3 ms·
This is cool, but note that it's only enough to feed the floating point units on about 1000 consumer grade GPUs. I know cloud is all the rage and stuff, but th
by hedora 4y ago
This is cool, but note that it's only enough to feed the floating point units on about 1000 consumer grade GPUs.
I know cloud is all the rage and stuff, but the thing that really surprised me from the article is at how (relatively) slow the internet backbone is.
- hnuser123456 4y agoI'm guessing you're talking VRAM bandwidth, which is just over 1 TB/s on a 4090, while the "internet backbone" is apparently ~1 Pb/s, lowercase B, so actually only 128 4090s have the memory bandwidth to match the internet backbone. Of course, they would fill up in 0.2 ms, at only 24GB each running in parallel.
- hedora 4y agoThose are over $2K. I meant "normal" consumer grade stuff in the $200-$400 range, as opposed to "enthusiast" stuff. Either way, it's no more than a few racks of server-grade GPUs, which is probably where applications would actually want 1PBit/sec of VRAM bandwidth.