Product Overview
The Kunlunxin P800 is the third-generation AI accelerator card from Kunlunxin Technology (a Baidu company). Based on the in-house XPU-P architecture, it delivers 345 TFLOPS of peak FP16 compute (surpassing the NVIDIA H20's 148 TFLOPS) at a TDP of about 400W, in the OAM module form factor; it launched in March 2024. It supports running the full-strength DeepSeek-V3/R1 671B on 8 cards in a single server, and multiple 10,000-card clusters have been delivered.
Key Positioning:
- Kunlunxin Gen 1 (2018): 14nm, deployed inside Baidu
- Kunlunxin Gen 2 (2021): 7nm, in-house Kunlun Core II, 256 INT8 TOPS
- Kunlunxin P800 (2024): XPU-P architecture, 345 TFLOPS FP16, OAM — this page
- Kunlunxin M100 (early 2026): inference-dedicated — existing page
- Kunlunxin M300 (early 2027): ultra-large-scale multimodal training
Core Specifications
| Parameter | Value |
|---|
| Architecture | In-house XPU-P architecture |
| Process | 7nm |
| Memory | 96 GB HBM3 |
| Memory Bandwidth | 2.4 TB/s |
| FP16 | 345 TFLOPS (surpasses the H20's 148 TFLOPS) |
| Low-power mode | 128 TFLOPS @ 120 W |
| INT8 | 820 TOPS (some reports cite 690–820 TOPS) |
| MoE Support | Native support for MoE architectures |
| TDP | 400 W |
| Form Factor | OAM module |
| Interconnect | XCCL (Kunlunxin interconnect), supports IB/ROCE |
| Release | 2024-03 |
| Mass Production | Launched in March 2024, large-scale delivery since 2025 |
| Cluster Scale | Supports 10,000-card clusters; an all-in-house 30,000-card cluster has been realized |
| Supernode | Tianchi 256 / Tianchi 512 |
| Supply Status | In short supply, constrained by foundry capacity |
Large-Model Adaptation
| Model | Deployment | Notes |
|---|
| DeepSeek-V3/R1 671B | 8 cards in a single server inference | Passed CAICT adaptation certification |
| DeepSeek MoE full-parameter training | 32 servers suffice | Supports MLA and multi-expert parallelism |
| ERNIE series | Native Baidu Cloud support | Main deployment on Baidu AI Cloud |
| Llama / Qwen / ChatGLM | Supported | Includes MoE distilled versions |
| Baichuan | Supported | Domestic model ecosystem |
CUDA compatibility: models that run on CUDA migrate to the P800 at low cost; open-source inference frameworks such as vLLM are supported.
| Parameter | Details |
|---|
| Company | Kunlunxin Technology (Beijing) Co., Ltd. |
| Parent Company | Baidu (57.67% stake) |
| Founded | April 2021 (spun off from Baidu) |
| P800 Launch | March 2024 |
| IPO Status | Started STAR Market IPO tutoring in May 2026 |
| Valuation | Over 10 billion RMB |
| Key Customers | Baidu AI Cloud, China Mobile (won the AI inference server centralized procurement) |
| Certification | CAICT five-star rating for "Stable Operation of Intelligent Computing Service Clusters" |
Use Cases
- ✅ Domestic large-model training (full-parameter training of DeepSeek, ERNIE, etc.)
- ✅ Large-model inference (671B on 8 cards in a single server)
- ✅ Baidu AI Cloud (core compute foundation of the Baige platform)
- ✅ Domestic intelligent computing centers (10,000-card clusters verified)
- ✅ MoE model inference (native hardware optimization)
- ❌ Deep CUDA ecosystem dependence (migration requires adaptation)
- ❌ Low-power edge deployment (400W TDP is high)
- ❌ International markets (restricted by export controls)
Key Timeline
| Date | Event |
|---|
| 2018 | Kunlunxin Gen 1 released (14nm) |
| 2021-04 | Kunlunxin Technology began independent operations |
| 2021 | Kunlunxin Gen 2 mass-produced (7nm Kunlun Core II) |
| 2024-03 | P800 officially launched (this page) |
| 2025-02 | Passed DeepSeek 671B adaptation certification |
| 2025 | Large-scale delivery of 10,000-card clusters |
| 2026-05 | Started STAR Market IPO |