MetaX XiYun C500 (2022)
Product Overview
XiYun C500 is MetaX Integrated Circuit's first training-inference integrated general-purpose GPU released in 2022, based on self-developed XCORE 1.0 architecture, equipped with 64GB HBM2e memory, supporting FP64/FP32/TF32/FP16/BF16/INT8 multi-precision mixed computing, FP16 compute 280 TFLOPS, INT8 compute 560 TOPS, interface supports PCIe Gen5 and MetaXLink multi-card interconnection, is the first product of MetaX's "XiYun" C series.
Positioning: Training-inference integrated GPU, balancing AI training and inference scenarios, performance better than NVIDIA H20 (according to third-party benchmarks).
Core Specifications
| Item | Parameter |
|---|---|
| Architecture | Self-developed XCORE 1.0 (dozens of core IPs) |
| Process | Not disclosed (estimated 7nm) |
| FP32 | 54 TFLOPS (vector 18 + matrix 36) |
| TF32 | 140 TFLOPS |
| FP16 | 280 TFLOPS |
| BF16 | 280 TFLOPS |
| INT8 | 560 TOPS |
| Memory Capacity | 64 GB HBM2e |
| Memory Bandwidth | Not disclosed (estimated ~1.6 TB/s) |
| TDP | 350 W (estimated) |
| Interconnect | MetaXLink (7 high-speed interconnect ports, up to 64 cards interconnection) |
| Interface | PCIe Gen5 + MetaXLink |
| FP64 Support | ✅ (scientific computing/meteorological prediction) |
| Release | 2022 |
| Mass Production | Since 2023 |
| Software Stack | MXMACA (CUDA compatible, migration cost reduced by 90%) |
⚠️ Specification Note: Process, TDP, and memory bandwidth not fully disclosed by official sources, subject to MetaX's subsequent official data sheet.
XiYun C Series Product Line
| Product | Architecture | Memory | FP16 TFLOPS | Release | Status |
|---|---|---|---|---|---|
| XiYun C500 | XCORE 1.0 | 64GB HBM2e | 280 TFLOPS | 2022 | On sale |
| XiYun C550 | XCORE 1.x | Not disclosed | Not disclosed | 2024 | On sale |
| XiYun C588 | XCORE 1.x | Not disclosed | Not disclosed | 2024+ | On sale |
| XiYun C600 | XCORE 1.5 | 144GB HBM3e | FP8 1000 TFLOPS | 2025 | Risk production |
Comparison with NVIDIA H20
| Metric | XiYun C500 | NVIDIA H20 | Difference |
|---|---|---|---|
| FP16 | 280 TFLOPS | ~300 TFLOPS | -7% (close) |
| INT8 | 560 TOPS | ~600 TOPS | -7% (close) |
| Memory | 64GB HBM2e | 96GB HBM3 | -33% |
| Interconnect | MetaXLink | NVLink | TBD |
| Ecosystem | MXMACA (CUDA compatible) | CUDA | H20 mature |
| Price | ~¥38,900/card | ~¥200,000/card | C500 80% cheaper |
Third-party benchmark: According to public information, XiYun C500 series training-inference integrated GPU performancebetter than H20.
MXMACA Software Ecosystem
| Layer | Tool | Description |
|---|---|---|
| Software Stack | MXMACA | MetaX unified computing architecture |
| AI Framework | PyTorch | Native support |
| Distributed | DeepSpeed | Distributed training |
| CUDA Compatibility | Automatic migration tool | Code migration cost reduced by 90%+ |
| Large Model | Supports domestic thousand-card cluster | Full parameter training verified |
Application Scenarios
- ✅ Domestic large model training (280 TFLOPS FP16, 64GB memory)
- ✅ AI inference as a service (560 TOPS INT8)
- ✅ Scientific computing (FP64 double-precision support)
- ✅ Meteorological prediction (HPC traditional scenarios)
- ✅ Domestic AI computing center (cost-performance advantage)
- ✅ CUDA migration scenarios (90%+ migration cost reduction)
- ❌ FP8 inference (no direct FP8 format support)
- ❌ Ultra-large-scale clusters (MetaXLink TBD vs NVLink)
Related Cards
- MetaX XiYun C600 - Next generation (FP8 1000 TFLOPS)
- Kunlun P800 - Domestic AI compute
- Huawei Ascend 910C - Domestic training card competitor
- NVIDIA H20 - Compliance version competitor