You are right. We compared vs 3 year old H100, not Blackwell. Still more expensive than four macbooks we used. Our biggest weakness is that we are comparing the max throughput against the published one-request number. We are waiting on M5 Ultra machines, and we hope to acquire H100/H200/B200 so we can do a better "Apple to NVIDIA" test. I hope our results are still impressive on their own?
Thanks, Kevin! We are hoping to get noticed, so we can buy more Macs and NVIDIA/AMD hardware. We created a practical Pipeline Parallelism for training and inferencing. But as of now, no one knows about it. So gotta do something. Posting on HN is one thing a person like me without a network can do.
5 comments
[ 0.27 ms ] story [ 3.4 ms ] threadComparing 2022 hardware to 2025 hardware ?
No mention of power consumption or cost ?
Comparing single requests vs multiple requests ?
How is this "apples to apples" ?
My overall point is about value.
A faster configuration for more $$$ might be valuable to a company, but out of reach for an individual.
A slower system, on older hardware, might cost 1/4th, and still achieve 90% ( im thinking of the RTX 3XXX/4XXX gpu unlocks ).
So, converting all this to dollars and power provides perspective on value.
Keep up the good work. Keep kicking the ball downfield and we all win.