With the power being 70W from the connector only, how feasible is it to have 3 per server and have effectively 48GB VRAM for tasks?
[1] https://www.maxsun.com/products/intel-arc-pro-b60-dual-48g-t...
Kind of. It's more two 24gb b60s in a trenchcoat. It connects to one slot but it's two completely separate gpus and requires the board to support pcie bifurcation.
And lanes. My board has two PCIe x16 slots fed by the CPU, but if I use both they'll only get x8 lanes each. Thus if I plugged two of these in there, I'd still only have two working GPUs, not four.
The biggest Deepseek V2 models would just fit, as would some of the giant Meta open source models. Those have rather pleasant performance.
In theory, how feasible is that?
I feel like the software stack might be like a Jenga tower. And PCIe limitations might hit pretty hard.
> "Because 48GB is for spreadsheets, feed your rendering beast with a buffet of VRAM."
Edit: I guess must just be a bad translation or sloppy copywriting, and they mean it's not for just spreadsheets rather than it is...