Modern inference stacks use different GPUs for decode versus prefill. This is a GPU suitable for prefill.
alt Hacker News