Received my Framework Desktop on Monday and been quite impressed with the unit.
Running some benchmarks I noticed a unexpected behavior related to the power level and boost behavior. The APU does not seem to boost past 115W and will settle on 100W after a short while. Thermals look decent and far away from any throttling. This seems to be at odds with the product spec that states: “Processor power level: 120W sustained, 140W boost power”.
I was not able to spot any setting in the firmware or the AMD software that would influence the power limit.
Framework Desktop Max+ 395 128GB
Windows 11 Business 24H2 10.0.26100.4946 fresh install
Power Scheme: High performance
Drivers: Framework_Desktop_AMD_Ryzen_AI_Max_300_driver_bundle_W11_v101_2025_07_10
BIOS: 3.02
To rule out something OS related I booted a Fedora 42 live environment and tested with s-tui. Same behavior: 115W boost that drops to 100W.
Are you able to test it with a mixed workload? Like if you run LM Studio so it loads up the GPU instead of the CPU, or for some models I’ve seen them split between the CPU & GPU, do you get any different results? Or gaming?
It will be a little frustrating if it turns out there’s no way to get it to run at higher power levels, since so many mini PCs these days have fully unlocked BIOS that let you tune everything to your liking. The 395+ has so much raw compute available that for maximum performance I would personally want the BIOS unlocked to let me set the PL1, PL2, PPT, temperature targets, seconds to PL2, etc. That’s what my current GMKtec PC does, and GMKtec also sells an AMD 395+ system. I just prefer how modular the Framework appears to be, which is why I ordered one of those.
I’ve had great success with “Universal x86 Tuning Utility” for setting custom power limits and temperature limits. You have to install an old version of .NET to make it work since it appears the software isn’t actively under development anymore, but even on my Ryzen 7 8845HS I’ve been able to adjust power limits. If you’re bored and want to try, then pershaps you can see if it works for raising the power limit on your Framework Desktop.
I do hope we’re not about to discover its the PSU which is the limiting factor.
Reviews of other manufacturers Max+ 395 mini-PCs show them working at 140W peak, 120W sustained so this is Framework-specific. It’d be VERY interesting to hear from anyone who bought a motherboard only as to whether they see the same with another PSU.
This is basically a deal-breaker for me as the entire reason for the Framework Desktop was to enable Strix Halo to be run at its full potential.
Edit - Eve clients don’t generally do much if docked up. Park them all outside Jita 4-4 and see what happens
First of all I highly doubt that the PSU is limiting anything here, the 400W unit Framework is providing is already at a much higher capacity than any of the other Strix Halo MiniPCs which are generally in the 200W-250W range.
I wouldn’t put too much stock into these numbers being measured from software. The numbers I am seeing here are consistent with what some reviewers have shown, however the best way to measure true power consumption is “from the wall”, and what I have seen is that these measurements are more consistent with the 120W sustained / 140W boosts mentioned in the marketing.
But even if these numbers did somehow imply that the APU is not being used to its full potential, what exactly would the performance difference be in real terms? Probably very little.
The version of Strix Halo in laptops is generally ran at considerably lower power limits than this, so even if we believe these are the actual power consumption numbers, the real world performance should be considerably better than the laptop counterparts.
As you asked about the cpu governor I assume you ask about it in Linux. I rebooted into a Fedora live environment, default was set to “powersave”. Changing it to performance did not make an impact .. please excuse the screen photo:
Most of my tests were under Windows 11 with power scheme set to “high performance”. I verified PERFEPP and PERFEPP1 in the power plan, both are set correctly:
> powercfg -q SCHEME_CURRENT SUB_PROCESSOR PERFEPP
Power Scheme GUID: 8c5e7fda-e8bf-4a96-9a85-a6e23a8c635c (High performance)
GUID Alias: SCHEME_MIN
Subgroup GUID: 54533251-82be-4824-96c1-47b60b740d00 (Processor power management)
GUID Alias: SUB_PROCESSOR
Power Setting GUID: 36687f9e-e3a5-4dbf-b1dc-15eb381c6863 (Processor energy performance preference policy)
GUID Alias: PERFEPP
Minimum Possible Setting: 0x00000000
Maximum Possible Setting: 0x00000064
Possible Settings increment: 0x00000001
Possible Settings units: %
Current AC Power Setting Index: 0x00000000
Current DC Power Setting Index: 0x00000000
> powercfg -q SCHEME_CURRENT SUB_PROCESSOR PERFEPP1
Power Scheme GUID: 8c5e7fda-e8bf-4a96-9a85-a6e23a8c635c (High performance)
GUID Alias: SCHEME_MIN
Subgroup GUID: 54533251-82be-4824-96c1-47b60b740d00 (Processor power management)
GUID Alias: SUB_PROCESSOR
Power Setting GUID: 36687f9e-e3a5-4dbf-b1dc-15eb381c6864 (Processor energy performance preference policy for Processor Power Efficiency Class 1)
GUID Alias: PERFEPP1
Minimum Possible Setting: 0x00000000
Maximum Possible Setting: 0x00000064
Possible Settings increment: 0x00000001
Possible Settings units: %
Current AC Power Setting Index: 0x00000000
Current DC Power Setting Index: 0x00000000
The choice of governor shouldn’t affect the power limits. I think @Alexandru_Stan was referring to the power mode which you can change from the top right menu.
Could you try with HWInfo and look up both CPU Power and look up these 3 under full load
it’s quite possible the Power limit is meant for full SoC, not just the CPU. The IOD will take it’s power share. On Desktop it usually accounts for 30W under load, and this counts towards the socket 200W power limit on 9950x for example.
Thank you for the clarification. Changing the power profile in Fedora indeed makes a difference. Set to performance its the package power does a quick burst to 150W, about 9 minutes of 140W before dropping to 120W under load.
So does this have an equivalent in Windows? Like does this mean it will run 100/115 watt in Windows using the Balanced profile, but if you set Windows to Performance then it too will run at the full, advertised power limits?
The equivalent settings are the power scheme and/or power mode. Both were set to “High performance.” Changing to “Balanced” dropped the package power under load to ~55–60 W.
I ended up resetting the power schemes, and that made all the difference: boost peaks to 150, ~140 for extended periods, and ~120 sustained for 60+ minutes. The boost behavior is now comparable to Fedora with the power mode set to Performance. I looked at a few YouTube reviews, and it seems I’m not alone: https://youtu.be/D7BehyQVVbU?t=1259
It might be an artifact of the Windows + driver installation. I’ll try to replicate it over the weekend.
Thank you for the rubber ducking.
Steps taken:
powercfg -restoredefaultschemes
Reboot, open Settings → System → Power, and change Power mode to “Best performance.”
P.S. I can now answer the “what exactly would the performance difference be in real terms?” question: about 12% in synthetic benchmarks and ~8% faster in my real-life workload.
Thank you very much for coming back with the info.
It does rather beg the question as to why Windows power scheme levels are set this way - I presume you did a fresh install etc? My guess is the drivers set it for sustained usage of 100W as they assume its a laptop.
Also means a lot of those reviews are invalid as they don’t show peak performance under Windows.
Edit - it could be the Adrenaline package on Windows which is doing it. If it reverts to the lower sustained TDP then that’d be my bet.
I have seen the same power level behaviour under Linux (openSuse Tumbleweed) without changing any power management settings from stock. Short boost to 115W and then continuous 100W TDP. In my case however that suits me really well. Generally I think 100W is at the upper end of an efficiency sweet spot of that APU.
It is nice to read that I merely loose 8% real world performance while saving around 20% power and heat. Given that my 395 is in a fully passive case, these things matter in my case.
