Is there any plans to update the bios on the framework desktop to allow for undervolting / overclocking? I saw a video yesterday of another max+ 395 desktop from a different manufacturer that was allowing AMD PBO options and would like to know if that’s something that has been considered and or if there’s a reason why its not already enabled in the bios? Thanks!
While more BIOS options are always appreciated, if you are talking about ETA Prime’s video, did anyone notice that he was pushing an extra 50w of power (175w?!?!) for a total of 4% increase in performance across the 4 benchmarks he ran…
Cinebench: +4.2%
Geekbench: +5.8%
Geekbench OPENCL: +2.5%
Timespy: +3.5%
I watch a lot of his videos, and he tends to think everything is awesome/amazing/super, but doesn’t actually talk about the metrics as a whole (he just flashes numbers without percentage increases or what not).
Anyway, it’s pretty well known at this point that whether AMD or Intel, yes you can overclock (PBO or manually), but to be honest the increase is extremely small (usually less than 10%) in most cases compared to the exponential increase in power draw.
1 Like
Yeah, I am mostly talking about that video. I had initially thought it was something that was not available from the platform, as I have seen some people talk about trying it with other software and it not working, particularly with Ryzen Master. What he showed was not amazing, I think he didn’t do it justice given how well the max+ chips seem to scale with power. My main interest for this in the bios is undervolting.
You don’t need bios settings to undervolt.
Also: Overclocking beyond the PBO range won’t net a lot from what i’ve tested.
The following chart shows the reported package power during a stress-ng -c 32 run. I limited the cpu clock to 600MHz (lowest clock) and in 100MHz steps between 1200MHz and 5100MHz.
When looking at the raw numbers, we’ve got between 60 - 75 MHz/W for clocks <= 4GHz. > 4GHz the efficiency drops and reaches 35MHz/W at 5GHz.
You get 4GHz for ~65W, but need 140W for 5GHz.
PS: when running avx512 operations, the number look slightly different, but the sweet spot remains. You need about 4W of power for every 100MHz step until you reach 4GHz. >4GHz you’ve got 2 more steps that each require 6W and hit the power ceiling of 140W at ~4.2GHz.
Disclaimer: This is a very unsientific way of testing Perf/W. Do not compare those numbers in absolute values. I run hand optimized per-core undervolts.
1 Like
It might not be needed but it makes it a lot easier without using a 3rd party application. Undervolting is not entirely about performance gains. For example the 7600x I had was previously able to undervolt to -20 all core stable not increasing performance all that much but cutting down on wasted power and decreasing temps significantly. I’m also less interested in overclocking / undervolting on the cpu. I don’t see why this wouldn’t be a solid add for users regardless of the results. For framework wanting to be the company that it’s portraying to be it’s a little weird for them to not support it from the start.
I’m confused. Were you not asking for exactly that?
Not entirely, no. But in this case, the CPUs performance is:
-
- limited by available power
-
- limited by thermals
So undervolting directly affects performance.
We have to differentiate undervolting the CPU from undervolting the iGPU. Those are separate things. Undervolting the iGPU is currently not even supported with ryzenadj, while undervolting the CPU is supported.
Ohter APUs have that iGPU undervolting supported with ryzenadj, so it is probably not impossible but it seems uncertain if such a feature will be yet added to ryzenadj.
Yes, BIOS undervolting would be more accessible but ryzenadj is great. It can do much more. I used it to reduce the Tjmax. You can basically reduce it to the temperature max you want to see (I set it to 88°C) and it does not move even a bit above those limits and will accordingly throttle. You can combine this with raising the sustained TDP to eg 140W btw and then it will not boost for just 10 min but as long as that healthier lowered Tjmax allows. This is ideally combined with undervolting of course.
1 Like
Your graph was showing the cpu scaling and I am more interested in what the igpu can achieve. Especially if you are able to shift more of the total system power over to the igpu.
Has anyone from framework ever commented on this? I really want this to get some traction.