I see huge gains in Q32T1: +40% reading, +60% writing

I'm fine with a delay. They should wring every drop out of the current gen systems before moving on.

The system cannot find the path specified. 25H2   26200.7462

Quote from: not spending 10k on Yesterday at 22:29:54for inferencing, only (V)RAM size and bandwidth matter

Wrong. Different inferencing applications have different hardware needs! The inferencing application I use does not care for VRAM size or bandwidth but cares for Nvidia library versions, numbers of all types of GPU cores, GPU clock, RAM size (if I let inferencing run for a long time on the same analysed object: typically >16GB for >5min, >32GB for >80min), and only then CPU clock and possibly more than 16 CPU threads.