Skip to main content

Hygon DCU Z100L

Product Overview​

The Hygon DCU Z100L is the trimmed GPGPU accelerator card in Hygon Information's DCU series, of the same "Shensuan No. 1" generation as the Z100, based on deep customization under AMD CDNA licensing (gfx906). Compared with the Z100, the Z100L is reduced in compute unit count, compute, and memory bandwidth, but retains the 32GB HBM2 large memory and the software ecosystem advantage of ROCm/DTK CUDA compatibility, targeting cost-effective HPC and small/medium model training and inference scenarios, and is one of the standard GPU options for Xinchuang servers (paired with Hygon C86 CPUs).

The Z100L is commonly seen in national supercomputing centers (Chengdu, Xi'an, etc.), telecom operator and finance industry domestic substitution projects, and ranks alongside the Iluvatar CoreX BI-V100 as a mainstream Chinese GPGPU choice.

Core Specifications​

ParameterValue
ArchitectureGPGPU, deeply customized based on AMD CDNA licensing (gfx906 / first-generation CDNA)
Process Node7nm
Compute Units60 compute units / 3840 compute cores (Baidu Baike)
FP64 Compute10.1 TFLOPS
FP32 Compute10.1 – 12.2 TFLOPS
FP16 / BF16 Compute20.2 – 24.5 TFLOPS
INT8 Compute40.5 – 49.1 TOPS
Memory Capacity32 GB
Memory TypeHBM2
Memory Bandwidth1024 GB/s (1 TB/s)
TDP250 W / 280 W (varies by source)
InterconnectxHMI interconnect (~184 GB/s); PCIe 4.0
InterfacePCIe 4.0 ×16
Cooling Form FactorAir cooling / cold-plate liquid cooling / immersion liquid cooling
Launch2022 (Shensuan No. 1)
Mass Production/AvailabilityIn mass production

⚠️ Specification notes: The Z100L's FP64 10.1 / FP32 10.1–12.2 / FP16 20.2–24.5 / INT8 40.5–49.1 TFLOPS, 32GB HBM2, 1024 GB/s, and xHMI 184 GB/s come from Hygon product materials and industry comparison tests (including a publicly available Z100L specification PDF); the compute unit count follows Baidu Baike's "60 CUs / 3840 cores" figure (some sources record 64 CUs; Baidu Baike prevails). TDP has two reported values, 250W and 280W; both are listed.

Key Features​

  • Large-memory cost-effectiveness: 32GB HBM2 + 1 TB/s bandwidth, suited to medium/large models and HPC
  • CUDA-like ecosystem: DTK (based on ROCm) is compatible with HIP/CUDA, with low migration cost
  • Full-precision capability: FP64/FP32/FP16/INT8 full precision, with double precision suited to scientific computing
  • Multiple cooling options: supports air cooling, cold-plate liquid cooling, and immersion liquid cooling for different data centers

Vendor Information​

ItemDetails
CompanyHygon Information Technology Co., Ltd.
HeadquartersTianjin
Founded2014
IPOSTAR Market 688041
Technology SourceAMD x86 / CDNA licensing + proprietary DCU architecture

Use Cases​

  • ✅ Scientific computing / HPC (FP64 double precision, benchmarked on some metrics against the AMD MI50/MI60 and NVIDIA V100)
  • ✅ Small/medium model training and inference (cost-effective scenarios)
  • ✅ Standard GPU for Xinchuang servers (Hygon C86 CPU + DCU systems)
  • ✅ Finance / telecom operator domestic substitution
  • ❌ Ultra-large model training (32GB memory, limited compute)
  • ❌ Ultra-low-power edge scenarios

References​