Hygon’s 128-Core & 512-Thread C86 CPU Targets Intel Xeon With 15% IPC Gain, As China Races to Cut Foreign Chip Reliance

Hygon, China's premier chipmaker, has presented its latest CPUs and GPUs for data centers, offering up to 128 cores & 512 threads.

Hygon, based in Beijing, China, has been hard at work producing high-quality chips for the domestic markets. Their previous C86 solutions, designed in collaboration with AMD, have seen major interest by domestic players, and now, the next-generation C86 processor development is underway.

Now, Hygon has shared plans to develop six chips that will be offered to domestic audiences in China. The leading chip in the lineup is the next-gen C86 CPU, designed for general-purpose computing. Currently, China offers its C86-4G series for mainstream and enterprise audiences. The next-gen C86-5G series aims to be bolder.

First Datacenter Chip With SMT4 Support: Up To 128 Cores & 512 Threads

Starting with the core highlights, the next-gen Hygon C86 series will feature a brand new architecture, offering an IPC uplift greater than 15%. The chip will also feature STM4 multi-threading support, which means that each CPU core can handle four concurrent threads, expanding its enterprise and server capabilities. The flagship chip will feature 128 cores & quadruple the thread count of 512.

There are also some major updates on the instruction support with AVX512 being added for enhanced vector compute, and AI acceleration from INT8 and BF16 is also a major selling point.

• New microarchitecture, IPC improved by 15%+
• Supports SMT4 multi-threading to meet high-concurrency business needs.
• Compatible with the AVX512 instruction set, enhanced vector computing power.
• New AI acceleration instructions, such as INT8 and BF16, have been added.

In terms of performance, Hygon is claiming that its next-gen C86 chips will rival Intel's Xeon 6 lineup. Each chip will offer 10 TFLOPs of FP64 performance. That in itself is very promising as it will bring the Chinese domestic markets on par with the big players. The platform itself will offer 104 PCIe 5.0 lanes, which will reduce the dependency on Broadcom's PCIe Switches.

2 of 9

The first of these chips is already in mass production. Server offerings include the 2U air-cooled dual-socket rack server H620G59, fanless liquid-cooled cold-plate high-density cabinet TC800G6 (1P total compute per cabinet, PUE 1.08), and immersion phase-change liquid-cooled ultra-high-density cabinet TC8600H G5 (80,000+ CPU cores per group).

GPU Development Going At A Steady Pace

But Hygon isn't stopping at CPUs; they have a whole ecosystem of chips in development. Next is a new GPU accelerator, "DCU," which is designed with AI training and compute in mind. This chip will feature a full-precision GPGPU architecture with support for ultra-high-speed inter-chip interconnect. Hygon will be leveraging the latest HBM memory technologies while also offering FP64/FP16/BF16 support. It is stated that the chip will be comparable to NVIDIA's A100 based on the Ampere architecture.

• Full-precision GPGPU architecture
• Supports FP64/PF16/BF16 multi-precision calculation
• Ultra-high-speed inter-chip interconnect
• Paired with high-speed HBM memory

A Fully Integrated Networking Solution For Data Centers

The next four chips include a PCIe 5.0 Switch, a Scale-Up Interconnect Switch, a 400G Network Interface chip, and a ScaleFabric 400/800G switch. These will be crucial for Hygon to develop its own AI & high-performance Enterprise ecosystem. The details of these chips are listed below:

Hygon PCIe 5.0 Switch:

• High-Speed I/O
• Targets Broadcom's high-end PCIe switching chips.
• 104 lanes

Scale-Up Interconnect Switch:

• Comparable to NVLINK + NVSwitch
• Ultra high-speed interconnect for multiple GPUs/CPUs

ScaleFabric 400G Network Interface Chip:

• 400Gb/s port speed, native RDMA protocol
• Credit-based flow control + lossless LLR characteristics
• Network card communication latency is 0.93µs, maximum QP is 256K.
• Plug and play, supports smooth evolution to 800G
• Comparable to NVIDIA InfiniBand NDR Network AICs.

ScaleFabric 400/800G Switch:

• 400/800Gb/s port speed, native RDMA
• 80*400Gb/s switching port density, 64Tb switching capacity
• 260ns switching latency, millisecond-level link failure switching
• Self-developed 112G high-speed SerDes IP, leading in signal driving capabilities.
• Comparable to NVIDIA InfiniBand NDR Switches

Hygon confirmed that its CPUs are now in mass production and are already shipping in various data-center aimed products. The GPUs would also enter production soon.

Follow Wccftech on Google to get more of our news coverage in your feeds.