Skip to main content

Baidu Kunlunxin P800

Vendor: Baidu

Category: GPU Graphics Processor

Architecture: XPU

Introduction​

Baidu Kunlunxin AI accelerator chip, incubated from Baidu. P800 series FP16 compute 345 TFLOPS, 2.3× that of NVIDIA H20. January 2026 launched STAR Market IPO, valuation expected at approximately $500 billion. Leveraging "application-driven" advantages and large-scale cluster technology, it has become a top-tier domestic AI chip. Approximately 116,000 units shipped in 2025.

Specifications​

ModelComputeMemoryInterfaceTDPProcess
Kunlunxin Gen 2256 TOPS (INT8) / 128 (FP16)32GB GDDR6 (512 GB/s)PCIe 4.0160W7nm
Kunlunxin Gen 3512 TOPS (INT8) / 256 (FP16)64GB HBM2eOAM400W5nm

Official Website​

Visit Official Website

Driver Downloads​

Linux​

OS Support​

WindowsLinuxmacOSAndroid
❌✅❌❌

Version History​

VersionRelease DateDescription
SDK 3.02024Gen 3 chip support + Paddle integration

Performance Benchmarks​

ModelTaskPerformance Metric
Kunlunxin Gen 3Llama 2 7B Inference~35 tok/s (INT8)
Kunlunxin Gen 2Paddle Model Inference~80% GPU efficiency
Kunlunxin Gen 3Natural Language Understanding (NLU)General AI inference

Pricing​

ModelReference PriceNotes
Kunlunxin Gen 3Contact vendorEnterprise customers
Kunlunxin Gen 2Contact vendorPrimarily via Baidu Cloud instances

Quick Installation​

Linux​

# 1. Install XPU driver and SDK
sudo rpm -ivh kunlun-driver-*.rpm
tar -xzf xpu-sdk-*.tar.gz && cd xpu-sdk && sudo ./install.sh

# 2. Verify
xpu-smi

Driver and SDK downloaded from Kunlunxin Official.

Code Examples​

Python (PaddlePaddle XPU)​

import paddle

# Check XPU availability
print(f"XPU available: {paddle.device.is_compiled_with_xpu()}")
paddle.set_device('xpu')

# Run simple model
x = paddle.randn([1024, 1024])
y = paddle.matmul(x, x)
print(f"XPU matrix multiply: {y.shape}")

Architecture Highlights​

  • XPU Architecture: Baidu proprietary AI accelerator architecture, purpose-built for deep learning, supporting training and inference
  • PaddlePilot Deep Integration: Native acceleration backend for Baidu PaddlePaddle
  • Kunlunxin Gen 3: Significantly improved compute and memory, supporting large model training scenarios

Model Compatibility​

Model/FrameworkSupportNotes
PaddlePaddle✅ NativeBest support
PyTorch⚠️Via XPU adapter plugin
PaddleOCR✅Officially recommended acceleration
Wenxin Large Models✅Native support
General Models⚠️Ecosystem expanding

If you're evaluating alternatives, the following products may also fit your scenario: