Complete side note, but I can’t work out how the author managed to mistype “at” as “Δt”.

Edit: reading fail on my part, nothing to see here.

Out of curiosity, what are you training with these cards?

Converting four RTX PRO 6000 Blackwell cards to waterblocks, finding a VRM choke loose on the workbench, and getting back to 41k tok/s.

> 4× RTX PRO 6000 Blackwell Workstation (GB202, 96 GB GDDR7, 600 W)

Those are SM120 so no tmem/tcgen05 and lack of support in main libraries (it's like everybody is focusing on B300/SM100).

For that money I'd buy a single B300, similar total AI TOPS, similar GPU bandwidth aggregated, and only 25% less total memory (probably saved in less implementation complexity), half the energy consumption...

Also by having all SMs local they have the special L1-level interconnect. SMs can collaborate on the same GEMM. And a bunch of other nice features.

Or, you know, rent it.

Ditch the tiny DC fans. Build a shroud and switch to a single ac-powered industrial blower / duct fan.

If you want ready, well engineered, water-cooled multi-GPU research workstations, my colleagues at https://comino.com build and sell them. Or you can purchase fitted waterblocks from them for many GPUs, and build your own.

You can tell this is AI slop by the horrible soldering descriptions - anyone with experience can look at that VRM and go "Oh, the solder is still in its stencil-applied state and has not flowed across the contacts at all, on either component or board. This is a reflow-in-oven issue from the manufacturer." This wasn't a cracking joint this was a poorly-done joint.

Signed, IPC-610 certified tech.

AI slop post.