MetaX Integrated Circuit's flagship iterative model based on XiYun C500/C550, released in September 2024, using self-developed XCORE 1.5 architecture, 12nm domestic process (SMIC), performance targeting international mainstream AI training cards.
Core Specifications
| Specification | Value |
|---|
| Architecture | XCORE 1.5 (XiHe architecture iteration) |
| Process | 12 nm (SMIC domestic process) |
| TDP | 350 W (estimated) |
| Memory | 64–96 GB HBM2e (estimated) |
| Memory Bandwidth | ~1800–2000 GB/s (estimated) |
| FP32 Compute | 18 TFLOPS (estimated) |
| FP16 / BF16 Compute | 35 TFLOPS (estimated) |
| INT8 Compute | 70 TOPS (estimated) |
| Interface | PCIe 5.0 x16 (estimated) |
| Interconnect | MetaXLink (estimated) |
| Release | 2024-09 (estimated) |
| Price | Not disclosed (estimated ¥55,000) |
Technical Highlights
- XCORE 1.5 architecture: Iterative upgrade of XiHe architecture, optimized for tensor operations
- 12nm domestic process: SMIC foundry, promoting supply chain domesticization
- Flagship performance: As the flagship iteration of C500 series, compute fully exceeds C550
- Full-process domestic production: From architecture design to manufacturing and packaging, gradually achieving full domestic production
- MetaXLink interconnect: Supports multi-card expansion, building large-scale training clusters
Product Positioning
XiYun C588 is the flagship iterative model of MetaX's C500 series, using XCORE 1.5 architecture and 12nm domestic process, performance targeting NVIDIA A100 80GB. As the transitional flagship between C500 and C600, C588 has significant improvements in compute and memory bandwidth.
Comparison with C500/C550
| Metric | XiYun C500 | XiYun C550 | XiYun C588 |
|---|
| Architecture | XCORE 1.0 | XCORE 1.0 | XCORE 1.5 |
| Process | 7 nm | 7 nm | 12 nm (domestic) |
| FP16 Compute | 15 TFLOPS | 22.4 TFLOPS | ~35 TFLOPS (estimated) |
| Memory Bandwidth | ~1200 GB/s | ~1600–1800 GB/s | ~1800–2000 GB/s (estimated) |
| Release Time | 2023 | 2024 | 2024 |
| Positioning | First training-inference integrated | C500 iteration | Flagship iteration |
Application Scenarios
- Large AI model training (targeting A100)
- High-performance AI inference
- Scientific computing and simulation
- Data center AI computing infrastructure
- Domestic A100 replacement scenarios
Reference Price
| Channel | Price | Description |
|---|
| Official pricing | Not disclosed | Released in 2024, estimated ¥50,000–60,000/card |
| Channel estimate | ≈ ¥55,000 | Estimated based on A100 pricing ratio |
References