Overview

• Chinese GPU manufacturer Moore Threads has unveiled its latest model, the MTT S4000, designed for AI and data center compute workloads.
• The MTT S4000 features 48GB of video memory and 768GB/sec of video memory bandwidth, and is compatible with both x86 and Arm.
• The new GPU supports the LLaMA and GPT models, among others, and is aimed at AI services.
• Moore Threads also revealed a 'kilocard cluster' that harnesses 1,000 of its GPUs, which has been used by China's Zhiyuan Research Institute to train a 70 billion parameter model in 33 days.
• Despite Moore Threads being on the US's entity list of companies that are persona non grata, the company's MUSIFY tool allows easy migration of CUDA code to the MTT S4000, potentially attracting patriotic Chinese developers.