Pinned post
Every "we have 600 GB of pooled VRAM" spreadsheet is fiction.
A model has to fit on ONE node before it can train on any of them. So the ceiling is your largest single GPU, not the sum.
I built a planner to compute this properly. Here's what it says. 馃У
A model has to fit on ONE node before it can train on any of them. So the ceiling is your largest single GPU, not the sum.
I built a planner to compute this properly. Here's what it says. 馃У