#asahi · 2022-10-27 09:29
chadmedgrub-install is unnecessary and that error is because u-boot does not provide uefi nvram#asahi · 2022-10-27 09:29
chadmedyep#asahi · 2022-10-27 09:28
chadmedhttps://github.com/chadmed/asahi-gentoosupport/blob/main/install.sh#L99 like this#asahi · 2022-10-27 09:28
chadmedor you can i guess, but you need to specify that its a removeable drive #asahi · 2022-10-27 09:27
chadmedoh right yeah dont do grub-install #asahi · 2022-10-27 09:26
chadmedthe overlay has the latest kernel, use it :)#asahi · 2022-10-27 04:49
chadmedi reckon the benchmark asm just needs to light up more of each core#asahi · 2022-10-27 04:48
chadmedwhen im at home and have a couple of days ill dig out the m1n1 experiment from the backup of my dev machine (RIP) and try to see what we were doing wrong there#asahi · 2022-10-27 04:48
chadmedSMC, freqbench has no way to access the counters since it relies on voltage_now and current_now from power supply class devices#asahi · 2022-10-27 04:47
chadmedyeah im still not happy with the lowest pstates given the wide variance/noise in the smc power measurements. sorry i didnt save the run logs#asahi · 2022-10-27 03:31
chadmedhttps://tpaste.us/X0Ye #asahi · 2022-10-27 03:29
chadmedyep #asahi · 2022-10-27 03:14
chadmed dont have time to do it again but i captured the results #asahi · 2022-10-27 03:14
chadmednvm it was saved in tmpfs -.- #asahi · 2022-10-27 03:13
chadmedyeah one second#asahi · 2022-10-27 03:09
chadmedso a little more sense #asahi · 2022-10-27 03:09
chadmedresults ranging from 100mW to 5W #asahi · 2022-10-27 03:08
chadmedok now things are making a little more sense #asahi · 2022-10-27 02:45
chadmedim going to run it one more time#asahi · 2022-10-27 02:41
chadmedits not really comparable#asahi · 2022-10-27 02:41
chadmedyeah i understand that but again, that measures whole soc power on cores that are tuned for different pstates at different V-P curves #asahi · 2022-10-27 02:41
chadmedthere is divergence at the top end of the pcore results at the highest pstates but below that its literally just noise and there is no way to verify that from userspace at this time #asahi · 2022-10-27 02:40
chadmed"that doesnt look right, the snapdragon uses more power" really doesnt invalidate the methodology i checked over multiple times and cross referenced with an entirely different one tbh #asahi · 2022-10-27 02:39
chadmedi think the cores really just are this efficient#asahi · 2022-10-27 02:39
chadmedno they shouldnt, which is why i use m1n1 and hit the energy counters inside the cores directly to avoid this exact thing happening, where the total system power turns the work a single core is doing into noise #asahi · 2022-10-27 02:38
chadmedthe system is doing a load of other stuff, and the work the single core is doing at 600mhz just falls into background noise#asahi · 2022-10-27 02:38
chadmedbecause its like i predicted#asahi · 2022-10-27 02:37
chadmedthe core creates energy #asahi · 2022-10-27 02:37
chadmedno as in NEGATIVE 300mW #asahi · 2022-10-27 02:30
chadmedjim keller found a way to create energy from running instructions #asahi · 2022-10-27 02:30
chadmed-300mW uh-huh yeah sure#asahi · 2022-10-27 02:30
chadmedsorry#asahi · 2022-10-27 02:30
chadmedyeah this is utterly useless, its just noise#asahi · 2022-10-27 01:44
chadmedim setting the options recommended in the readme#asahi · 2022-10-27 01:44
chadmeddont have the userspace scheduler enabled in my normal build, and my timer frequency is 1khz#asahi · 2022-10-27 01:39
chadmedyouve nerd sniped me though, building a kernel now :P #asahi · 2022-10-27 01:33
chadmedso the scaling between them should be consistent#asahi · 2022-10-27 01:33
chadmedand like if we're going to say that what matters is the scaling between cores for the EM and the actual value doesnt matter, then it doesnt matter if theyre too low either since all cores and pstates were tested exactly the same way #asahi · 2022-10-27 01:32
chadmedprobably not, but if its going to get the power value out of the SMC then its going to show the same thing i got in m1n1 when i did it, which matched the values from the joule counters, which is what is in the DT#asahi · 2022-10-27 01:28
chadmedbut it wouldnt work because it has no conception of the SMC and im not going to add it just to arrive at the same conclusion i did in m1n1#asahi · 2022-10-27 01:27
chadmedor rather, if it _did_ work, it would give the value out of the SMC which will include the entire system power draw and not just one core#asahi · 2022-10-27 01:26
chadmedok no this isnt going to work for power sampling#asahi · 2022-10-27 01:19
chadmednot really willing to scratch up traces on the logic board of my daily driver to find out though :P #asahi · 2022-10-27 01:19
chadmedyeah i know, at first i thought we might not be initialising the counters properly, but like i said it lines up with an earlier methodology where i would use the SMC's integrated power meter. the values are within a few % of each other#asahi · 2022-10-27 01:16
chadmediirc the max opp-microwatt value for firestorm was ~27k or something ridiculous without floating point instructions#asahi · 2022-10-27 01:15
chadmedi dont think the methodology itself is _particularly_ flawed, for example there is a very noticeable increase in power use when you light up the FPUs#asahi · 2022-10-27 01:15
chadmed(in m1n1)#asahi · 2022-10-27 01:14
chadmedbasically what i did was read the joule counters, run a mix of integer and floating point instructions, measure how long they took, read the joule counters again, take deltaE and then divide by the time the instructions took #asahi · 2022-10-27 01:13
chadmedill spend some time refining the methodology when i have a few days to dedicate to it. fwiw i dont entirely disagree, and like i said we werent really sure ourselves if they were legit or not. #asahi · 2022-10-27 01:11
chadmedwhich doesnt actually tell us much about what the core clusters are doing