Skip to main content

Hygon DCU Z100

Product Overview​

The Hygon DCU Z100 is a GPGPU accelerator card in Hygon Information's DCU series (of the "Shensuan No. 1" generation), based on deep customization under AMD CDNA technology licensing (architecture code gfx906, first-generation CDNA), and is the core of Hygon's "CPU + DCU" heterogeneous computing solution. Descended from the AMD ecosystem, the Z100 can be natively compatible with the ROCm open computing platform, and through Hygon's proprietary DTK (DCU Toolkit) enables low-cost migration of CUDA applications, making it one of the most mature domestic "CUDA-like" ecosystem accelerator cards.

According to Peking University's high-performance computing system procurement award announcement, the Z100 has 8192 general-purpose compute cores, 32GB HBM2 memory, and FP64 compute of 10.8 TFLOPS, with hardware parameters broadly at the same generation level as the NVIDIA A100 and AMD MI100. The Z100 and the trimmed Z100L belong to the same Shensuan No. 1 product family, widely deployed in national supercomputing centers (such as Kunshan, Chengdu, and Zhengzhou) and in telecom operators' and banks' intelligent computing centers.

Core Specifications​

ParameterValue
ArchitectureGPGPU, deeply customized based on AMD CDNA licensing (gfx906 / first-generation CDNA)
Process Node7nm (estimated, based on public information for the same-generation Z100L)
Compute Cores8192 general-purpose compute cores (Peking University award announcement)
FP64 Compute10.8 TFLOPS (Peking University HPC system award announcement, authoritative source)
FP32 ComputeNot disclosed (estimated ~10.8 TFLOPS, on par with FP64, a CDNA-like architecture trait)
FP16 / BF16 ComputeNot disclosed (estimated ~21.6 TFLOPS)
INT8 ComputeNot disclosed (estimated ~43.2 TOPS)
Memory Capacity32 GB
Memory TypeHBM2
Memory BandwidthNot disclosed (Z100L is 1024 GB/s; the Z100 should be higher, estimated ~1.2 TB/s)
TDP~250 W (undisclosed, estimated from product materials)
InterconnectxHMI interconnect (DCU series, bandwidth ~184 GB/s); PCIe 4.0
InterfacePCIe 4.0 ×16
Launch2022 (Shensuan No. 1)
Mass Production/AvailabilityIn mass production

⚠️ Specification notes: FP64 10.8 TFLOPS / 8192 cores / 32GB HBM2 come from Peking University's high-performance computing system procurement award announcement (HCZB-2021-ZB0364), an authoritative public source. FP32/FP16/INT8 compute, memory bandwidth, and TDP are not listed separately by the vendor; the estimates in the table derive from industry comparison materials and same-generation Z100L parameters, and are marked as estimates, for reference only. Hygon DCUs excel at full precision, with double-precision (FP64) performance benchmarked against the A100/MI100.

Key Features​

  • CUDA-like ecosystem: based on the ROCm open ecosystem, with the DTK toolchain supporting automatic HIP/CUDA migration
  • Full-precision computing: FP64/FP32/FP16/INT8 full-precision capability, with double precision especially strong
  • x86 synergy: paired with Hygon C86 series CPUs to form a fully domestic "CPU+DCU" heterogeneous system
  • Large model adaptation: already adapted to mainstream large models such as LLaMa, GLM, Tongyi Qianwen, and DeepSeek

Vendor Information​

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

Use Cases​

  • ✅ High-performance computing (HPC) / scientific computing (FP64 double-precision advantage)
  • ✅ Large model training and inference (low ROCm/CUDA migration cost)
  • ✅ National supercomputing / intelligent computing centers (domestic, self-controlled)
  • ✅ Finance / energy / telecom operator data centers
  • ❌ Ultra-low-power edge scenarios (high TDP, data center positioning)
  • ❌ Scenarios with strong native CUDA dependence unwilling to migrate

References​