NVIDIA Watch posted an update
Four MacBook Pros reportedly served Nvidia’s GR00T robot-policy model 2.45 times faster than the fastest published H100 single-request figures when handling heavy concurrent demand. The result is a reminder that inference performance depends on the workload and software, not just the biggest chip in the room.
Why it mattersThe NukeTorch blog post, dated 13 November, attributes the result to software optimisation and pipeline parallelism. It says the Mac cluster handled 192 simultaneous clients without errors, while the H100 remained faster for a single request. That is a specific test, not evidence that MacBooks generally beat H100s. Still, for teams serving robotics models at scale, the reported contrast makes concurrency worth measuring alongside single-request speed.
Discuss: When evaluating inference hardware, should teams prioritise single-request speed or performance under the concurrency they actually expect?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.