Specifications

Overview

Moore Threads is China’s leading general-purpose GPU vendor, built on its in-house MUSA architecture and now on its third GPU generation. For AI buyers outside the mainstream NVIDIA stack, the relevant parts are the MTT S4000 datacenter accelerator and the KUAE training cluster it powers — a fully domestic, 1,000-card AI training system.

The MTT S4000 pairs 48 GB of high-bandwidth memory at 768 GB/s with a PCIe 5.0 ×16 host interface and rates 25 TFLOPS FP32, 50 TFLOPS TF32, 200 TFLOPS FP16/BF16 and 200 TOPS INT8 at a 450 W TGP. It supports multi-modal inference and AIGC serving alongside graphics workloads, carries hardware virtualization with SR-IOV isolation and GPU elastic partitioning, and exposes a MUSA Security Engine 2.0 with TEE support. Chip-to-chip communication runs over MTLink, and the card provides 240 GB/s of I/O bandwidth with 48-channel 1080p30 encode and 96-channel decode.

At the lower end, the MTT S30 is a 40 W, slot-powered single-slot half-height card with 1,024 shaders and 4 GB of GDDR6, suited to display and light edge inference, while the MTT S80 brings 4,096 MUSA cores, 14.7 TFLOPS FP32 and 16 GB of GDDR6 over PCIe Gen5. All parts are programmed through the MUSA software stack, with tooling to ease migration from CUDA. Availability and export terms are quoted on request.