by HygonCN
Give Chinese data centres a home-grown accelerator that runs CUDA-style AI software with little porting, reducing dependence on restricted Nvidia chips.
Deep Computing 3 runs Tencent's 295B Hunyuan model (May 2026); Deep Computing 4 still in development.
Updated 9 May 2026Checked 11 Oct0 updates this week
Like a GPU, a DCU splits big matrix maths across many small cores working at once, the shape of both neural networks and physics simulations.
Stacked HBM next to the processor feeds the cores fast enough; for large AI models, memory speed matters as much as raw compute.
Hygon's DTK mirrors AMD's ROCm, whose HIP language closely tracks CUDA, so frameworks like PyTorch and existing GPU code move over cheaply.
Hygon's own scale-up switch and ScaleFabric network link many DCUs, and Sugon builds them into dense supernode servers.
| Spec | Hygon DCU accelerators |
|---|---|
| Deep Computing 2 vs gen 1 | Over 100% fasterR (reported) |
| Number formats | FP64, FP16, BF16R (reported) |
| Large model adapted (May 2026) | 295B parameters (Tencent Hy3 preview)R (reported) |
| ScaleFabric network | 400 Gb/s, 0.93 us latencyR (reported) |
R reported by the company
Hygon
x86 server CPUs, DCU AI accelerators
Hygon Information Technology is China's main maker of x86-compatible server CPUs, based on technology licensed from AMD in 2016, and of DCU accelerators for AI and scientific computing. It listed on Shanghai's STAR Market in 2022.
