MetaX Xiyun C700
Product Overview
The MetaX Xiyun C700 is the next-generation training flagship GPU of MetaX's Xiyun C series, with development initiated in 2025. It adopts a new-generation GPU architecture and instruction set (not the C600's XCORE 1.5), targets performance approaching the NVIDIA H100, and further extends support for lower-precision computing such as FP4 to meet the mixed-precision training needs of next-generation large models. Compared with the mass-produced C600, the C700 delivers major improvements in compute, memory, communication, and energy efficiency, and is developed on a higher-node domestic advanced process.
According to MetaX's IPO prospectus and securities research reports, the C700 project represents a total investment of about RMB 2.04 billion; core design and functional verification are largely complete, with tape-out testing expected in 2026 and mass production in the second half of 2027. It is the key test of MetaX's architectural iteration capability, and its mass-production arrival would mean MetaX formally enters the ranks of global high-end GPU competitors.
Core Specifications
| Parameter | Value |
|---|---|
| Architecture | New-generation GPU architecture and instruction set (not XCORE 1.5) |
| Process Node | Higher-node domestic advanced process (specific node not disclosed) |
| FP32 Compute | Not disclosed (targets H100-class performance) |
| FP16 / BF16 Compute | Not disclosed |
| FP8 Compute | Not disclosed (expected support, further improved over the C600) |
| FP4 Compute | Extended support (new-generation low precision, specifics not disclosed) |
| INT8 Compute | Not disclosed |
| Memory Capacity | Not disclosed |
| Memory Type | Not disclosed (expected HBM3e or higher) |
| Memory Bandwidth | Not disclosed |
| TDP | Not disclosed |
| Interconnect | MetaXLink high-speed inter-card interconnect (expected to support larger-scale supernodes) |
| Interface | Not disclosed (expected PCIe 5.0 / CXL, etc.) |
| Launch | 2025 (development initiated) |
| Mass Production/Availability | Tape-out expected 2026, mass production expected H2 2027 (in development) |
⚠️ Specification Notes: The C700 is currently in development and MetaX has not published any complete specifications. Information such as "targets the H100", "extended FP4", "domestic advanced process", and "2026 tape-out / H2 2027 mass production" comes from MetaX's IPO prospectus and public research reports by Donghai Securities / Huachuang Securities. All compute, memory, bandwidth, and power fields are not disclosed and should not be treated as released specifications.
Key Features (Planned Direction)
- H100 benchmark: targets performance approaching the NVIDIA H100, entering the global high-end GPU ranks
- New-generation architecture: a brand-new GPU architecture and instruction set, iterating beyond XCORE 1.5
- FP4 low precision: extended lower-precision computing for next-generation large-model mixed-precision training
- Domestic advanced process: built on a higher-node domestic process, improving compute/energy efficiency
- Supernode scaling: works with MetaXLink and the Xijing S supernode, evolving toward 10,000-card clusters
Vendor Information
| Parameter | Value |
|---|---|
| Company | MetaX Integrated Circuits (Shanghai) Co., Ltd. |
| Headquarters | Shanghai |
| Founded | 2020-09 |
| Listing | STAR Market 688802 |
| R&D Investment | Total C700 project investment of about RMB 2.04 billion (IPO prospectus) |
Use Cases
- ✅ Large-model training (H100 benchmark, planned positioning)
- ✅ Large-scale intelligent computing centers (supernode clusters)
- ✅ Next-generation mixed-precision training (FP4/FP8)
- ❌ No product currently on sale (in development, specs undetermined)
- ❌ Selection reference must await official tape-out/mass-production data
Latest Development Progress (2026-09-14 Interim Results Briefing)
MetaX updated the C700's progress at its 2026 interim results briefing:
| Parameter | Latest Status |
|---|---|
| Current Status | Core design and functional verification near completion |
| Next Steps | Advancing tape-out → software adaptation → customer testing → mass-production introduction |
| Project Initiation | April 2025 |
| Performance Target | Overall performance benchmarked against the NVIDIA H100 |
| New Precision | Added support for low-precision computing such as FP4 (for large-model inference) |
| Supply Chain | Reliable domestic solutions for wafer foundry, packaging, and high-capacity memory, building a closed domestic supply chain loop from design → manufacturing → packaging and testing |
| R&D Strategy | Adhering to "mass-produce one generation, develop one generation, plan one generation" |
| Expected Milestones | Tape-out testing in 2026, mass production in the second half of 2027 |
📌 Same-Generation Product Comparison: The Xiyun C600, based on the domestic 7nm process, entered mass production in May 2026, equipped with 144GB HBM3e and FP8 compute of about 1,000 TFLOPS, and has completed national security and stability testing. The proprietary MetaXLink high-speed interconnect has been applied at 1,000-card scale with bandwidth approaching H100 levels, and 10,000-card-scale deployment is underway.
Key Insight: The C700's core selling point is not just "matching the H100" but supply chain self-sufficiency — MetaX CTO and chief hardware architect Peng Li explicitly stated that the company has domestic solutions for key links such as wafer foundry, packaging, and high-capacity memory, and that "even if geopolitical tensions intensify in the future, the impact on the company will be limited". In 2026, with HBM constrained and channel prices up 20~30% in three months, that statement carries more weight than any compute figure.
Related Cards
- MetaX Xiyun C600 — Flagship on sale (XCORE 1.5, HBM3e, mass-produced in 2026)
- MetaX Xiyun C500 — Mass-production workhorse (XCORE 1.0)
- MetaX Xiyun C290 — Early XCORE 1.0 model
- Biren BR100 — Domestic general-purpose GPU competitor