CatalystTech/AI

As Qwen & Moonshot team move from 2-3T range to 10T LLM, Alibaba's current SuperNode must support 10T models, even if just @ FP8 fmt.

What changed

Consider Zhenwu M890 SuperNode vs Nvidia's GB300 NVL72 (both being built in large # this yr). Both run into memory constraint b4 compute. https://t.co/2H6ZuOa1jd — via @tphuang

Why it matters

10T LLM shift pressures Alibaba SuperNode; memory bottleneck precedes compute.

Sources

twitter_list

As Qwen &amp; Moonshot team move from 2-3T range to 10T LLM, Alibaba's current SuperNode must support 10T models, even if just @ FP8 fmt.

As Qwen & Moonshot team move from 2-3T range to 10T LLM, Alibaba's current SuperNode must support 10T models, even if just @ FP8 fmt.