Skip to main content

MetaX C600 (2025)

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​

ParameterValue
ArchitectureIn-house XCORE 1.5 (dozens of core IPs)
Process7nm (domestic SMIC process)
FP81000 TFLOPS (rivals the H100's 989 TFLOPS)
FP3215 TFLOPS
FP16 / BF16~500 TFLOPS (shares Tensor Cores with FP8 at a 1:2 ratio)
INT82000 TOPS (official release figures)
Precision SupportFP8 / FP16 / BF16 / FP32 / INT8 / INT4 mixed precision
Memory Capacity144GB HBM3e
Memory Bandwidth3.6 TB/s
TDP400 W
Energy Efficiency2.5 TFLOPS/W (FP8)
InterconnectIn-house MetaXLink high-speed interconnect, 320 GB/s bidirectional per card, flexible networking of 2-64 cards
Maximum Cluster256-card supernode (supports interconnection of 1,024+ cards in a single cluster)
ReliabilityECC memory correction + RAS fault tolerance + multi-tenant hardware isolation
Packaging2.5D advanced packaging (JCET / Tongfu Microelectronics)
EDADomestic solutions such as Empyrean
ReleaseJuly 2025 (debut at WAIC 2025) / official release in October 2025
Mass ProductionRisk production by end of 2025
Security CertificationNational 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:

StageDomestic SolutionDescription
IP DesignMetaX in-house XCORE 1.5Fully independent from instruction set to compute units
Wafer ManufacturingSMIC and othersDomestic 7nm process
Packaging & TestingJCET, Tongfu Microelectronics2.5D advanced packaging
EDA ToolsEmpyrean and othersDomestic EDA tools being adopted step by step
Software StackMXMACAIn-house, CUDA compatible

Strategic significance: completely eliminating dependence on overseas supply chains and building secure, reliable, independent compute infrastructure.

MXMACA Software Ecosystem​

LayerToolDescription
Software StackMXMACAMetaX unified compute architecture
AI FrameworksPyTorchNative support
DeepSpeedDistributed training
CUDA CompatibilityAutomatic migration toolsCuts code migration cost by over 90%
Large ModelsSupports domestic thousand-card clustersCompleted 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

Vendor Information​

ParameterDetails
CompanyMetaX Integrated Circuits (Shanghai) Co., Ltd.
Nanjing TeamR&D team of nearly 100 people ("strongest brains")
Project InitiationFebruary 2024 (C600 initiated, investment of about 1.37 billion RMB)
Tape-OutOctober 2024
ReleaseOctober 2025
Mass ProductionRisk production by end of 2025
IPO StatusCompleted two rounds of inquiry responses, pursuing a STAR Market listing
Revenue320 million RMB in H1 2025, with 1.43 billion RMB in orders on hand
Deployed25,000+ GPUs
PricingC500 about 38,900 RMB per card (1/3 of the A100)
C700 PlanProduction 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​

MetricMetaX C600NVIDIA H100Difference
Process7nm (domestic)TSMC 4nmH100 more advanced
FP81000 TFLOPS989 TFLOPSC600 +1%
FP16Not disclosed (estimated ~300 TFLOPS)989 TFLOPSTo be confirmed
FP3215 TFLOPS67 TFLOPSH100 4.5x
Memory144GB HBM3e80GB HBM3C600 +80%
Bandwidth3.6 TB/s3.35 TB/sC600 +7%
Energy Efficiency2.5 TFLOPS/W~1.4 TFLOPS/WC600 has the edge (FP8)
EcosystemMXMACA (CUDA compatible)CUDAH100 more mature
PriceTo be announced (C500 is 1/3 of the A100)~$30k+C600 expected to be cheaper
Domestic Production✅ 100%❌Unique to C600

Key Timeline​

DateEvent
2024-02C600 project initiated (investment of 1.37 billion RMB)
2024-10C600 tape-out delivered
2025-07C600 Baidu Baike entry created
2025-10C600 officially released in Nanjing (this page)
End of 2025C600 risk production
2027 H2C700 targeted for production

Production Status Confirmation (2026-09 Update)​

ParameterStatus
Mass Production TimeMass production achieved in May 2026
ProcessDomestic 7nm
Memory144 GB HBM3e
FP8 ComputeAbout 1,000 TFLOPS
ComplianceCompleted the national security and stability testing
InterconnectIn-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.