Moore Threads · Module / Kit
MTT E300
(Translation draft; official site has no standalone English E300 page) Built on Moore Threads' self-developed ChangJiang AI SoC, the MTT E300 integrates a full-function GPU, big-core CPU and energy-efficient NPU in a heterogeneous multi-core design, delivering up to 50 TOPS dense INT8 with INT8/FP16/FP32/FP64 support and stable operation from -20°C to +65°C. It supports mainstream AI frameworks, cloud-edge collaboration and lightweight deployment with a full toolchain for reliable domestic edge AI.
Key Highlights
- (Translation draft) Self-developed ChangJiang SoC combining GPU, big-core CPU and NPU
- up to 50 TOPS dense INT8
- up to 64GB LPDDR5X unified memory at up to 136.5GB/s
- wide temp range -20°C to +65°C, up to 45W
- MXM 314-pin 60×82mm module
- CUDA/MUSA compatible with Torch-MUSA and vLLM-MUSA (Ollama/llama.cpp planned)
- open-source TOP50 vision ModelZoo
- modular design for easy integration.
Technical Specifications
| SoC/Chip | ChangJiang (GPU+CPU+NPU heterogeneous) |
|---|---|
| INT8 算力/INT8 Compute | up to 50 TOPS dense |
| FP16/FP32 | 12 TFLOPS FP16 / 3 TFLOPS FP32 |
| 统一内存/Unified Memory | 16/32/64GB LPDDR5/LPDDR5X |
| 内存带宽/Memory BW | up to 136.5GB/s (LPDDR5X 8533MT/s) |
| 总线/Bus | PCIe 5.0 x14 Lane |
| 形态/Form Factor | MXM 314-pin, 60×82mm |
| 温度/Temp | -20°C~+65°C |
| 功耗/Power | up to 45W |
| 大模型上限/LLM Size | up to 7B/30B/80B with 16/32/64GB |
Typical Use Cases
(Translation draft) Embedded edge AI, robotics, smart cameras, edge gateways, industrial inspection, on-board and unmanned devices requiring low-latency domestic AI inference and on-device deployment.
官网无独立英文产品页（en 路由回退为英文首页），英文列为翻译稿；量化规格来自官网中文规格区块；图片由官网 hero banner 裁取模组 PCB 主体（本地 row 图为生态/架构示意、非硬件本体，未采用）。
Related Products
Moore Threads
MTT S5000
Universal GPU for AI Training and Inference View details →
Moore Threads
MTT S4000
Large-model AI inference and training GPU with Tensor Cores View details →
Moore Threads
MTT S3000
Full-featured server-class GPU View details →
Moore Threads