Skip to main content

Hygon DCU BW1000 (3rd-Gen Shensuan)

Hygon's third-generation flagship DCU (Shensuan No.3 / DCU-3, production model BW1000, series code DCU 8300) for large-model training, scientific computing, and government workloads. Deployed at scale on the Info High-Speed Rail (XinxigaoTie) AI computing platform — the first platform to schedule BW1000 clusters in production.


Core Specifications

SpecValue
ArchitectureGPGPU (CUDA-like, DTK/HIP software stack, Chiplet packaging)
ProcessUndisclosed
FP16480 TFLOPS (per Moark deployment docs; Hygon has not published official figures)
FP3260 TFLOPS
FP6430 TFLOPS (strength: scientific computing / CAE simulation)
Precision supportFP16 / BF16 / TF32 / FP32 / FP64 / INT8
FP8 supportNot supported (quantized alternatives required for FP8 models)
Memory64 GB HBM2e
Memory bandwidth1.6 TB/s
TDP~300 W (typical)
InterconnectPCIe 5.0 x16 (host interface)
LaunchReleased 2025, scaling through 2026
PriceUndisclosed

📌 Sourcing note (2026-09 cross-validation): FP16/FP32/FP64/memory/bandwidth/TDP figures are taken from the Moark (moark.com) deployment documentation, corroborated by Baidu Baike; both sources agree. Hygon has never officially published compute figures — media claims of 5nm process and 800 TFLOPS conflict with each other and are not adopted. Process node is left unlisted (only Chiplet packaging is confirmed).


Highlights

  • CUDA-like ecosystem: full DTK/HIP software stack compatible with ROCm/CUDA programming models; native support for PyTorch, TensorFlow, and PaddlePaddle with operator coverage over 99% of CUDA equivalents
  • Full-precision scientific computing: 30 TFLOPS FP64 is top-tier, excelling at CAE simulation, weather modeling, and computational fluid dynamics
  • Seamless migration: DeepSeek models adapted on launch day; Tencent Hunyuan, DeepSeek V3/R1, and Qwen3 fully supported
  • Cluster-ready: BW1000 clusters live on the Info High-Speed Rail platform, supporting 50+ industry large-model programs

Positioning

The BW1000 is Hygon's third-generation DCU flagship: Shensuan No.1 (V100-class) → Shensuan No.2 (A100-class, ~180 TFLOPS FP16 range) → Shensuan No.3 BW1000 (H800/H200 inference-subset class). Paired with Hygon C86 CPUs in a CPU+DCU heterogeneous architecture, it targets finance, government, and energy xinchuang (domestic substitution) deployments.


Use Cases

  • Large-model training and inference (hundred-billion-parameter scale; near-A100 LLM training performance)
  • Scientific computing and CAE simulation (FP64 strength; 256-card parallel runs show up to 700x speedup)
  • Government/finance xinchuang compute centers (Info High-Speed Rail cluster scheduling)
  • ❌ Direct FP8 model deployment (conversion required)

References