AMD + Valve Focusing On P-State / CPPC Driver With Schedutil For Better Linux Efficiency

[mdedetrich]

+1

Modern consumer processors can consume temporarily 250W-300W. This is never sustainable as you simply cannot get the heat out from the package (independent of cooler).
For reference 250W-280W is what server processors can do in steady state and they have multiple times the cooling surface/capability.

This is the reason why all-core boost is different from single-core boosts nowadays and is also the reason why the "performance governor" can nowadays hurt performance.
It's not always evident in benchmarks especially when you hit all cores with heavy load anyway.

Correction, modern Intel CPu's like Rocket Lake can consume 250-300W (at least if we are talking about desktop "consumer" CPU's). AMD's ryzen's top end CPU's (i.e. Ryzen 5950x) consume ~110W. Ryzen server CPU's (i.e. epyc) end up consuming ~240W, but we are looking at CPU's that have up to 64 cores and 128 threads here.

Even without the improvements to p-state. modern AMD laptop CPU's are far superior to Intel's when it comes to power/efficiency, Intel CPU's are just much smarter in reducing the power use when they think nothing is happening.

[jaxa]

Correction, modern Intel CPu's like Rocket Lake can consume 250-300W (at least if we are talking about desktop "consumer" CPU's). AMD's ryzen's top end CPU's (i.e. Ryzen 5950x) consume ~110W. Ryzen server CPU's (i.e. epyc) end up consuming ~240W, but we are looking at CPU's that have up to 64 cores and 128 threads here.

Even without the improvements to p-state. modern AMD laptop CPU's are far superior to Intel's when it comes to power/efficiency, Intel CPU's are just much smarter in reducing the power use when they think nothing is happening.

Based on this, more like 140 Watts for CPUs like the 5900X and 5950X when maxed out. Still a good result.

Intel's Alder Lake could push power consumption even further than Rocket Lake, at least for the top i9-12900K. AMD is rumored to move the TDP up from 105 Watts to 170 Watts for top Raphael (Zen 4) CPUs on the AM5 socket, so you could expect those to hit up to 200 Watts if cooling is sufficient.

[mdedetrich]

Based on this, more like 140 Watts for CPUs like the 5900X and 5950X when maxed out. Still a good result.

Yeah hence the approximate 110, you can push it further if you run a power virus but then if you do the same with a rocket lake CPU you can get above 300W if you unlock the power delivery from the motherboard.

As you pointed out though, there is a world of difference in power usage between the 2 CPU's

Intel's Alder Lake could push power consumption even further than Rocket Lake, at least for the top i9-12900K. AMD is rumored to move the TDP up from 105 Watts to 170 Watts for top Raphael (Zen 4) CPUs on the AM5 socket, so you could expect those to hit up to 200 Watts if cooling is sufficient.

Will be interesting to see, one reason why Rocket Lake is so bad is that its still 14nm (with a ported 10nm architecture). Theoritically on Intels 10nm (now called Intel 7) it should be more efficient, but we will have to see.

[mppix]

Correction, modern Intel CPu's like Rocket Lake can consume 250-300W (at least if we are talking about desktop "consumer" CPU's). AMD's ryzen's top end CPU's (i.e. Ryzen 5950x) consume ~110W. Ryzen server CPU's (i.e. epyc) end up consuming ~240W, but we are looking at CPU's that have up to 64 cores and 128 threads here.

Even without the improvements to p-state. modern AMD laptop CPU's are far superior to Intel's when it comes to power/efficiency, Intel CPU's are just much smarter in reducing the power use when they think nothing is happening.

Sure AMD is on a better node but PBO will also use every bit of the power and thermal envelope that it can. Also some models have many more cores.
The discussion here is that modern processors cannot sustain their boost clocks and pinning them to high boosts does not give highest performance. It is better to 'save' some 'thermals' for when you really need them.

[mdedetrich]

The discussion here is that modern processors cannot sustain their boost clocks and pinning them to high boosts does not give highest performance. It is better to 'save' some 'thermals' for when you really need them.

Well actually the point I was making is that Intel CPU's cannot maintain boost clocks under normal circumstances (by default iirc Intel's boost only lasts ~50 seconds), AMD can assuming that your thermals are in check. AMD CPU's (and NVidia GPU's actually, there its called dynamic boost) dynamically boost the clock depending on the power/thermal budget being allocated to the CPU (simplifying here). The ironic issue here with AMD CPU's is that on typical motherboards/cooling systems they typically hit the thermal limit faster than any power limit which basically means for the most part the clock speed is typically static.

Gamers nexus does a really good video on this https://www.reddit.com/r/Amd/comment...oes_precision/. Its also worth reading the comments in this reddit thread because the behavior can also differ wildly depending on your motherboard (mainly because mobo manufactures really like screwing with power/voltage as an advertisement gimmick)

[numacross]

Will be interesting to see, one reason why Rocket Lake is so bad is that its still 14nm (with a ported 10nm architecture). Theoritically on Intels 10nm (now called Intel 7) it should be more efficient, but we will have to see.

Granted this is just a rumor, but it seems that Intel has increased the requirements for the power subsystem for Alder Lake.

[mppix]

Well actually the point I was making is that Intel CPU's cannot maintain boost clocks, AMD can assuming that your thermals are in check. AMD CPU's (and NVidia GPU's actually, there its called dynamic boost) dynamically boost the clock depending on the power/thermal budget being allocated to the CPU (simplifying here). The ironic issue here with AMD CPU's is that on typical motherboards/cooling systems they typically hit the thermal limit faster than any power limit which basically means for the most part the clock speed is typically static.

Gamers nexus does a really good video on this https://www.reddit.com/r/Amd/comment...oes_precision/. Its also worth reading the comments in this reddit thread because the behavior can also differ wildly depending on your motherboard (mainly because mobo manufactures really like screwing with power/voltage as an advertisement gimmick)

Yes, it is actually quite interesting how intel managed to remain at least somewhat competitive on the desktop.

Re AMD clock speed: I think even on AMD single core boosts are higher than multicore bosts, especially on 5900x and 5950x. For lower core CPUs the clock speeds may be closer if not the same..

[cyring]

Those would have been done more than a year ago if the AMD specifications were more than a preview of Registers subset.
The old way, BKDG for Zen families, won't happen unfortunately.

I have developed some TSC based virtual PMC like counters to measure the residency elapsed cycles when kernel is routed to my idle driver: https://github.com/cyring/CoreFreq

It gives somehow a hint but it is not as accurate as the fixed performance C3, C6, C7 ... C10 and PCn provided with the Intel MSR.
Wish AMD give some details about Package and Core C-States counters.

[brucethemoose]

Well actually the point I was making is that Intel CPU's cannot maintain boost clocks under normal circumstances (by default iirc Intel's boost only lasts ~50 seconds), AMD can assuming that your thermals are in check. AMD CPU's (and NVidia GPU's actually, there its called dynamic boost) dynamically boost the clock depending on the power/thermal budget being allocated to the CPU (simplifying here). The ironic issue here with AMD CPU's is that on typical motherboards/cooling systems they typically hit the thermal limit faster than any power limit which basically means for the most part the clock speed is typically static.

Gamers nexus does a really good video on this https://www.reddit.com/r/Amd/comment...oes_precision/. Its also worth reading the comments in this reddit thread because the behavior can also differ wildly depending on your motherboard (mainly because mobo manufactures really like screwing with power/voltage as an advertisement gimmick)

Right... in desktops with aftermarket cooling. That's a whole different ballgame.

My laptop can not sustain its boost under load, even with the dGPU totally off and no major graphics load. OEM desktops are probably somewhere between those two, and IGP-only systems with a graphics load will be even more extreme than my laptop.