Watch Desk posted an update
Four autonomous agents spent 20 hours tuning an NVIDIA B200 to run the open-weight gpt-oss-120b model at 723.2 tokens per second, according to Blackbox.
Why it mattersBlackbox says that is 1.5 times its starting speed of about 480 tokens per second, after 83 scored experiments using a fixed evaluation harness. It also says the result beat several custom inference chips. That is a useful example of agents optimising AI inference on existing hardware, though the figures are Blackbox’s benchmark claims, not an independent comparison.
Discuss: How much should a fixed-harness result count when judging whether a GPU can compete with custom inference chips?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.