A server-grade, data center AI accelerator card from the Chinese company Moore Threads (founded in 2020 by the former General Manager of NVIDIA China). Announced in December 2023, the card is positioned as a solution for training and inference of Large Language Models (LLMs).
Let’s take a closer look at the accelerator’s features.
Manufacturer: Moore Threads Technology Co., Ltd. (Beijing, China). Announcement Date: December 19, 2023. Segment: Data Center / AI Accelerator. Architecture: MUSA 3rd Gen (Chunxiao core).
The MTT S4000 is part of Moore Threads’ comprehensive KUAE data center platform. This platform integrates hardware servers (e.g., MCCX D800), high-speed MTLink interconnects (analogous to NVLink), distributed storage, and cluster management software.
Within the KUAE ecosystem, support for popular AI frameworks is implemented, along with the MUSIFY tool, which allows adapting code written for NVIDIA CUDA to the MUSA architecture. Thanks to this, the MTT S4000 accelerator can be used for training and inference of large language models with minimal migration costs, and it integrates into scalable clusters for trillion-parameter tasks. The company emphasizes the rapid deployment and high efficiency of corporate AI infrastructures, offering turnkey solutions for large clients.
Technical Specifications: Moore Threads MTT S4000 48GB
Architecture: MUSA Gen 3 (Chunxiao core)
Interface: PCIe 5.0 x16
Process Node: 12nm; Transistors: 22 billion; Memory Bus: 384-bit
Core Clock: 1500 MHz; Memory Clock: 2000 MHz (16 Gbps effective)
Compute Cores: 8192
Memory: 48 GB GDDR6
Tensor Cores: 128
TMU / ROP: 512 / 512
Memory Bandwidth: 768 GB/s
FP32 Performance: 25 TFLOPS
TF32 Performance: 50 TFLOPS
FP16 / BF16 Performance: 100 TFLOPS
INT8 Performance: 200 TOPS
L2 Cache: 4 MB
Simultaneous Streams: Up to 96x 1080p
Form Factor: 2-slot, 266mm length
Interconnect: MTLink 1.0
TDP: 450 W; Power: 2x 8-pin (12VHPWR)
Cooling: Passive
Video Outputs: 4x Display (up to 8K)
Codecs: Hardware video encode/decode
Accelerator Architecture Features
MUSA 3rd Gen Architecture MUSA (Moore Threads Unified System Architecture) is the company’s proprietary architecture, now in its third generation. The Chunxiao core is manufactured on a 12nm process and contains 22 billion transistors.
Key architectural decisions:
8192 compute cores — significantly more than the S3000 predecessor.
128 Tensor Cores — hardware blocks for accelerating matrix operations in AI workloads.
512 TMU + 512 ROP — texture and raster units.
4 MB L2 cache.
Precision support: FP32, TF32, FP16, BF16, INT8, INT4.
The MUSA 3.0 architecture and Chunxiao chip implement not only high performance but also a comprehensive set of software-hardware solutions for corporate and cloud AI infrastructures. A key element of the ecosystem is the MUSA SDK software stack, which includes muDNN (analogous to cuDNN), muFFT libraries, and MCCL — a library for collective GPU communications, which is critical for scaling and cluster training.
MTLink 1.0 Interconnect Allows combining multiple S4000 cards into a single compute pool with high-speed connectivity. Supports scaling up to 10,000 GPUs in a single cluster.
MUSIFY CUDA Compatibility To simplify migration from NVIDIA CUDA, the company offers MUSIFY, torch_musa, and musa_converter tools, which allow launching PyTorch projects with minimal changes and automatically translating CUDA code into MUSA calls. The MUSIFY toolkit ensures code translation from NVIDIA CUDA to MUSA. This is a critical feature: developers can port existing CUDA applications to the S4000 without rewriting the code entirely. According to Moore Threads, the migration cost is zero.
Furthermore, the accelerator supports hardware virtualization (SR-IOV), a trusted execution environment (TEE), and MUSA Safety Engine 2.0 security mechanisms, making it suitable for cloud and corporate tasks with high requirements for isolation and data protection. All of this underscores Moore Threads’ strategic focus on creating a full-fledged, independent, and scalable AI platform capable of competing with Western analogues not only in hardware but also in the maturity of its software ecosystem.
Passive Cooling The card features passive cooling, which is standard for server GPUs. Cooling is provided by the server chassis.
Preliminary Performance Assessment: Moore Threads MTT S4000 48GB
Memory: 48 GB GDDR6; Bandwidth: 768 GB/s
FP32: 25 TFLOPS; FP16: 100 TFLOPS; INT8: 200 TOPS
TDP: 450W; Process: 12nm
Analogues:
NVIDIA A100 80GB: 80 GB HBM2e; 2039 GB/s; 19.5 TFLOPS FP32; 312 TFLOPS* FP16; 624 TOPS* INT8; 400W; 7nm.
NVIDIA H100 SXM: 80 GB HBM3; 3350 GB/s; 67 TFLOPS FP32; 990 TFLOPS* FP16; 1979 TOPS* INT8; 700W; 4nm.
Huawei Ascend 910C: 64 GB HBM2e; ~1600 GB/s; ~32 TFLOPS FP32; ~640 TFLOPS FP16; ~1280 TOPS INT8; 350W; 7nm. * with Tensor Core / Sparse
Conclusion on specs: The S4000 significantly lags behind the H100 and is roughly comparable to the A100 on certain FP32 metrics, but loses to the A100 in memory bandwidth (768 vs 2039 GB/s) and AI Tensor Core operations. However, for the market of alternative Chinese GPUs, this is one of the most mature solutions.
Practical Accelerator Benchmarks
Kua’e Qianka (KUAE) — LLM Cluster Training Moore Threads demonstrated training a 3-billion parameter model on a KUAE cluster based on S4000. According to Tom’s Hardware (May 2024), the results were comparable to unspecified NVIDIA solutions — presumably A100. The card took 3rd place in AI testing, outperforming some NVIDIA GPU clusters.
DeepSeek-R1 — Inference In February 2026, Moore Threads reported the successful deployment of the DeepSeek-R1-Distill-Qwen-7B model on the S4000 (and even on the consumer S80). Rating — excellent inference performance (Tom’s Hardware). This is important: the S4000 supports fresh open-source models.
Features of Moore Threads MTT S4000 48GB
PCIe 5.0: The only one among Chinese GPU competitors at the time of announcement.
MTLink 1.0: Scaling up to 10,000 GPUs in a cluster.
MUSIFY (CUDA compatibility): A critical factor for porting existing AI pipelines.
48 GB GDDR6: Sufficient for inference of 7B–30B parameter models (with quantization).
Passive Cooling: Standard for server deployments.
Video Outputs: An unusual capability for an AI card to handle graphics (up to 8K, 96 streams of 1080p).
Full-fledged Ecosystem: MCCX D800 servers, KUAE cluster platform.
Use Cases
Optimal Scenarios:
LLM Inference 7B–30B (★★★★☆): 48 GB is enough for most models in this range (FP16/INT8).
Fine-tuning (LoRA/QLoRA) (★★★☆☆): Supported by the MUSA + MUSIFY ecosystem.
LLM Training (distributed) (★★★☆☆): Works in the KUAE cluster, but slower than NVIDIA.
Computer Vision (★★★☆☆): CV model inference, image processing.
Video Processing (★★★★☆): 96 streams of 1080p — a strong point of the card.
Cloud VDI (★★★★☆): Video outputs + GPU acceleration for virtual desktops.
HPC (Scientific Computing) (★★☆☆☆): FP32/FP64 performance is limited.
Graphics Rendering (★★★☆☆): 8K output, DirectX 12 Ultimate.
Recommended models for inference on MTT S4000:
LLaMA 2 7B/13B (FP16)
Qwen 7B/14B
ChatGLM 3/4 6B
DeepSeek-R1-Distill 7B/14B
Baichuan 2 7B/13B
With INT8/GPTQ quantization: models up to 30B parameters.
Accelerator Testing Conducted
1. KUAE Cluster — 3B LLM Training (2024) Moore Threads demonstrated a run on the Kua’e Qianka (千卡集群) cluster, consisting of 1000 S4000 cards (125 MCCX D800 servers). Result: The card showed performance comparable to the NVIDIA A100 in an LLM training task. Tom’s Hardware noted that the S4000 took third place in AI testing, ahead of some NVIDIA clusters.
2. DeepSeek-R1 Inference (February 2026) Moore Threads demonstrated excellent performance in DeepSeek R1 Distill Qwen 7B inference on both the S4000 and the consumer S80. This is the first public example of a modern reasoning model running on Moore Threads Chinese GPUs.
Projects and Deployments
KUAE AI Cluster Lineup KUAE (快算) is Moore Threads’ flagship product based on the S4000:
KUAE Kilocard Cluster — 1000 GPU (125 MCCX D800 servers)
KUAE Myriacard Cluster — 10,000 GPU (1250 servers)
Each MCCX D800 server: 8x S4000, 2x Xeon Gen 4, 1 TB RAM, 15.36 TB NVMe, 2x 400 Gb/s network.
Model Adaptation According to GamingDeputy and a Moore Threads report, the company has completed three strategic project signings for cluster deployment:
¥660 million (~$90 million) contract for a 10,000 GPU KUAE cluster (Q1 2026).
Deployments in several Chinese provinces.
DeepSeek-R1 — inference confirmed (February 2026).
ChatGLM (智谱 AI) — support via the MUSA ecosystem.
Qwen (Alibaba) — compatibility confirmed.
LLaMA — via MUSIFY/CUDA translation.
Card Rating Based on the information gathered, ChaiTex assigns the MTT S4000 the following ratings:
Performance (★★★☆☆): Comparable to A100 in specific tasks, but without HBM.
Memory (★★★★☆): 48 GB is good for inference, low for training.
Scalability (★★★★☆): MTLink + clusters up to 10K GPU.
Software Stack (★★★☆☆): MUSIFY works, but doesn’t cover all CUDA.
Open-source Ecosystem (★★☆☆☆): Exists, but the community is small.
Production Readiness (★★★☆☆): Contracts exist, but few independent verifications.
Market Availability (★★★★☆): Available through Chinese suppliers.
Energy Efficiency (★★☆☆☆): 12nm is a serious drawback.
Overall Score: 3.0 / 5
The accelerator is available for pre-order in the Chaitex catalog.
Comparison within the Moore Threads lineup:
MTT S3000 (32 GB) — outdated, for light tasks.
MTT S4000 (48 GB) — the main workhorse for AI.
MTT S5000 (80 GB, Pinghu architecture) — the new flagship, 1000 TFLOPS, displacing the S4000 from the high-end segment.
Card Energy Efficiency
Moore Threads MTT S4000 48GB: TDP: 450W; FP32/Watt: 0.056 TFLOPS/W; INT8/Watt: 0.44 TOPS/W; Process: 12nm.
NVIDIA A100 80GB: TDP: 400W; FP32/Watt: 0.049 TFLOPS/W; INT8/Watt: 1.56 TOPS/W; Process: 7nm.
NVIDIA H100 SXM: TDP: 700W; FP32/Watt: 0.096 TFLOPS/W; INT8/Watt: 2.83 TOPS/W; Process: 4nm.
Conclusions and Summary
The Moore Threads MTT S4000 is a critically important element of the Chinese GPU ecosystem, but with existing compromises.
What works:
For LLM inference 7B–30B: The S4000 is a working tool. 48 GB of GDDR6 is sufficient, the MUSA ecosystem supports main Chinese models (DeepSeek, Qwen, ChatGLM), and MUSIFY provides partial CUDA compatibility.
For large-scale deployments: The KUAE cluster platform with MTLink allows building clusters of up to 10,000 GPUs. These are real, not theoretical solutions — contracts worth hundreds of millions of yuan have been signed.
For the Russian market: The S4000 is available through Chinese distributors, is not subject to US export restrictions, and has a mature (albeit not ideal) software platform.
The MTT S4000 is suitable as a primary card for AI inference and fine-tuning in Russian data centers where NVIDIA A100/H100 are unavailable. The recommended configuration is MCCX D800 servers (8 cards), which provides 384 GB of pooled GPU memory. For tasks requiring higher compute density, the MTT S5000 (80 GB, Pinghu architecture, 1000 TFLOPS) should be considered — the new flagship, which is already available for order.
The Moore Threads MTT S4000 is one of the most mature and technologically rich Chinese AI accelerators to emerge amid global restrictions on Western solutions. The card stands out with its substantial memory capacity (48 GB GDDR6), support for modern compute formats, developed MTLink cluster infrastructure, and its own MUSA software ecosystem, which includes CUDA migration tools (MUSIFY), AI libraries, and virtualization means. Despite lagging behind NVIDIA flagships in several technical parameters (memory bandwidth, energy efficiency, Tensor operation performance), the S4000 represents a real and economically viable alternative for inference and fine-tuning of medium and large language models, as well as for building scalable AI clusters in Russia and China. It is especially important that the S4000 is already being deployed in real projects, compatibility with leading Chinese and international open-source models is confirmed, and the software stack is actively developing.
In the current environment, the Moore Threads MTT S4000 is the optimal choice for Russian data centers and corporate clients who cannot access NVIDIA accelerators. The card is well-suited for inference and fine-tuning of models up to 30 billion parameters, as well as for building own cluster solutions based on the KUAE platform. For tasks requiring maximum performance and energy efficiency, it is recommended to monitor the release of new models (e.g., S5000); however, in terms of the totality of characteristics, ecosystem maturity, and availability, the S4000 is currently one of the best options on the market of alternative AI accelerators.
Contact our experts to select the optimal hardware configuration and calculate rental pricing for your AI workloads.