Hygon Shensuan DCU
Vendor: Hygon
Category: ASIC Dedicated Accelerator
Architecture: GPGPU (ROCm-compatible)
Introduction
Hygon Shensuan series DCU (GPGPU). The only domestic vendor with both x86 CPU and GPGPU DCU dual-chip products. Shensuan Z100/ZK2 series is compatible with ROCm/CUDA ecosystem, with FP64 double precision at approximately 10 TFLOPS, suitable for HPC+AI hybrid scenarios. Approximately 83,000 units shipped in 2025, market cap of 330 billion RMB.
Specifications
| Model | Compute | Memory | Interface | TDP | Process |
|---|---|---|---|---|---|
| Shensuan II (K100-AI) | 192 TFLOPS (FP16) / 49 (FP32) | 64GB HBM2 | OAM | 300W | 7nm |
| Shensuan I (Z100) | 21.6 TFLOPS (FP16) / 10.8 (FP32) | 32GB HBM2 | PCIe 4.0 | 250W | 7nm |
Official Website
Driver Downloads
Linux
Related Documentation
OS Support
| Windows | Linux | macOS | Android |
|---|---|---|---|
| ⚠️ (Partial) | ✅ | ❌ | ❌ |
Version History
| Version | Release Date | Description |
|---|---|---|
| DTK 3.0 | 2024 | Z200 support + CUDA compatibility optimization |
| DTK 2.5 | 2023 | PyTorch compatibility layer enhancement |
Performance Benchmarks
| Model | Task | Performance Metric |
|---|---|---|
| Shensuan II Z200 | BF16 Training | Near MI250X efficiency |
| Shensuan I Z100 | General AI Inference | Compatible with ROCm ecosystem |
Pricing
| Model | Reference Price | Notes |
|---|---|---|
| Z200 | Contact vendor | Domestic alternative to AMD Instinct |
| Z100 | Contact vendor | In mass production |
Quick Installation
Linux (Kylin / UOS / Ubuntu)
# Hygon DCU is compatible with ROCm ecosystem
# Install DCU driver (similar to ROCm)
sudo ./dcu-driver-*.run --install
# Verify
dcu-smi
Code Examples
Python (Hygon DCU PyTorch)
import torch
# Hygon DCU is compatible with ROCm/CUDA backend
assert torch.cuda.is_available()
print(f"DCU: {torch.cuda.get_device_name(0)}")
Architecture Highlights
- GPGPU Architecture: Based on AMD Zen architecture license + proprietary DCU design, compatible with ROCm ecosystem
- Domestic Substitution: Independent R&D based on AMD license, targeting the Xinchuang/domestic substitution market
- ROCm-like Ecosystem: Can directly use ROCm toolchain and PyTorch ROCm versions
Model Compatibility
| Model/Framework | Support | Notes |
|---|---|---|
| PyTorch | ✅ ROCm | ROCm backend compatible |
| Kylin/UOS | ✅ Native | Domestic OS adaptation |
| General AI | ⚠️ | Via ROCm ecosystem |
| Large Model Training | ⚠️ | Gradually improving |
Related Products
If you're evaluating alternatives, the following products may also fit your scenario:
- NVIDIA GPU / CUDA — NVIDIA (GPU Graphics Processor)
- AMD ROCm / GPU — AMD (GPU Graphics Processor)
- Iluvatar Tianguai 100 — Iluvatar (GPU Graphics Processor)
- Huawei Ascend — Huawei (NPU Neural Processor)
- Cambricon Siyuan MLU — Cambricon (ASIC Dedicated Accelerator)
- Enflame Blaze T20/T21 — Enflame (ASIC Dedicated Accelerator)
- Intel Gaudi (Habana) — Intel (ASIC Dedicated Accelerator)