AVX2 optimized SF+NNUE and processor temperature

[Gregory Owett]

In conclusion of all this, I reduced the number of cores to 12, so the temperature is 76 ° C on average now.
And thank you for all this info!

[daws]

I do not like noisy water cooler.

I can hardly hear my Corsair H115i water cooler. That's with my I9 7980XE running Stockfish AVX2 using 16 real cores ( hyperthreading off ). The cpu temperature doesn't get much above 50 C. My graphics card ( 2080 TI ) gets a bit hotter when running lc0 type engines, about 70 C but that's not watercooled. I also have a large case with plenty of fans giving good circulation and I don't overclock.

[corres]

Hi,
Do you keep your machines running all the time? Me, I do it approx. 14-16 h. max per day. I have a Ryzen 3900x, and I use 16 threads (temp. approx. 77-81 ° C, 4000 MHz according to Ryzen Master).

If you switch off SMT in BIOS you can use all 12 physical cores at lower temperatures.
Using SMT/HT it generates more heat what is surplus.

[corres]

I am running a 3970x with an air cooler (this one: plus 8 case fans), and it is running about 79-81 Celcius under full load. I have run it fully loaded for hours at a time.

I think you do not use your machine in a 30 grade Celsius room...
The temperature of room is very important!

[elpapa]

I was obliged to lessen the settings of my Ryzen 9 3950x because after some time it was frozen during test run (SF+NNUE against SF+NNUE) because of over-heating the CPU (CPU temperature was more than 90 degrees Celsius with Noctua NHD15 SE AM4 air cooler)
Now the CPU clock is 16 x 4.0 GHZ (It was 16 x 4.40 GHz with CPU Core Voltage = 1.450) and the Core Voltage is 1.400 Volt, nominally.
Power consumption of CPU was ~200 Watts, now it is about 160 Watts and now the temperature of CPU is about 75 degrees Celsius.

1.4 Volts is way too high. At 4.0 GHz 1.2 Volts should be enough. You can probably do 4.2 GHz at 1.3 Volts.

If it doesn't seem to be stable, drop the frequency or increase voltage in small steps until you're good (or vice versa if it is stable and you want to maximize performance).

You can use prime 95 with small datasets/AVX off to stress test the CPU. Passing prime 95 with AVX on is usually overkill, but at least check that it doesn't crash immediately.

[corres]

I was obliged to lessen the settings of my Ryzen 9 3950x because after some time it was frozen during test run (SF+NNUE against SF+NNUE) because of over-heating the CPU (CPU temperature was more than 90 degrees Celsius with Noctua NHD15 SE AM4 air cooler)
Now the CPU clock is 16 x 4.0 GHZ (It was 16 x 4.40 GHz with CPU Core Voltage = 1.450) and the Core Voltage is 1.400 Volt, nominally.
Power consumption of CPU was ~200 Watts, now it is about 160 Watts and now the temperature of CPU is about 75 degrees Celsius.

1.4 Volts is way too high. At 4.0 GHz 1.2 Volts should be enough. You can probably do 4.2 GHz at 1.3 Volts.

If it doesn't seem to be stable, drop the frequency or increase voltage in small steps until you're good (or vice versa if it is stable and you want to maximize performance).

You can use prime 95 with small datasets/AVX off to stress test the CPU. Passing prime 95 with AVX on is usually overkill, but at least check that it doesn't crash immediately.

For me the 16 x 4000 MHz is the most important. If I would be more lucky I may get a better CPU on "silicon lottery". But I am forced to cook what I have.
For Ryzen gen. 1 the max (official) Core Voltage is 1.450 Volts. For Ryzen gen. 2 AMD did not give "official" max Core Voltage, so I used this old value.
The 1.400 Volt is the namely Voltage, in reality it is about 1.365 Volts. I think the Core temperature (~75 Celsius) is more important than the Core Voltage.

[Alayan]

More voltage means more power consumption. The minimum voltage to achieve CPU stability depends on the target clockspeed, but you should try to have the lowest voltage that doesn't hurt stability at your target clockspeed.

Less power consumption means your temperatures will drop, too. Pumping a too high voltage might also cause long-term CPU degradation, but the limit for Zen 2 is unclear.