Re:AMD Launches Instinct MI300X AI Accelerator, Up to 60% Fa...
by sbininit on Wednesday December 06, 2023 @05:28PM
(#38550372)
I suspect that those who would otherwise be buying thousands of H100s would simply put in the dev work to make the software function well on AMD hardware
Re:AMD Launches Instinct MI300X AI Accelerator, Up to 60% Fa...
by idonotknowwhy on Wednesday December 06, 2023 @05:34PM
(#38550424)
I'm hoping for the same thing, but with Intel ARC.
Intel gpus are cheaper, but have even worse ML software for consumer projects
Re:AMD Launches Instinct MI300X AI Accelerator, Up to 60% Fa...
by frognumber on Wednesday December 06, 2023 @06:41PM
(#38551010)
I'm optimistic.
* Intel has a history of getting this type of software working.
* AMD doesn't.
I don't think I'll ever trust AMD after... experiences, but Intel I'm bullish on Intel.
Arc A770 seems like by far the bang for the buck once software lands. 16GB for $300 if you find a good deal. Decent performance. Stick four of them in a system, and you've got 64GB for $1200 and a pretty decent amount of compute. Buy five systems like that, and you can run the largest open-source LLMs out there with Alpa for ~$6K.
That brings Facebook's GPT3-grade OPT models from corporate down to serious enthusiast level.
I know they are "technically supported" by onyx/pytorch etc. But almost every project I've seen there's a bunch of asterisks, a bunch of people needing different work arounds, and subpar performance compared to the hardware.
Maybe it's because most people have their consumer cards? Nvidia however doesn't have this issue, people on pascal are only just now having compatibility issues with not supporting a high enough cuda version.
The dream is they release a 4090-type competitor for a few grand, it doesn't even need full 4090-level performance, just strap 48gb+ of HBM and it would be a absolutely incredible deal for individuals learning/researching -> who then use the full enterprise stuff for full training/commercial inference.