Hygon Unveils 128-Core C86 CPU and Full-Stack Data Center Platform
Hygon has announced its next-generation data center silicon portfolio, expanding beyond server processors into a comprehensive infrastructure ecosystem that includes GPU accelerators, high-speed networking, and advanced cooling solutions. The latest C86 server CPU architecture delivers a claimed 15%+ IPC improvement while scaling to 128 cores and 512 simultaneous threads through four-way simultaneous multithreading (SMT4).
Rather than focusing solely on CPU performance, Hygon’s roadmap reflects a broader strategy: building a vertically integrated platform that combines general-purpose compute, AI acceleration, high-bandwidth networking, and deployment-ready server systems. This approach aims to reduce dependence on external ecosystem vendors while targeting hyperscale cloud, HPC, and AI infrastructure deployments.
π Next-Generation C86 Server CPU #
The newest C86 architecture is designed for enterprise-scale concurrency, cloud-native applications, and AI-assisted workloads. Alongside higher core density, the platform significantly expands vector processing capabilities and I/O bandwidth.
Core Specifications #
| Feature | Specification |
|---|---|
| IPC Improvement | Over 15% |
| Maximum Configuration | 128 CPU Cores / 512 Threads |
| Simultaneous Multithreading | SMT4 (4 Threads per Core) |
| AI Instructions | AVX-512, INT8, BF16 |
| Peak FP64 Performance | Up to 10 TFLOPS |
| PCI Express | 104 PCIe 5.0 Lanes |
SMT4 Increases Compute Density #
One of the most significant architectural enhancements is the adoption of four-way simultaneous multithreading. While most modern server CPUs utilize SMT2, Hygon allows each physical core to execute four independent hardware threads.
Potential advantages include:
- Higher utilization under cloud-native workloads
- Increased throughput for microservices
- Better container consolidation
- Improved virtualization density
- Greater parallel request handling
This design primarily benefits workloads where thread-level parallelism is more valuable than single-thread performance.
Integrated AI Execution #
The processor also extends its capabilities into AI inference by incorporating native support for:
- AVX-512 vector instructions
- INT8 acceleration
- BF16 acceleration
These instruction extensions enable CPUs to perform lightweight inference workloads without relying exclusively on discrete GPU accelerators. This approach is particularly suitable for:
- Edge inference
- Recommendation systems
- Search services
- AI-enhanced databases
- Enterprise analytics
Although large-scale LLM training still requires GPUs, many production inference workloads can execute efficiently on modern vector-enabled CPUs.
High-Bandwidth I/O #
The platform exposes 104 PCIe 5.0 lanes directly from the processor, enabling dense system configurations that may include:
- Multiple GPU accelerators
- Large NVMe storage arrays
- High-speed network adapters
- FPGA accelerator cards
- SmartNIC deployments
Reducing dependence on external PCIe switch fabrics helps lower latency while simplifying server platform design.
π₯οΈ Full-Stack Data Center Silicon Strategy #
Beyond CPUs, Hygon introduced several complementary products that collectively form a complete data center infrastructure stack.
These additions target three primary domains:
- AI acceleration
- High-speed networking
- Fabric interconnects
The objective is to optimize both compute performance and communication efficiency across large distributed clusters.
DCU GPU Accelerator #
Hygon’s latest Deep Computing Unit (DCU) is a general-purpose accelerator targeting both scientific computing and AI training workloads.
Key capabilities include:
- Native FP64 support
- FP16 acceleration
- BF16 acceleration
- High Bandwidth Memory (HBM)
- High-speed chip-to-chip interconnect
This combination enables the accelerator to address both traditional HPC simulations and modern AI training pipelines.
The use of HBM significantly improves memory bandwidth compared to conventional GDDR memory, helping alleviate bottlenecks commonly encountered in large-scale model training.
π High-Speed Fabric and Networking #
Large AI clusters require efficient communication between thousands of GPUs and CPUs. To address this challenge, Hygon introduced several networking products spanning PCIe expansion, RDMA networking, and switch fabrics.
| Product | Key Specifications | Target Segment |
|---|---|---|
| PCIe 5.0 Switch | 104 PCIe 5.0 lanes | High-density server expansion |
| Scale-Out Fabric Switch | Proprietary node interconnect | Distributed AI clusters |
| 400G Smart NIC | 400 Gb/s, RDMA, 0.93 ΞΌs latency, 256K Queue Pairs | High-performance networking |
| 400G / 800G Ethernet Switch | Up to 64 Tb switching capacity, 260 ns latency | AI fabric infrastructure |
RDMA for AI Clusters #
Native RDMA support allows compute nodes to exchange data directly between memory spaces while bypassing the operating system networking stack.
Benefits include:
- Lower communication latency
- Reduced CPU overhead
- Higher throughput
- Improved GPU utilization
- Better scaling across distributed training jobs
As AI models continue to grow in parameter count, network latency increasingly becomes a limiting factor for overall cluster efficiency.
π§ Enterprise Server Platforms and Cooling #
To complement the new silicon portfolio, Hygon also introduced several server reference platforms optimized for different deployment environments.
H620G59 #
A conventional 2U dual-socket air-cooled server intended for enterprise racks using traditional cooling infrastructure.
Ideal deployment scenarios include:
- Enterprise virtualization
- Database servers
- General cloud infrastructure
- Private cloud deployments
TC800G6 #
A fanless cold-plate liquid-cooled platform featuring a reported Power Usage Effectiveness (PUE) of 1.08.
Cold-plate cooling transfers heat directly from processors and accelerators to circulating liquid, enabling higher thermal efficiency while reducing overall power consumption.
TC8600H G5 #
An immersion phase-change liquid-cooled rack designed for ultra-high-density computing environments.
According to Hygon, a fully populated deployment can scale beyond 80,000 CPU cores within a single clustered infrastructure, making it suitable for HPC and AI supercomputing installations.
βοΈ Deployment Status #
Hygon confirmed that the new C86 processor family has already entered mass production and is shipping within production-ready server systems.
The deployment roadmap currently includes:
| Component | Status |
|---|---|
| C86 Server CPU | Mass Production |
| Enterprise Server Platforms | Available |
| DCU GPU Accelerator | Entering Mass Production |
| Network Switch Portfolio | Availability Pending |
Although commercial launch dates for the networking products have not yet been announced, the roadmap indicates continued expansion toward a fully integrated data center platform.
π Building an End-to-End Data Center Ecosystem #
Rather than competing solely on CPU performance, Hygon is positioning itself as a provider of an integrated infrastructure stack that spans compute, acceleration, networking, and deployment platforms.
Its latest portfolio combines:
- High-core-count server processors
- AI-focused GPU accelerators
- PCIe switching infrastructure
- RDMA-capable networking
- High-capacity Ethernet fabrics
- Air-cooled and liquid-cooled server platforms
As hyperscale operators increasingly prioritize vertically optimized hardware ecosystems, this strategy aligns with broader industry trends toward tightly integrated infrastructure designed for cloud computing, HPC, and large-scale AI workloads.