1 / 5
| Product Introduction: | The Moore Thread MTT S4000 is a full-featured elemental computing card based on the Qu Yuan architecture, designed for training, inference, and hybrid computing scenarios of large-scale language models at the hundred-billion parameter level. |
| Name: | Moore Threads MTT S4000 48GB |
| INT8: | High Performance200 TOPS |
| FP16/FP32: | 25 TFLOPS |
| TF32: | 50 TFLOPS |
| VRAM type: | HBM2E (High Bandwidth Memory, with approximately 30% higher bandwidth compared to HBM2, suitable for the high bandwidth demands of large model inference) |
| Chip fabrication process: | TSMC 7nm process, paired with 2.5D CoWoS advanced packaging (ensuring core heat dissipation and signal stability) |
| DirectX support: | Does not support DirectX or OpenGL for non-graphics rendering positioning, no game/graphics related API adaptation |
| Video memory frequency: | Effective frequency 9.375Gbps. By combining the HBM2 memory type with a 1.2TB/s bandwidth, the typical frequency of HBM2 memory in the same industry level is inferred. |
| Memory Bus Width: | 1024bit Based on a memory bandwidth of 1.2TB/s and an HBM2 effective frequency of approximately 9.375Gbps (bandwidth = bus width × frequency ÷ 8), it complies with the HBM2 specification characteristics. |
| Video memory speed: | 7ns (based on the physical characteristics of HBM2E memory chips, corresponding to an equivalent frequency of 4000 MHz, with latency better than GDDR6) |
| Output Interface: | No traditional video output interfaces (DP/HDMI/VGA), focused on data center inference/computing scenarios |
| Output interface type: | The core interface is PCIe Gen4.0 x16 (bidirectional bandwidth 64GB/s), supporting x86/ARM host interconnection. |
| Form Factor: | FHFL dual - slot |
| Weight: | 1064g |
Designed for 24x7 enterprise data center operations, it features enterprise-grade components and energy-efficient hardware optimized for deployment at scale.
Supports Secure Boot with Hardware Root of Trust technology, provides an extra layer of security, and meets NEBS Level 3 standards to meet new data center standards.
The passively cooled, energy-efficient, dual-socket design fits a wide range of Tesla-Certified systems from OEM vendors and is easy to integrate into existing data center infrastructure.