwidth lead |
| TDP | 400W | 400W | Comparable |
| Interface | PCIe 5.0 ×16 | PCIe 5.0 | PPU interface more advanced |
| Single-card Cost | -40% vs H20 | Baseline | PPU significant cost advantage |
| Ecosystem | Self-developed HALO stack | CUDA | H20 ecosystem mature |
| Comprehensive Performance | Close to H20 | Baseline | PPU is important breakthrough |
Deployment Status#
- China Unicom Sanjiangyuan Green Computing Power Center: Alibaba Cloud signed 1,024 devices, 16,384 Pingtouge compute cards, contributing 1,945 P compute (largest brand in project procurement)
- Alibaba Internal: Taobao search recommendation, Alibaba Cloud PAI platform, Cainiao logistics scheduling (Incang 800 already deployed, PPU will gradually replace)
Strategic Significance#
- Domestic substitution: Affected by US sanctions, transferred to domestic production, important breakthrough in domestic AI chip supply chain autonomy
- Cost advantage: Single-card BOM cost 40% lower than H20, driving Alibaba Cloud inference instance prices down 50%
- Performance breakthrough: Comprehensive performance close to H20, significant upgrade from Incang 800 (12nm, 820 TOPS)#
Manufacturer Information#
| Item | Content |
|---|---|
| Manufacturer | Alibaba Pingtouge Semiconductor (T-Head) |
| Product Page | https://www.t-head.cn/ |
| Software Stack | HALO (Hanguang Accelerated Linear Operator) |
| Ecosystem Positioning | Alibaba Cloud internal deployment + future external sales |
Related Products#
- Alibaba Incang 800 - Previous-generation chip (2019 release)#
- NVIDIA H20 - Primary competitor
- Huawei Ascend 910C - Same-period domestic AI chip
- Full comparison table