Moore Threads Unveils MTT S4000 GPU: Equipped With 48 GB Memory, 200 TOPS AI Compute, Gen5 Ready

Moore Threads, the Chinese GPU maker, has unveiled its brand new MTT S4000 GPU which offers 200 TOPs of AI compute & 48 GB memory for LLMs.

At an event, Moore Threads introduced its brand new MTT S4000 GPU which uses its third Gen MUSA core architecture. The card is specifically designed to power AI workloads and offers very large memory capacities for Large Language Models.

Coming to the specifications, the Moore Threads MTT S4000 features 48 GB of GDDR6 memory clocked at 16 Gbps to provide 768 GB/s bandwidth. The GPU comes with the latest MTLink 1.0 interface technology which allows customers to run multiple cards simultaneously. Think of it as an NVLINK solution for Moore Threads GPUs. The card is also based on the PCIe Gen5 protocol & the company is so far the only one that is offering consumer-level hardware with Gen5 compliance.

According to the company itself, some of the compute figures shared showcase 25 TFLOPs of FP32, 50 TFLOPs of TF32, 100 TFLOPs of FP16/BF16, and 200 TOPS on INT8 performance. That's 5x more than the fastest NPU+CPU+GPU combination available on AI PCs such as AMD's Ryzen 8040 series and Intel Core Ultra series. Unfortunately, the core count and other technical aspects have not been shared by the company.

One interesting thing to note is that despite being an AI accelerator card, the MTT S4000 does have four display outputs and can support up to 8K displays. The card also supports 96 simultaneous 1080p streams and is outfitted with the latest USIFY development tools that can make full use of NVIDIA's CUDA-based software. Summing up the specs:

The card itself comes in a standard two-slot passive-cooled solution and makes use of a 12VHPWR power connector to boot. For comparison, the previous-gen MTT S3000 offers 32 GB of memory & peak FP32 compute of 15.2 TFLOPs. So that's a 50% increase in memory capacity and 64% increase in FP32 compute capabilities.

The Moore Threads MTT S4000 GPU is also being integrated into the KUAE computing solutions that are similar to NVIDIA's DGX systems. The Kuae MCCX D800 system makes use of 8 MTT S4000 GPUs and also comes with seamless expansion from a single machine to multiple cards and multiple AI systems. Both MTT S4000 GPUs and the Kuae systems support the latest LLMs such as LLaMA, GLM, Aquila, Baichuan, GPT, Bloom, Yuyan, and can handle 130 Billion parameters with ease. The first 1000 MTT S4000 GPUs will be housed within China's first large-scale computing cluster to power AI workloads.

Just like the other two Moore Threads GPUs, the MTT S80 and MTT S70, the MTT S4000 is expected to be available at very competitive prices. The performance on the other hand is something that we will only really know about once actual results become available since the first two cards were very underwhelming despite the company boasting a lot about their gaming performance capabilities.