Skip to main content

Cambricon Siyuan MLU

Vendor: Cambricon

Category: ASIC Dedicated Accelerator

Architecture: MLUarch

Introduction

Cambricon domestic AI accelerator cards. The Siyuan series covers both training and inference scenarios. Siyuan 590 (MLU590) uses 7nm Chiplet packaging, FP16 compute 256 TFLOPS, INT8 compute 512 TOPS, 96GB HBM2e. The proprietary Bangware software stack achieves CUDA compatibility, accessed via the NeuWare SDK. In 2025, approximately 116,000 units were shipped, placing it in the domestic third tier.

Specifications

ModelComputeMemoryInterfaceTDPProcess
Siyuan 590256 TFLOPS (FP16) / 512 (INT8) / 80 (FP32)96GB HBM2e (2.7 TB/s)OAM350W7nm
Siyuan 370128 TOPS (INT8)24GB LPDDR5PCIe 4.0150W12nm
Siyuan 27064 TOPS (INT8)8GB DDR4PCIe 3.070W16nm

Official Website

Visit Official Website

Driver Downloads

Linux

OS Support

WindowsLinuxmacOSAndroid

Version History

VersionRelease DateDescription
Neuware 5.02024590 support + CNN/Transformer dual optimization
BANG C 2.02023CUDA programming model compatibility

Performance Benchmarks

ModelTaskPerformance Metric
Siyuan 590BERT-Large Training~80% A100 efficiency (official data)
Siyuan 590INT8 Inference512 TOPS
Siyuan 370Inference ThroughputINT8 optimized
Siyuan 270Edge Inference64 TOPS, 70W

Pricing

ModelReference PriceNotes
Siyuan 590Contact vendorEnterprise training card
Siyuan 370Contact vendorInference accelerator
Siyuan 270Contact vendorEdge/endpoint

Quick Installation

Linux (Ubuntu 22.04 / Kylin)

# 1. Install NeuWare driver
sudo ./neuware-driver-*.run --install

# 2. Install NeuWare SDK
sudo ./neuware-sdk-*.run --install

# 3. Verify
cnmon

NeuWare is downloaded from the Cambricon Developer Community.

Code Examples

Python (Cambricon PyTorch)

import torch
import torch_cambricon # Cambricon NPU extension

# Use MLU backend
device = torch.device("mlu")
x = torch.randn(1024, 1024, device=device)
y = torch.matmul(x, x)
print(f"MLU matrix multiply: {y.shape}")

Bang C (Cambricon CUDA-compatible)

// Bang C API is highly compatible with CUDA
__mlu_entry__ void kernel(float *input, float *output) {
int idx = get_thread_id();
output[idx] = input[idx] * 2.0f;
}

Architecture Highlights

  • MLUarch: Cambricon proprietary AI accelerator architecture. 590 uses 7nm Chiplet packaging, 96GB HBM2e large memory
  • Bang C: CUDA-compatible programming language, enabling low-cost CUDA code migration
  • NeuWare: Complete software stack — driver + runtime (CNRT) + operator library (CNBL/CNCL) + compiler

Model Compatibility

Model/FrameworkSupportNotes
PyTorch✅ torch_cambriconMLU backend
Bang C✅ NativeCUDA-compatible
MindSpore⚠️Under adaptation
Llama / Qwen and similar LLMs⚠️Ecosystem gradually maturing
Model QuantizationINT8 optimized
Domestic OSKylin/UOS adapted

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