2026-04-30 12:24 UTC
Replies (1)
-
@lina@vt.social 2026-04-30 15:59
@ben The hardware is completely different. LLMs et al run (both training and inference) on GPU clusters (or fancy TPUs). They do extremely wide matrix multiplication that is amenable to GPU parallelization. What Valve has there is a dense CPU cluster (I don't think you can even put GPUs into blades like that). I haven't watched the video, but my intuition is most of Valve's compute there is going into the parallelism of handling many matches at once (which GPUs cannot handle, as GPUs can only effectively parallelize the *same* operation many times, while individual games don't run in lock-step). The actual machine learning model is probably very simple and lightweight for a single instance relative to LLMs (hence why it can just train on a random workstation) and therefore itself wouldn't benefit from GPU parallelism much. Re floor footprint, still, one 19" rack has more footprint than two towers. It's about 60cm by 100cm (200cm tall). That's at least 4 very chunky towers, easily 6 more reasonable sized ones. And that's a single rack (Valve has two).