Your home ESX server lab hardware specs?

[applegrcoug]

for AI tasks the bandwidth in the PCI-E slot is not such a big deal.

Months ago I ran a test looking at that and if placing cards on the impact of mining risers.

3x3090s all at pcie3 x4 (32GT/s)took 0:34 to load llama 3. I get 13.6 t/s

3x3090 all at pcie1 x4 (8GT/s) took 1:13 to load llama 3. I get 13.5 t/s

If I were to run a mining riser, it would (most likely) run at pcie3 x1, which works out to 8GT/s.

Dropping back to only one lane gimps performance to load the model into the VRAM which makes sense. If the model were loading from a spinner, then it really wouldn't matter.

Once it is loaded, it gimps it a little bit, but not enough to matter.