3 ms·
No surprise. I imagine they'll just buy older and/or more stable stock. Worked on a three 50U rack recently with quad T4's per 2U and we're not even through 10
by iszomer 4y ago
No surprise. I imagine they'll just buy older and/or more stable stock.
Worked on a three 50U rack recently with quad T4's per 2U and we're not even through 10% of that order yet.
- totoglazer 4y agoT4 is great for some stuff. And total trash for many other things you might want A100/H100 for.
- iszomer 4y agoThat wasn't my point. We build servers, test them, and ship them (in the US). No fucks given on the technical superiority of the `A100/H100` than parts that may be a generation older.
- ShamelessC 4y agoYeah I mean, great for running inference (and obviously training smaller models). The problem is that Scaling Laws for Transformers show us that scaling model parameter count _up_, is a sure way to improve the model's performance over your task. So when you're playing nation-state hard-ball, since the architectures are _somewhat_ trivially copied upon being published, that means the best checkpoint goes to the one with the fastest GPU's. If China wasn't behind in deep learning, they will be.