RobertJasiek wrote:
AFAIK, tensor cores are essential also at runtime to get fast speed and RTX 3000 is better than RTX 2000 at that.
I know nothing about the lag issue.
As to Ryzen, I consider 3700X, 3900X, some XT, 5800X, 5900X or await a still unannounced 5700X. XT may require a suitable mainboard or UEFI update presuming an earlier CPU. Ryzen 5000, too, a new mainboard launched later or a rare expensive mainboard with UEFI update without CPU. Compared to current Ryzen 3000 prices, Ryzen 5000 are somewhat expensive. I do not know how much acceleration we get from the latter. The AMD presentation gave 1080p figures +19%~+28% but these are the least meaningful for us. 4K figures are nearly identical from 3600X to 5950X in 3D gaming. Mevertheless, more cores and higher IPC should mean something for us, but how much? +5%? +20%? In go AI, the (if only one) GPU is the bottleneck.
3950X and 5950X are too expensive for my taste.
Thanks for the answer. I am fairly certain tensor cores are essential, just wanted to confirm that with an expert. The lag issue was really a big turn-off for me, I even filed a bug report for Leela, but it turned out the issue was known and not a bug. Maybe the newer generations do not suffer from that anymore... would be great, cause that is the only reason for me to use AMD cards at the moment.
From what I understood in the above linked blog from Tim Dettmer who is apparently quite an expert in the field, the CPU won't make much of a difference. All the tough computing is done on the GPU and the CPU is basically just dealing out the tasks and data (also a question to @lightvector if that is really the case). So I wouldn't invest too much into that, any newer 6-8 core will do fine as far as I understand. During analysis even my old 4-core i7 is running at 20% while GPU is under full load.
Guess I just wait a little longer, see how the prices go after AMDs GPU announcement in November, and maybe get some nice christmas deal for a 20XX or even a 30XX.