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 memory, 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
| Item | Parameter |
|---|
| Architecture | MUSA (4th generation, Pinghu) |
| Process | TSMC 6nm (estimated) |
| GPU Cores | 4096 MUSA Cores (self-developed ISA) |
| Memory | 80GB memory |
| Memory Bandwidth | 1.6 TB/s |
| FP32 | 62.5 TFLOPS (estimated) |
| BF16 / FP16 | 500 TFLOPS (estimated) |
| INT8 | 2,000 TOPS (estimated) |
| TDP | 300 W |
| PCIe | PCIe 4.0 ×16 |
| Interconnect | MUSA Link (self-developed, similar to NVLink) |
| Form Factor | OAM / PCIe |
| Release | 2025-02-12 (parameters disclosed) |
| Mass Production | 2025-Q1 (parameters disclosed) |
| Unit Price (OAM) | ~$4,000-6,000 |
📌 Data correction (2026-08 online verification):
- Memory type: This site previously recorded "GDDR6X", which contradicts the 1.6 TB/s bandwidth (GDDR6X caps at ~1 TB/s, and 80GB is not a typical GDDR configuration). Moore Threads officially announced only "80GB memory, 1.6 TB/s bandwidth" without disclosing the chip type. Now labeled as reported.
- Compute precision: The official "1000 TFLOPS dense per-card AI compute" refers specifically to FP8 precision (liquid-cooled version), not FP16. FP16 is about 500 TFLOPS. The precision was previously unlabeled, easily misread as FP16 performance and overestimated.
- Interconnect: Officially MTLink 784 GB/s (previously recorded as "MUSA Link" with no figure).
- Compute figures re-verified (2026-09 cross-validation): The official datasheet figures are BF16/FP16 400 TFLOPS, FP32 100 TFLOPS, INT8 800 TOPS, TDP 500W (market-reported); the previously listed 500/62.5/2000 extrapolated from FP8 at 1:2, and 300W, are deprecated.
MTT S5000 Parameter Evolution (2024 → 2025 Version)
| Metric | MTT S5000 | MTT S4000 | Improvement |
|---|
| Process | 7nm | 12nm | New generation |
| Core count | 4096 | 2048 | 2× |
| Memory | 48GB GDDR6 | 24GB GDDR6 | 2× |
| Bandwidth | 700 GB/s | 448 GB/s | 1.56× |
| FP32 | 25 TFLOPS | 12 TFLOPS | 2.08× |
| BF16 | 50 TFLOPS | 24 TFLOPS | 2.08× |
| Interconnect | MUSA Link 800 GB/s | 400 GB/s | 2× |
| TDP | 300W | 250W | +20% |
MUSA Architecture
Core Components
| Component | Description |
|---|
| MUSA Core | Self-developed SIMT core (similar to CUDA Core) |
| Tensor Core | Self-developed matrix unit (similar to Tensor Core) |
| SFU | Special Function Unit (transcendental functions) |
| RT Core | Hardware ray tracing core |
| MUSA Link | 8-card full interconnect, 800 GB/s bidirectional |
Differences from NVIDIA CUDA
| Dimension | MUSA | CUDA |
|---|
| Core Architecture | SIMT | SIMT |
| Instruction Set | Self-developed (similar to PTX) | PTX / SASS |
| Thread Model | 32 threads / Warp | 32 threads / Warp |
| Software Stack Maturity | 3-4 years | 18 years |
| Ecosystem | MUSIFY (similar to CUDA) | cuDNN / cuBLAS / NCCL |
| Developer Base | ~10K developers | 4M+ developers |
Software Stack MUSIFY
| Layer | Tool | Targeting NVIDIA |
|---|
| AI Framework | PyTorch-MUSA | PyTorch + CUDA |
| TensorFlow-MUSA | TensorFlow |
| MindSpore | MindSpore compatible |
| Compiler | MUSA CC | nvcc |
| Runtime | MUSA Runtime | CUDA Runtime |
| Math Library | MUSBlas | cuBLAS |
| Deep Learning Library | MUDNN | cuDNN |
| Communication Library | MUSA CC | NCCL |
| Graphics API | Vulkan / OpenGL / DirectX | Same |
⚠️ 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.
| Item | Content |
|---|
| Company | Moore Threads Intelligent Technology (Beijing) Co., Ltd. |
| Founder | Zhang Jianzhong (former NVIDIA China region GM) |
| Founded | 2020-10 |
| Funding | $500M+ (Series A 2021, Series B 2022, Series C 2023) |
| Valuation (2025) | ~¥35B |
| 2025 Revenue | ~¥2.2B |
| Headquarters | Chaoyang District, Beijing |
| Official Website | https://www.mthreads.com |
| Status | Preparing for STAR Market IPO (2026-2027 estimated) |
| Employees | ~2000 people |
| Major Customers | China Mobile, Inspur, Lenovo, ByteDance, Zhipu AI |
Product Line
| Product Line | Positioning | Representative Model |
|---|
| MTT S Series | Data center AI training | S5000, S4000, S3000 |
| MTT G Series | Consumer graphics card | MTT S80, S70, S50 |
| MTT K Series | Workstation professional card | K5000, K4000 |
| MTT E Series | Embedded / Edge | E3000 |
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
- 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
| Company | Positioning | Representative Product | Funding |
|---|
| Moore Threads | Full-function GPU + AI | MTT S5000 | $500M+ |
| Biren Technology | Data center AI | BR104 | $700M+ |
| Jingjia Micro | Military + civilian GPU | JM9 | Public |
| Xinyuan Microelectronics | IP + design services | Multiple IPs | Public |
| Iluvatar | Data center AI | MR 100/200 | $400M+ |
References