M ANSARI wrote: ↑Wed Sep 28, 2022 12:05 pm
In my days of overclocking we did do weird stuff to try and have better CPU to heat sink heat flow. I remember meticulously polishing and flattening the heat sink to a mirror finish so that there would be maximum efficiency of heat transfer. As for water cooling ... yes of course in time the water reservoir will reach a uniform temperature ... but the surface area of the cooling is dramatically increased as you can have a very large radiator or even 2 radiatiors. I remember once using a huge water cooler reservoir that was completely passive ... not a single fan ... and it would easily keep things under control even at max CPU utilization on all cores. Can't remember the name of that cooler but I do remember that even Vas from Rybka got one of those.
I had two of those in my system, one for CPU@65W and one + 120mm radiator+fan@1000RPM for GPU@275W.
Nowadays people buy AIO, all in one, water-cooling-systems for their CPUs, some are modular, I like that ones, with fittings and so on, so you can maintain them by yourself.
M ANSARI wrote: ↑Wed Sep 28, 2022 12:05 pm
I haven't built a new computer in like 10 years ... so am way behind the curve. I did upgrade my PC with a 2080 Ti card when Lc0 started showing some interesting chess. I feel it is time to make a nice new build. As you get older you tend to not accept stuff you would accept when you are younger. Sound for me is very important and am willing to sacrifice some performance for some peace and quiet!
Interesting: at 88W package power (65W "TDP"), the 7950X manages 31308 points in Cinebench R23 multicore, compared to 38291 at 230W [1]. That's 18% less performance at 62% less power draw. Meanwhile, Apple's M1 Ultra has 70W package power (for the CPU), i.e. roughly comparable to the 88W setting of the 7950X and gets 21732 points [2]. That's in the same ballpark, with the 7950X being a bit ahead, now that AMD also is at 5nm at TSMC.
So much for the "x86 is soooo inefficient" trope, now that AMD is also at 5nm with TSMC.
Could you recompile vnni256 with Clang15 with -O3 -march=znver3 -fomit-frame-pointer
Should see an increase.
Also I wonder how much comes from Zen4 vs generally using much faster ram compared to DDR4.
Keep in mind that DDR5 has some other improvements - its not only transfer rate.
