Chinese GPU manufacturer 'Moore Threads' announces GPU 'MTT S4000' for AI training and data centers
On December 19, 2023, the emerging Chinese GPU manufacturer '
A large scale model exhibition of the first thousand squares of math in the head of the motor line, and the full acceleration of the development of calculation power |
https://www.mthreads.com/news/146
Nvidia's biggest Chinese competitor unveils cutting-edge new AI GPUs — Moore Threads S4000 AI GPU and Intelligent Computing Center server clusters using 1,000 of the new AI GPUs | Tom's Hardware
https://www.tomshardware.com/pc-components/gpus/nvidias-biggest-chinese-competitor-unveils-cutting-edge-new-ai-gpus-moore-threads-s4000-ai-gpu-and-intelligent- computing-center-server-clusters-using-1000-of-the-new-ai-gpus
Moore Threads introduces MTT S4000 48GB AI GPU with MTLink and zero-cost NVIDIA CUDA framework translation - VideoCardz.com
https://videocardz.com/newz/moore-threads-introduces-mtt-s4000-48gb-ai-gpu-with-mtlink-and-zero-cost-nvidia-cuda-framework-translation
Moore Threads Unveils MTT S4000 GPU: Equipped With 48 GB Memory, 200 TOPS AI Compute, Gen5 Ready
https://wccftech.com/moore-threads-mtt-s4000-gpu-48-gb-memory-200-tops-ai-gen5-ready/
Moore Threads is a GPU company founded in October 2020 by Zhang Jianzhong, former global vice president of NVIDIA. In the past, Moore Threads has partnered with many other Chinese companies, including Lenovo, to release various hardware and software.
The major feature of the recently announced 'MTT S4000' is that it is equipped with a core that uses the third generation of the proprietary architecture 'Moore Threads Unified System Architecture (MUSA).'
Below is a table comparing MTT S4000, ' MTT S3000 ' announced in November 2022, and ' MTT S2000 ' announced in April 2022.
| MTT S4000 | MTT S3000 | MTT S2000 | |
| architecture | 3rd generation MUSA | 2nd generation MUSA | 1st generation MUSA |
| Computing performance | not clear | 15.2TFLOPS | 12TFLOPS |
| memory | 48GB GDDR6 | 32GB GDDR6 | 32GB |
| form factor | dual slot passive | dual slot passive | single slot passive |
| interface | PCIe Gen5 x16 | PCIe Gen4 x16 | |
| API support | not clear | DirectX, Vulkan, OpenGL, OpenGL ES | |
| OS support | not clear | X86/ARM/LoongArch | |
| connectable display | Up to 4 DisplayPort connections | DisplayPort 1.4 (up to 8K resolution) | |
On the other hand, according to overseas media Tom's Hardware, the MTT S4000 does not match the performance of NVIDIA's `` NVIDIA Ampere Architecture '' released in 2020 and `` NVIDIA Ada Lovelace Architecture '' in 2022. Still, Tom's Hardware says, ``The MTT S4000 has a memory capacity of 48 GB and a memory bandwidth of 768 GB per second, so it will be useful for training some AI and large language models.''
Moore Threads also said about the MTT S4000, ``The unique MUSIFY tool included in the MTT S4000 is fully compatible with NVIDIA's CUDA, making it easy to migrate code written on the CUDA platform to MUSA's platform. 'This allows us to reduce migration costs and eliminate performance degradation.'
Furthermore, Moore Threads announced the 'KUAE Intelligent Computing Center', an AI training cluster equipped with approximately 1000 MTT S4000s. This KUAE Intelligent Computing Center combines the MTT S4000 with RDMA networking, distributed storage, and integrated cluster management software.
Regarding the performance of KUAE Intelligent Computing Center, Moore Threads says that when training a large-scale language model ' Aquila2 ' with approximately 70 billion parameters on KUAE Intelligent Computing Center, it takes 33 days to complete the training.
According to Moore Threads, there are no plans to sell the MTT S4000 alone, but rather to large customers interested in the KUAE Intelligent Computing Center.
◆Forum now open
A forum related to this article has been set up on the GIGAZINE official Discord server . Anyone can write freely, so please feel free to comment! If you do not have a Discord account, please create one by referring to the article explaining how to create an account!
• Discord | 'Do you think any GPU manufacturer will catch up with NVIDIA and AMD?' | GIGAZINE
https://discord.com/channels/1037961069903216680/1186960836162490378
Related Posts:
