Skip to main content

Hygon DCU 4 (ShenSuan No. 4)

Product Overview​

Hygon's ShenSuan No. 4 DCU is the newest generation of GPGPU coprocessor in the ShenSuan line from Hygon Information (688041.SH), and the next-generation AI compute core of the company's "CPU + DCU" dual-chip strategy. On September 29, 2026, Hygon confirmed on its investor interaction platform that the new generation of general-purpose processors and coprocessors is progressing on schedule and in line with expectations, with ShenSuan No. 4 expected to launch in H2 2026.

Four facts are currently public: roughly 2x the performance of the prior generation, FP4 support, pairing with the second-generation Scale up switch chip HySwitch 2.0, and 192-card interconnect per chip. The 192-card interconnect is the most consequential architectural number of this generation — combined with the Scale up capability of HySwitch 2.0, it means just two network tiers can support clusters of tens of thousands of cards, a step-change in Hypon's large-scale training fabric and a direct competitive answer to the thousand-card supernode roadmap taken by other domestic AI chip vendors.

The DCU line sits on a mature software foundation. DCU products are built on a GPGPU architecture with full-precision compute capability, and the in-house DTK heterogeneous computing platform and DAS AI software stack have integrated over 2,000 operators and support more than 100 mainstream AI tools. As of H1 2026, Hygon had completed full adaptation and optimization for over 400 mainstream large models including DeepSeek, Qwen3, Hunyuan, and Zhipu, covering 99% of non-closed-source models worldwide. DCU products have landed in more than 20 key industries and over 300 application scenarios, serving government departments, state-owned banks, the three major telecom carriers, and internet companies. In 2025, DCU shipments reached roughly 80,000 units, up about 150% year over year, placing the line among the leaders in domestic AI accelerators.

Hygon's recent financials underline the shipment base for the new product. In H1 2026, revenue reached RMB 9.099 billion, up 66.52% year over year, with RMB 1.798 billion in net profit attributable to shareholders, up 49.69%. Second-quarter net profit grew 61% quarter over quarter, and R&D spending of RMB 2.652 billion equaled 29.15% of revenue.

⚠️ Disclosure status (important): Hygon has not published ShenSuan No. 4's compute figures, memory capacity, memory type, memory bandwidth, or TDP. Every one of those fields in the table below is marked "Not disclosed", with no speculative fill. Circulating claims about FP6 support and 100,000-card clusters come from unofficial channels and are not adopted here. The only quantifiable official statement is "roughly 2x the performance of the prior generation" — note this is a relative multiple, not an absolute compute figure, and Hygon has not specified which baseline part or test methodology it refers to.

Key Specifications​

ParameterValue
ArchitectureGPGPU (ShenSuan series, CUDA-compatible ecosystem)
ProcessNot disclosed
MemoryNot disclosed
Memory TypeNot disclosed
Memory BandwidthNot disclosed
FP8Not disclosed
FP4Supported
FP16Not disclosed
INT8Not disclosed
TDPNot disclosed
InterconnectHySwitch 2.0 (192-card interconnect per chip)
Release / ProductionExpected to launch in H2 2026

📌 On undisclosed fields: process node, memory capacity, memory type, memory bandwidth, FP8 / FP16 / INT8 compute, and TDP — eight fields in total — have not been disclosed by Hygon and are all marked "Not disclosed" with no speculation.

📌 The one quantifiable metric: the official wording is "roughly 2x the performance of the prior generation", with no absolute compute figure, baseline part, or test methodology given. The preceding generation is the ShenSuan No. 3 (DCU K100 AI version, 2024, 192 TFLOPS FP16 / 392 TOPS INT8 / 64GB HBM2e).

📌 On FP4: Hygon has made no standalone statement on FP4 support for ShenSuan No. 4; the capability comes from market interpretation of the new part's positioning and the industry-wide FP4 trend. Under a strict evidentiary standard, treat it as unconfirmed by the vendor.

Key Features​

1. HySwitch 2.0 and 192-Card Scale up Interconnect​

  • ShenSuan No. 4 pairs with the second-generation Scale up switch chip HySwitch 2.0
  • A single chip supports 192-card interconnect — a key metric for the DCU line's Scale up direction, far exceeding the interconnect scale of prior generations
  • Two network tiers support clusters of tens of thousands of cards, sharply reducing both the tier count and latency overhead of large-scale clusters
  • Separating Scale up (chip-to-chip) from Scale out (node-to-node) is the cluster-scaling route widely adopted by domestic AI chip vendors

2. "CPU + DCU" Dual-Chip Synergy​

  • Hygon is one of the few domestic firms able to develop and commercially operate both x86 high-end general-purpose processors (CPU) and CUDA-like ecosystem coprocessors (DCU) at scale
  • The CPU is x86 instruction-set compatible, with a significant ecosystem maturity advantage; the DCU is GPGPU-based with full-precision compute capability
  • This "dual-chip synergy" gives Hygon integrated delivery capability at the whole-system and cluster level

3. DTK and DAS Software Ecosystem​

LayerContent
Heterogeneous computing platformDTK (DCU Toolkit), built on the open-source ecosystem, supporting mainstream AI tools
AI software stackDAS, with over 2,000 operators integrated
ToolingSupport for more than 100 mainstream AI tools
Model adaptationFull adaptation and optimization completed for over 400 mainstream large models including DeepSeek, Qwen3, Hunyuan, and Zhipu
CoverageCovers 99% of non-closed-source large models worldwide
Ecosystem allianceThe "Guanghe Organization" has aggregated over 6,000 ecosystem partners

4. Deployment and Customers​

  • Landed in more than 20 key industries and over 300 application scenarios
  • Customers span government departments, state-owned banks, the three major telecom carriers, and internet companies
  • Deep partnerships established with leading internet firms including ByteDance, Tencent, Alibaba, and Baidu for joint technology R&D and custom services
  • Applied across big data processing, artificial intelligence, and commercial computing

5. H1 2026 Financials​

MetricH1 2026Change
RevenueRMB 9.099 B+66.52% YoY
Net profit attributable to shareholdersRMB 1.798 B+49.69% YoY
Net profit excluding non-recurring itemsRMB 1.631 B+49.67% YoY
Q2 net profitRMB 1.111 B+61% QoQ
R&D spendingRMB 2.652 B+55.05% YoY
R&D as share of revenue29.15%—
Basic EPSRMB 0.78+50% YoY
  • Full-year 2025 R&D expense was RMB 4.145 billion, up 42.45% year over year
  • 3,007 R&D personnel, 80.83% of total headcount, of whom 78.65% hold a master's degree or above
  • 2,099 intellectual property projects cumulatively, including 1,204 invention patents

Use Cases​

  • ✅ Domestic AI compute infrastructure buildout (serving both the Xinchuang and general internet markets)
  • ✅ Training clusters from thousands to tens of thousands of cards (HySwitch 2.0, 192-card Scale up, two network tiers)
  • ✅ Large-model training and inference (400+ adapted models, covering 99% of non-closed-source models)
  • ✅ Scientific and commercial computing (GPGPU full-precision compute)
  • ✅ Integrated CPU + DCU delivery (x86 plus CUDA-like dual-chip synergy)
  • ❌ Procurement evaluation that requires concrete peak compute or memory figures (key specs not yet disclosed)
  • ❌ Deployment before H2 2026 (the product is expected to launch in H2 2026)

Comparison​

Data Sources​