Mthreads Expands AI Product to Ten Thousands of Cards Scale with Core Characteristics
TapTechNews July 3rd, Mthreads announced today that its AI flagship product, KUAE Intelligent Computing Cluster Solution, has expanded from the current thousands of cards level to the ten thousands of cards scale.
The Mthreads KUAE ten thousands of cards intelligent computing cluster, with the all-functional GPU as the base, builds a domestic general acceleration computing platform capable of carrying ten thousands of cards scale and having ten thousands of P-level floating-point computing power, specifically designed for the training of complex large models with trillions of parameters.
The KUAE ten thousands of cards intelligent computing solution has the following core characteristics:
Ten thousands of cards and ten thousands of P: The KUAE intelligent computing cluster achieves a single cluster scale of over ten thousands of cards, and the floating-point computing ability reaches 10 Exa-Flops, achieving a PB-level total capacity of super-large video memory, a PB-level ultra-high speed inter-card interconnect total bandwidth per second, and a PB-level ultra-high speed node interconnect total bandwidth per second.
Long and stable training: The Mthreads KUAE ten thousands of cards cluster has an average trouble-free operation time of more than 15 days, and can longest achieve stable training of large models for more than 30 days. The weekly average training efficiency is above 99%, far exceeding the industry average level.
High MFU: Through a series of optimizations at the system software, framework, and algorithm levels in the KUAE ten thousands of cards cluster, efficient training of large models is achieved, and the MFU (a common indicator for evaluating the efficiency of large model training) can reach up to 60%.
Eco-friendly: It can accelerate different architectures and modalities of large models such as LLM, MoE, multimodal, Mamba. Based on the MUSA programming language, the complete compatibility with CUDA capabilities and the automated migration tool Musify, it accelerates the Day0 level migration of new models.
TapTechNews learned that Mthreads will carry out three ten thousands of cards cluster projects, namely the ten thousands of cards cluster project in Qinghai Zero Carbon Industrial Park, the KUAE ten thousands of cards cluster project in Qinghai Plateau, and the ten thousands of cards cluster project in Guangxi ASEAN.