Hygon C86: The x86 China Still Makes

Written and maintained by Haink's infrastructure team · Compiled from Hygon and Sugon documentation, published benchmark results and toolchain records, 5 September 2026

Nearly every domestic processor programme in China builds on Arm, RISC-V or an architecture of its own. Hygon's C86 line is the exception: it is x86-64, and that single fact does more for a buyer's software estate than any specification on its datasheet.

It is also the host processor in Sugon's superpods — one C86-4G/H per compute node in a scaleX40, ten to a pod. This page covers the lineage, how the parts are named, what the line actually measures at, and the one feature set that has no equivalent in a Western x86 part.

The lineage

Hygon's x86 rights come from a 2016 licensing arrangement with AMD, structured as two joint ventures — Chengdu Haiguang Microelectronics (AMD 51% / Hygon 49%) and Chengdu Haiguang Integrated Circuit Design (AMD 30% / Hygon 70%) — for which AMD received $293 million to provide localised first-generation Zen cores for the domestic Chinese market.

Two boundaries on that arrangement matter. The licence covered Zen 1 only — not Zen 2, not later. And further licensing ended in 2019, so the line has now been developed independently for longer than the licensing relationship itself lasted. Hygon lists on the STAR Market and reported roughly ¥14.4 billion of revenue with ¥4.6 billion of R&D spending for 2025, which is the scale at which independent CPU development is an activity rather than a claim.

Readers arriving from the accelerator side should note it is the same arrangement: Hygon's DCU accelerators descend from AMD CDNA with a ROCm-derived software stack. One licensing event produced both product lines.

Why x86 is the whole argument

A domestic Arm or RISC-V processor asks an organisation to recompile, revalidate and in places rewrite. A domestic x86-64 processor asks it to do none of those things, because the binaries already exist.

For a mixed estate this is the difference between a processor that slots into what exists and one that starts a parallel platform. It is also why the C86 turns up as the host in accelerator systems: the accelerator is the unfamiliar part of the machine, and there is no reason to make the host unfamiliar too.

How the parts are named

The naming looks confusing and is not. It carries two independent pieces of information:

C86-{generation}G{series}

• Generation — 3G, 4G, 5G. The microarchitecture revision. C86-4G is the current server generation.
• Series number — 3000 for workstations, 5000 for general-purpose servers, 7000 at the top. Positioning within the generation.

So a C86-4G7460 is a fourth-generation, 7000-series part: 48 cores and 96 threads at 2.1 GHz, typically deployed two to a server. A C86 3350 is a 3000-series workstation part. The Sugon superpod's C86-4G/H is a fourth-generation part with a variant marker rather than a series number.

Within the 4G generation there are three published sub-models, and the difference between them is the single most actionable fact on this page.

If a workload needs AVX-512, it needs an M7. AVX512VNNI and AVX512BF16 in particular matter for CPU-side inference and mixed-precision work, and software that dispatches on CPUID will silently take a slower path on the other two. This is the question to put on a purchase order, and it is answerable before anything is shipped.

All 4G parts support DDR5 and PCIe 5.0.

Toolchain support

On 29 April 2026 the GCC compiler code base merged support for the C86-4G line, shipping with GCC 17. It provides three new targets — -march=c86-4g-m4, -march=c86-4g-m6 and -march=c86-4g-m7 — and -march=native detection.

That is a more meaningful signal than it looks. Upstream compiler support means the parts are describable to a build system by name rather than through generic x86-64 flags, that code can be tuned per sub-model, and that the work was done in public where anyone can inspect it. For a processor family whose main argument is ecosystem compatibility, upstreamed toolchain support is the argument being made concretely.

On the operating system side, the line is adapted to openKylin, Kylin and UOS, and Hygon reports more than 15,000 completed software and hardware compatibility certifications.

What it actually measures at

Unlike the accelerator line, the C86 has published results in standardised, independently administered benchmarks. That is worth stating plainly, because it is unusual in this segment.

Read the last row alongside the first two. SPEC Cloud and TPCx-IoT are throughput benchmarks on multi-socket systems, and the C86 does well in them. The CPU Mark comparison is against an Intel part from 2014, and the C86 3250 loses. Both are true, and together they describe the line accurately: competitive on aggregate throughput in server configurations, well behind current Western parts on per-core performance.

Hygon reports generational gains of 62% single-core and over 135% multi-core for its latest workstation parts against the previous generation — a vendor figure, and a large one, which is consistent with a line that started from Zen 1 and has had several independent revisions since.

SPEC and TPC publish their results publicly, so any of these can be checked at source rather than taken on trust.

The security block, which has no Western equivalent

This is the part of the C86 that is not a re-implementation of anything, and for buyers in its home market it is often the deciding feature.

• Cryptographic coprocessor (CCP) with hardware SM2, SM3 and SM4 — China's national cryptographic algorithms. Hygon reports SM4 encryption three times faster and decryption ten times faster than software implementations, and SM3 around 45% faster.
• CSV3 — real-time memory encryption with virtual machine integrity protection, locking the host out of modifying guest page tables. It is the basis for confidential-computing instances offered by Chinese cloud providers.
• Trusted key management (TKM) and a dedicated security processor implementing the Chinese trusted computing standard, with key material never leaving the security processor in plaintext.
• HCT, a software suite exposing all of it through OpenSSL-compatible interfaces rather than a proprietary API.

For a workload with domestic compliance requirements this is not a nice-to-have — it is frequently the requirement, and no Intel or AMD part satisfies it. Outside that context it is a differentiator with no buyer.

The C86-4G/H, in the scaleX40

One C86-4G/H per compute node, alongside four accelerators, across ten compute nodes. Each node is single-socket — one host processor feeding four accelerators, which is a tighter ratio than a conventional dual-socket eight-accelerator server and is why host cycles are a scarce resource in this design.

• Up to 12 DIMM slots supporting DDR5 at 6400 MT/s
• Four E1.S bays and two M.2 slots, the latter supporting hardware RAID1
• Four HHHL PCIe 5.0 slots for the cluster network, plus full-height slots designated for storage networking

That configuration also explains a figure that gets misread. Sugon quotes an aggregate memory access bandwidth above 8 TB/s for the pod, and it is host DDR5, not accelerator HBM. The arithmetic: twelve channels at 6400 MT/s and eight bytes per transfer is 614 GB/s per socket, and ten sockets is 6.1 TB/s of theoretical peak. Sugon's figure sits somewhat above that, so it evidently aggregates more than the DIMM channels alone — but it is the same order of magnitude, and nowhere near what forty HBM stacks deliver. Anyone quoting it as the accelerators' memory bandwidth is quoting the wrong subsystem.

When a C86 is not your choice

Where single-thread performance decides the workload. Databases with serialised hot paths, latency-bound transaction processing and licence-per-core software all reward fast cores, and the per-core comparison against current Western parts is where this line is weakest.

Where the software vendor certifies specific silicon. Enterprise applications with hardware certification lists are a support problem before they are a technical one. Check the list, not the architecture.

Where AVX-512 is required and the quote is not an M7. Two of the three 4G sub-models do not have it, and code that dispatches on CPUID will quietly run slower rather than fail. Specify the sub-model.

Where it deserves consideration: as the host processor in a domestic accelerator system, where it removes an entire category of porting work; in mixed estates where x86-64 binary compatibility avoids a parallel platform; in throughput-oriented server workloads, which is where the published benchmark results are strongest; and anywhere domestic cryptographic or confidential-computing requirements apply.

Frequently asked questions

What is the Hygon C86?

An x86-64 server and workstation processor line from Hygon Information Technology (海光信息), descended from first-generation AMD Zen cores licensed in 2016 and developed independently since 2019. It is the host processor in Sugon's scaleX superpods and is deployed widely in Chinese government, finance, telecoms and cloud infrastructure.

How are C86 parts named?

C86-{generation}G{series}. The generation — 3G, 4G, 5G — is the microarchitecture revision; the series number indicates positioning, with 3000 for workstations, 5000 for general-purpose servers and 7000 at the top. A C86-4G7460 is a fourth-generation 7000-series part with 48 cores and 96 threads at 2.1 GHz. Within the 4G generation there are three sub-models: M4, M6 and M7.

Does a C86 support AVX-512?

Only the C86-4G-M7. The M4 and M6 sub-models support AVX, AVX2, SHA and BMI2; the M7 adds full AVX-512 including AVX512F, AVX512BW, AVX512VNNI and AVX512BF16. Software that dispatches on CPUID will take a slower path on M4 and M6 rather than failing, so the sub-model belongs on the purchase order.

Does standard x86 software run on it?

Yes — it is x86-64, so operating systems, commercial binaries, hypervisors and build toolchains work without a port. GCC 17 added dedicated targets for the 4G sub-models in April 2026, and the line is adapted to openKylin, Kylin and UOS with over 15,000 reported compatibility certifications. The caveat is vector extensions: check AVX-512 requirements against the sub-model.

How fast is a C86 compared with Intel or AMD?

It depends which question you are asking. On aggregate server throughput it does well in standardised tests — a dual-socket C86-4G scored 154.9 on SPEC Cloud IaaS 2018 in February 2026, and a C86 system with TimechoDB took first place in TPCx-IoT in January 2026 at 24.65 million IoTps. On per-core performance it is well behind current Western parts: an 8-core C86 3250 measures roughly 17–18% slower than an Intel Core i7-5960X, a 2014 processor. Both readings are accurate and describe different workloads.

What security features does it have that x86 competitors do not?

Hardware implementations of China's SM2, SM3 and SM4 cryptographic algorithms via an on-die coprocessor, CSV3 real-time memory encryption with virtual machine integrity protection, trusted key management with a dedicated security processor, and the HCT software suite exposing these through OpenSSL-compatible interfaces. Where domestic compliance requirements apply this is often the deciding feature; elsewhere it has no buyer.

Which C86 is in the Sugon scaleX40?

The C86-4G/H — a fourth-generation part — one per compute node alongside four accelerators, across ten compute nodes. Each node is single-socket, so one host processor serves four accelerators. Each supports up to twelve DDR5-6400 DIMM slots, four E1.S bays and two M.2 slots with hardware RAID1, plus four HHHL PCIe 5.0 slots used for the cluster network.

Related

• Hygon DCU and DeepComputing 3 — the accelerator side of the same licensing arrangement
• Sugon scaleX40-3G — where the C86-4G/H is deployed, in full
• Building AI infrastructure without NVIDIA — where the host processor sits in the stack
• Sugon (中科曙光) · AI cluster architecture

Sources

• Hygon Information Technology (AMD joint-venture structure, $293m licence, Zen 1 scope, end of licensing in 2019, STAR Market listing)
• GCC project — C86-4G support merged 29 April 2026, shipping in GCC 17 (-march=c86-4g-m4/m6/m7 targets; the M4/M6/M7 instruction set split is documented in the patch)
• SPEC — Cloud IaaS 2018 results, February 2026 (dual-socket C86-4G platform score, scalability and instance ready time; published in SPEC's public results registry)
• TPC — TPCx-IoT results, January 2026 (C86 with TimechoDB, 24.65 million IoTps)
• Sugon — scaleX40 product page (C86-4G/H per compute node, node composition, DDR5-6400 support)
• Sugon — scaleX40-3G server user manual, 207 pp., and solution manual, 45 pp. (Hygon C86-4G/H series designation, storage and PCIe slot allocation, aggregate memory access bandwidth figure)
• Hygon — STAR Market IPO prospectus (processor product lines)
• Hygon — 2025 annual results (revenue and R&D spending); Hygon product materials (sub-model designations, cryptographic and confidential-computing features, generational performance claims)