Product Overview
The MetaX C600 is the first end-to-end domestic general-purpose GPU, released by MetaX Integrated Circuits in October 2025. Based on the in-house XCORE 1.5 architecture and a domestic 7nm process (SMIC), it integrates 144GB HBM3e memory (3.6 TB/s bandwidth) with peak FP8 compute of 1000 TFLOPS (benchmarked against the NVIDIA H100's 989 TFLOPS FP8), 400W TDP, and the MXMACA software stack compatible with the CUDA ecosystem.
It is the first domestic general-purpose GPU with an end-to-end domestic production chain (IP design → wafer manufacturing → packaging and testing), achieving 100% de-Americanization from chip design to manufacturing.
MetaX Product Lines:
- N Series: inference chips
- C Series: general-purpose GPUs (training + inference) — C500 (64GB HBM) → C600 (this page) → C700 (planned for 2027 H2)
- G Series: graphics rendering
Core Specifications
| Parameter | Value |
|---|
| Architecture | In-house XCORE 1.5 (dozens of core IPs) |
| Process | 7nm (domestic SMIC process) |
| FP8 | 1000 TFLOPS (rivals the H100's 989 TFLOPS) |
| FP32 | 15 TFLOPS |
| FP16 / BF16 | ~500 TFLOPS (shares Tensor Cores with FP8 at a 1:2 ratio) |
| INT8 | 2000 TOPS (official release figures) |
| Precision Support | FP8 / FP16 / BF16 / FP32 / INT8 / INT4 mixed precision |
| Memory Capacity | 144GB HBM3e |
| Memory Bandwidth | 3.6 TB/s |
| TDP | 400 W |
| Energy Efficiency | 2.5 TFLOPS/W (FP8) |
| Interconnect | In-house MetaXLink high-speed interconnect, 320 GB/s bidirectional per card, flexible networking of 2-64 cards |
| Maximum Cluster | 256-card supernode (supports interconnection of 1,024+ cards in a single cluster) |
| Reliability | ECC memory correction + RAS fault tolerance + multi-tenant hardware isolation |
| Packaging | 2.5D advanced packaging (JCET / Tongfu Microelectronics) |
| EDA | Domestic solutions such as Empyrean |
| Release | July 2025 (debut at WAIC 2025) / official release in October 2025 |
| Mass Production | Risk production by end of 2025 |
| Security Certification | National Security and Reliability Assessment Level I (2026-05-26, the first fully domestic flagship GPGPU on the list) |
End-to-End Domestic Production
The MetaX C600 is the first general-purpose GPU that is domestic across the entire chain:
| Stage | Domestic Solution | Description |
|---|
| IP Design | MetaX in-house XCORE 1.5 | Fully independent from instruction set to compute units |
| Wafer Manufacturing | SMIC and others | Domestic 7nm process |
| Packaging & Testing | JCET, Tongfu Microelectronics | 2.5D advanced packaging |
| EDA Tools | Empyrean and others | Domestic EDA tools being adopted step by step |
| Software Stack | MXMACA | In-house, CUDA compatible |
Strategic significance: completely eliminating dependence on overseas supply chains and building secure, reliable, independent compute infrastructure.
MXMACA Software Ecosystem
| Layer | Tool | Description |
|---|
| Software Stack | MXMACA | MetaX unified compute architecture |
| AI Frameworks | PyTorch | Native support |
| DeepSpeed | Distributed training |
| CUDA Compatibility | Automatic migration tools | Cuts code migration cost by over 90% |
| Large Models | Supports domestic thousand-card clusters | Completed full-parameter training of multiple models in cooperation with the Chinese Academy of Sciences |
| Ecosystem Strategy | "1+6+X" | GPU foundation empowering finance, healthcare, energy, education/research, transportation, and entertainment |
| Parameter | Details |
|---|
| Company | MetaX Integrated Circuits (Shanghai) Co., Ltd. |
| Nanjing Team | R&D team of nearly 100 people ("strongest brains") |
| Project Initiation | February 2024 (C600 initiated, investment of about 1.37 billion RMB) |
| Tape-Out | October 2024 |
| Release | October 2025 |
| Mass Production | Risk production by end of 2025 |
| IPO Status | Completed two rounds of inquiry responses, pursuing a STAR Market listing |
| Revenue | 320 million RMB in H1 2025, with 1.43 billion RMB in orders on hand |
| Deployed | 25,000+ GPUs |
| Pricing | C500 about 38,900 RMB per card (1/3 of the A100) |
| C700 Plan | Production in 2027 H2, over 80 billion transistors |
Use Cases
- ✅ Domestic large-model training (thousand-card clusters verified for full-parameter training)
- ✅ Generative AI inference (extra-large 144GB HBM3e memory)
- ✅ Domestic intelligent computing centers (end-to-end domestic production; de-Americanization is a hard requirement)
- ✅ Government/SOE projects (mandatory independent and controllable requirements)
- ✅ FP8 inference (1000 TFLOPS, rivaling the H100)
- ✅ CUDA migration scenarios (90%+ reduction in migration cost)
- ❌ Graphics rendering (general-purpose but not graphics-dedicated; the G Series is a better fit)
- ❌ Native CUDA ecosystem (migration via MXMACA required)
- ❌ Early-stage production supply (risk production only by end of 2025)
Comparison with NVIDIA H100
| Metric | MetaX C600 | NVIDIA H100 | Difference |
|---|
| Process | 7nm (domestic) | TSMC 4nm | H100 more advanced |
| FP8 | 1000 TFLOPS | 989 TFLOPS | C600 +1% |
| FP16 | Not disclosed (estimated ~300 TFLOPS) | 989 TFLOPS | To be confirmed |
| FP32 | 15 TFLOPS | 67 TFLOPS | H100 4.5x |
| Memory | 144GB HBM3e | 80GB HBM3 | C600 +80% |
| Bandwidth | 3.6 TB/s | 3.35 TB/s | C600 +7% |
| Energy Efficiency | 2.5 TFLOPS/W | ~1.4 TFLOPS/W | C600 has the edge (FP8) |
| Ecosystem | MXMACA (CUDA compatible) | CUDA | H100 more mature |
| Price | To be announced (C500 is 1/3 of the A100) | ~$30k+ | C600 expected to be cheaper |
| Domestic Production | ✅ 100% | ❌ | Unique to C600 |
Key Timeline
| Date | Event |
|---|
| 2024-02 | C600 project initiated (investment of 1.37 billion RMB) |
| 2024-10 | C600 tape-out delivered |
| 2025-07 | C600 Baidu Baike entry created |
| 2025-10 | C600 officially released in Nanjing (this page) |
| End of 2025 | C600 risk production |
| 2027 H2 | C700 targeted for production |
Production Status Confirmation (2026-09 Update)
| Parameter | Status |
|---|
| Mass Production Time | Mass production achieved in May 2026 |
| Process | Domestic 7nm |
| Memory | 144 GB HBM3e |
| FP8 Compute | About 1,000 TFLOPS |
| Compliance | Completed the national security and stability testing |
| Interconnect | In-house MetaXLink, already applied at thousand-card scale (bandwidth close to H100 level), now pushing toward 10,000-card deployments |
📌 Procurement positioning: the C600 is MetaX's currently purchasable flagship on sale. Until the next-generation C700 enters production in the second half of 2027, the C600 serves as the transitional supply of high-end domestic GPUs — for domestic compute projects that need to land right now, it is one of the options worth evaluating.