Skip to main content

Moore Threads MTT S5000 (Domestic GPU Training)

Product Overview

Moore Threads (Moore Threads) is a Chinese full-function GPU startup company, founded in October 2020, with the founder being former NVIDIA China region executive Zhang Jianzhong. MTT S5000 is a training+inference integrated GPU AI computing card based on fourth-generation MUSA "Pinghu" architecture, with parameters publicly disclosed on 2025-02-12: single card AI compute 1000 TFLOPS, 80GB GDDR6X, 1.6 TB/s bandwidth. Equipped with self-developed MUSA unified system architecture + MUSIFY software stack.

Strategic Positioning: Compared to Huawei Ascend's focus on AI training, Moore Threads follows the full-function GPU roadmap (graphics + AI + general-purpose computing), and is a domestic GPU startup company targeting NVIDIA, listed alongside Jingjia Micro, Xinyuan Microelectronics, Enflame, and Biren as the "Domestic GPU Five Tigers".

Core Specifications

ItemParameter
ArchitectureMUSA (4th generation, Pinghu)
ProcessTSMC 6nm (estimated)
GPU Cores4096 MUSA Cores (self-developed ISA)
Memory80GB GDDR6X
Memory Bandwidth1.6 TB/s
FP3262.5 TFLOPS (estimated)
BF16 / FP16500 TFLOPS (estimated)
INT82,000 TOPS (estimated)
TDP300 W
PCIePCIe 4.0 ×16
InterconnectMUSA Link (self-developed, similar to NVLink)
Form FactorOAM / PCIe
Release2025-02-12 (parameters disclosed)
Mass Production2025-Q1 (parameters disclosed)
Unit Price (OAM)~$4,000-6,000

MTT S5000 Parameter Evolution (2024 → 2025 Version)

MetricMTT S5000MTT S4000Improvement
Process7nm12nmNew generation
Core count40962048
Memory48GB GDDR624GB GDDR6
Bandwidth700 GB/s448 GB/s1.56×
FP3225 TFLOPS12 TFLOPS2.08×
BF1650 TFLOPS24 TFLOPS2.08×
InterconnectMUSA Link 800 GB/s400 GB/s
TDP300W250W+20%

MUSA Architecture

Core Components

ComponentDescription
MUSA CoreSelf-developed SIMT core (similar to CUDA Core)
Tensor CoreSelf-developed matrix unit (similar to Tensor Core)
SFUSpecial Function Unit (transcendental functions)
RT CoreHardware ray tracing core
MUSA Link8-card full interconnect, 800 GB/s bidirectional

Differences from NVIDIA CUDA

DimensionMUSACUDA
Core ArchitectureSIMTSIMT
Instruction SetSelf-developed (similar to PTX)PTX / SASS
Thread Model32 threads / Warp32 threads / Warp
Software Stack Maturity3-4 years18 years
EcosystemMUSIFY (similar to CUDA)cuDNN / cuBLAS / NCCL
Developer Base~10K developers4M+ developers

Software Stack MUSIFY

LayerToolTargeting NVIDIA
AI FrameworkPyTorch-MUSAPyTorch + CUDA
TensorFlow-MUSATensorFlow
MindSporeMindSpore compatible
CompilerMUSA CCnvcc
RuntimeMUSA RuntimeCUDA Runtime
Math LibraryMUSBlascuBLAS
Deep Learning LibraryMUDNNcuDNN
Communication LibraryMUSA CCNCCL
Graphics APIVulkan / OpenGL / DirectXSame

⚠️ Ecosystem Limitation: MUSIFY ecosystem has only 3-4 years of development, operator coverage ~70-80% (vs CUDA 99%+), complex LLM models require extensive manual optimization or fallback to CPU.

Company Information

ItemContent
CompanyMoore Threads Intelligent Technology (Beijing) Co., Ltd.
FounderZhang Jianzhong (former NVIDIA China region GM)
Founded2020-10
Funding$500M+ (Series A 2021, Series B 2022, Series C 2023)
Valuation (2025)~¥35B
2025 Revenue~¥2.2B
HeadquartersChaoyang District, Beijing
Official Websitehttps://www.mthreads.com
StatusPreparing for STAR Market IPO (2026-2027 estimated)
Employees~2000 people
Major CustomersChina Mobile, Inspur, Lenovo, ByteDance, Zhipu AI

Product Line

Product LinePositioningRepresentative Model
MTT S SeriesData center AI trainingS5000, S4000, S3000
MTT G SeriesConsumer graphics cardMTT S80, S70, S50
MTT K SeriesWorkstation professional cardK5000, K4000
MTT E SeriesEmbedded / EdgeE3000

Key Features

  • Full-function GPU: Graphics + AI + general-purpose computing (GPGPU) + ray tracing
  • Domestic production rate 60%: HBM/memory from Samsung/SK Hynix, CPU domestic (Zhaoxin), packaging domestic
  • Multi-precision support: FP32 / FP16 / BF16 / INT8 / INT4
  • Multi-card interconnect: MUSA Link 8 cards, 800 GB/s bidirectional
  • PCIe 4.0: One generation behind PCIe 5.0
  • Drawback: Compared to NVIDIA H100 (989 BF16 TFLOPS) compute 1/20, large ecosystem gap

LLM Training Performance Reference

  • LLaMA-2 7B training: MTT S5000 8 cards ≈ H100 1/4 speed (BF16 optimized)
  • Stable Diffusion XL: MTT S5000 1 card ≈ RTX 4090 50% speed
  • Qwen 1.5 14B fine-tuning: MTT S5000 4 cards ≈ A100 60% speed
  • Inference (70B Q4): MTT S5000 1 card ≈ RTX 4090 1.2× speed (bandwidth advantage)

Application Scenarios

  • ✅ Chinese market LLM training and inference
  • ✅ Domestic production replacement projects
  • ✅ Government, state-owned enterprise AI projects
  • ✅ AI computing center construction
  • ✅ Edge AI (embedded MTT E series)
  • ✅ Graphics rendering (consumer-grade MTT G series)
  • ❌ International market
  • ❌ Top-tier frontier model training (ecosystem + compute limitations)
  • ❌ FP8 training (only supports BF16)

Domestic GPU Five Tigers

CompanyPositioningRepresentative ProductFunding
Moore ThreadsFull-function GPU + AIMTT S5000$500M+
Biren TechnologyData center AIBR104$700M+
Jingjia MicroMilitary + civilian GPUJM9Public
Xinyuan MicroelectronicsIP + design servicesMultiple IPsPublic
IluvatarData center AIMR 100/200$400M+

References