MetaX XiSi N100
MetaX's first AI inference GPU, equipped with MXN100 heterogeneous GPGPU processor, mass-produced in June 2023, INT8 compute 160 TOPS, targeting cloud data center AI inference scenarios.
Core Specifications
| Specification | Value |
|---|---|
| Architecture | MXN100 heterogeneous GPGPU (MetaX self-developed) |
| Process | 7 nm (TSMC, estimated) |
| TDP | 75 W (estimated) |
| Memory | 8 GB HBM2E |
| Memory Bandwidth | 410 GB/s (estimated) |
| FP32 Compute | Not disclosed |
| FP16 Compute | 80 TFLOPS |
| INT8 Compute | 160 TOPS |
| Interface | PCIe 4.0 x16 (estimated) |
| Video Processing | 128-channel encoding / 96-channel decoding, up to 8K |
| Launch Date | 2023-06-07 (mass production) |
| Price | Not disclosed (estimated ¥15,000) |
Technical Highlights
- MXN100 architecture: MetaX self-developed heterogeneous GPGPU processor, optimized for AI inference, supporting mixed-precision computing
- High-density video processing: Supports 128-channel encoding and 96-channel decoding, compatible with HEVC, H.264, AV1, AVS2 and other video formats, up to 8K resolution support
- MXMACA software stack: Companion self-developed software stack, supporting various mainstream AI frameworks, computer vision processing and multimedia processing frameworks
- HBM2E memory: Uses high-bandwidth HBM2E memory, meeting memory bandwidth requirements for large language model inference
- Mass production achieved: Announced mass production in June 2023, applied in security, smart city and other fields
Product Positioning
XiSi N100 is MetaX's first AI inference GPU, targeting NVIDIA L4/T4 inference cards, focusing on cloud AI inference scenarios. As the inference-specialized product in MetaX's product matrix, N100 complements XiYun C series (training-inference integrated).
Application Scenarios
- Smart city/security AI inference (video analytics)
- Cloud CV/NLP inference services
- Intelligent video processing (transcoding, analytics)
- Edge inference servers
- Large model inference acceleration (INT8 quantization)
Reference Price
| Channel | Price | Description |
|---|---|---|
| Official pricing | Not disclosed | Mass-produced in 2023, estimated ¥10,000–20,000/card |
| Channel estimate | ≈ ¥15,000 | Estimated based on L4 pricing ratio |
Related Products
- XiYun C500 — MetaX training-inference integrated GPU
- XiYun C600 — MetaX fully domestic general-purpose GPU