Chinese chipmaker Hygon has announced a new lineup of hardware designed for the data center and artificial intelligence markets. The launch features a high-core-count CPU, a dedicated AI accelerator, and comprehensive networking infrastructure, representing an effort to boost China’s technological independence and compete directly with established industry leaders like Intel and NVIDIA.

Next-Generation C86-5G Processors

Moving away from its consumer-focused predecessor (the C86-4G), the new Hygon C86-5G CPU family targets high-density server environments. Built on an updated microarchitecture, the processors deliver a claimed Instructions Per Cycle (IPC) improvement of over 15%. A key feature is the integration of Simultaneous Multithreading 4 (SMT4), which enables each CPU core to process four threads concurrently.

The flagship model in this family features 128 cores and 512 threads, achieving up to 10 TFLOPs of FP64 computing performance. It also offers 104 PCIe 5.0 lanes. To compete effectively against Intel’s Xeon 6 processors, Hygon has integrated AVX512 instruction support alongside native AI acceleration for INT8 and BF16 data formats. These CPUs are already in mass production, powering hardware ranging from standard rack servers to advanced liquid-immersion cooling cabinets that can host up to 80,000 cores per cluster.

AI Accelerators and Proprietary Infrastructure

Alongside the CPUs, Hygon introduced its new Deep Computing Unit (DCU) GPU accelerator. Designed for AI training and High-Performance Computing (HPC), this GPGPU architecture supports FP64, FP16, and BF16 precision formats. Equipped with High Bandwidth Memory (HBM) and high-speed interconnects, the DCU is positioned to compete directly with NVIDIA’s Ampere-based A100 GPU.

To complete its ecosystem, Hygon developed four infrastructure chips, including a 104-lane PCIe 5.0 switch and a Scale-Up Interconnect Switch designed to function like NVIDIA’s NVLink. Furthermore, its ScaleFabric networking line offers 400 Gb/s speeds with 0.93-microsecond latency, alongside switches supporting up to 800 Gb/s ports, aiming to provide an alternative to NVIDIA’s InfiniBand technology.

Filed in . Read more about AI (Artificial Intelligence), CPU, GPU (Graphics Processing Unit), Intel and NVIDIA.