Moore Threads · Accelerator Card
MTT S4000
MTT S4000 is a powerful GPU for large models built on the third-generation MUSA architecture with Tensor Cores, 48GB VRAM and 768GB/s memory bandwidth. Proprietary MTLink technology supports multi-GPU interconnect for thousands-card clusters and trillion-parameter LLMs. It also delivers graphics rendering, video codec and UHD 8K HDR display, and supports smooth migration to the MUSA platform.
Key Highlights
- Advanced-generation Tensor Cores supporting FP64/FP32/TF32/FP16/BF16/INT8
- 48GB memory at 768GB/s
- MTLink supports 2/4/8-card bridging with 240GB/s inter-card I/O bandwidth
- over 91% linear speedup in thousand-card clusters
- MUSA Security Engine 2.0 with TEE
- hardware virtualization, elastic GPU partitioning and SR-IOV isolation.
Technical Specifications
| 显存/Memory | 48GB, 768GB/s |
|---|---|
| 显存带宽/Memory BW | 768GB/s |
| 卡间 I/O 带宽/Inter-card I/O BW | 240GB/s (MTLink) |
| 总线/Bus | PCIe 5.0 x16 |
| 显示/Display | 4× DP 1.4a |
| 最大功耗/Max TDP | 450W |
| 尺寸/Dimensions | 266×112×39mm |
| 精度/Precision | FP64/FP32/TF32/FP16/BF16/INT8 |
| 安全/Security | MUSA Engine 2.0, TEE |
| 编解码/Codec | up to 48ch 1080p30 encode, 96ch 1080p30 decode |
Typical Use Cases
Training, fine-tuning and inference of mainstream LLMs (LLaMA/GPT/ChatGLM/Qwen/Baichuan)single-node 8-GPU and multi-node distributed training (0.6B to 100B parameters)graphics rendering, multimedia codec and 8K HDR displaydeployed in the KUAE thousand-card cluster.
英文取自官网英文页；图片为官网产品封面图 cover.jpg；官网未公开单卡 TFLOPS/TOPS 量化算力值。
Related Products
Moore Threads
MTT S5000
Universal GPU for AI Training and Inference View details →
Moore Threads
MTT S3000
Full-featured server-class GPU View details →
Moore Threads
MTT S2000
Low-power full-featured server GPU for cloud-native workloads (translation draft) View details →
Moore Threads