Skip to main content

Moore Threads MTT S5000

Vendor: Moore Threads

Category: GPU Graphics Processor

Architecture: MUSA

Introduction

Moore Threads full-function GPU, using MUSA unified architecture compatible with CUDA ecosystem. Flagship MTT S5000 training-inference unified GPU is in mass production, supporting trillion-parameter large model training. MTT S4000 (FP16 100T, 48GB memory) priced at approximately 30,000 RMB. Listed on STAR Market in December 2025, first quarterly profit in Q1 2026.

Specifications

ModelComputeMemoryInterfaceTDPProcess
MTT S50001,000 TFLOPS (FP8)80GB GDDR6 (1.6 TB/s)OAM350W12nm
MTT S4000128 TFLOPS (BF16) / 256 (INT8)48GB GDDR6 (768 GB/s)PCIe 4.0450W12nm
MTT S8030 TFLOPS (FP32)16GB GDDR6PCIe 4.0255W12nm

Official Website

Visit Official Website

Driver Downloads

Linux

Windows

OS Support

WindowsLinuxmacOSAndroid

Version History

VersionRelease DateDescription
MUSA 2.02024Domestic graphics rendering + AI compute

Performance Benchmarks

ModelTaskPerformance Metric
MTT S5000Llama 2 7B Inference~40 tok/s (FP16, official data)
MTT S4000ResNet-50 Inference~5000 img/s
MTT S80Rendering/General ComputeRuns in CUDA-compatible mode

Pricing

ModelReference PriceNotes
MTT S5000Contact vendorTraining-inference flagship
MTT S4000~¥30,000Inference accelerator
MTT S80~¥3,000Consumer graphics card

Quick Installation

Linux (Ubuntu 22.04)

# 1. Install MUSA driver
sudo apt update
sudo dpkg -i mthreads-driver_*.deb

# 2. Install MUSA SDK
tar -xzf musa-sdk-*.tar.gz
cd musa-sdk && sudo ./install.sh

# 3. Verify installation
mthreads-smi

MUSA SDK is downloaded from Moore Threads Developer Center.

Code Examples

Python (MUSA PyTorch)

import torch

# MUDA is Moore Threads' CUDA-compatible backend
assert torch.cuda.is_available(), "Moore Threads GPU not found"
print(f"GPU: {torch.cuda.get_device_name(0)}")

# MUDA API is consistent with CUDA, no code changes needed
x = torch.randn(1024, 1024).cuda()
y = torch.matmul(x, x)
print(f"MUSA matrix multiply: {y.shape}")

Architecture Highlights

  • MUSA Unified Architecture: Moore Threads proprietary GPU architecture, compatible with CUDA programming model, supporting CUDA source-level porting
  • S5000 Flagship: Designed for large model training and inference, OAM interface suitable for cluster deployment
  • Dual-Track Strategy: Simultaneously advancing graphics rendering (MUSA Graphics) and AI computing (MUSA Compute)

Model Compatibility

Model/FrameworkSupportNotes
PyTorch⚠️ MUDA BackendCUDA-API compatible, requires MUSA PyTorch
TensorFlow⚠️ LimitedVia MUDA compatibility layer
Llama and similar LLMs⚠️Community adaptation in progress, large model support gradually improving
General InferenceResNet/YOLO inference is mature
Graphics RenderingDirectX/Vulkan/OpenGL support

If you're evaluating alternatives, the following products may also fit your scenario: