#
#asahi-dev · 2026-06-04 11:46
chadmedthe measure is per mhz, you need to peg one core to its top pstate and then clamp the coremark task to it
#asahi-dev · 2026-06-04 11:45
yukaAh, no
#asahi-dev · 2026-06-04 11:45
jannaudid you normalize them to the (max) frequency
#asahi-dev · 2026-06-04 11:44
yukaSo then my E-cores are 591
#asahi-dev · 2026-06-04 11:43
chadmedthat is the correct way. the fastest cores in the system are by convention always 1024, and the others are normalised relative to that
#asahi-dev · 2026-06-04 11:43
yukayou took the coremark number and then relativized them to 1024?
#asahi-dev · 2026-06-04 11:34
jannaubad performance on the boot core is not related to core switching
#asahi-dev · 2026-06-04 11:10
jannausurprisingly quite a difference between t8122 and t6030
#asahi-dev · 2026-06-04 11:09
jannauchaos_princess, yuka: see commit messages how I determined capacity-dmips-mhz
#asahi-dev · 2026-06-04 11:03
jannauit's only core 4 (active at m1n1 startup) which is inexplicable slow on M3. 17688 vs. 42851 Iterations/Sec for core 4 vs. core 5, 6, 7
#asahi-dev · 2026-06-04 10:01
chadmedyeah it sounds reasonable to me, especially if its reproducible
#asahi-dev · 2026-06-04 09:58
jannauso the 83% I get from coremark could be correct
#asahi-dev · 2026-06-04 09:57
jannauthat would put the e-cores at 80% of the p-cores based on our m1 capacity-dmips-mhz
#asahi-dev · 2026-06-04 09:54
jannau"The CPU performance cores and efficiency cores are 30 percent and 50 percent faster than those in M1"
#asahi-dev · 2026-06-04 09:51
chadmedbecause thats what crappy ecores usually come out at :P
#asahi-dev · 2026-06-04 09:51
jannauwhy had I 300-400 in mind for the e-cores?
#asahi-dev · 2026-06-04 09:50
chadmedblizz is like 74% of avalanche
#asahi-dev · 2026-06-04 09:50
chadmedyeah
#asahi-dev · 2026-06-04 09:50
jannauoh, I should have looked at the m1/m2 capacity-dmips-mhz values. m2 with 756 / 1024
#asahi-dev · 2026-06-04 09:49
chadmedit wouldnt surprise me tbh. firestorm and icestorm were pretty close, so were avalanche and blizzard
#asahi-dev · 2026-06-04 09:44
jannauand 56% of the absolute values with both at their max freqency
#asahi-dev · 2026-06-04 09:42
jannauno, p-cores are faster both absolute and relative to their max freq but I'm not trusting the results. e-core at 83% of p-core at iso frequency seems to high
#asahi-dev · 2026-06-04 09:39
chadmedecore still faster than pcore?
#asahi-dev · 2026-06-04 09:31
jannaum3 pro looks better but I doubt 8.784 vs 10.570 coremark iterations/s per MHz are expected
#asahi-dev · 2026-06-04 09:23
sveni don't have high hopes but it's worth a try
#asahi-dev · 2026-06-04 09:23
svenyeah
#asahi-dev · 2026-06-04 09:23
NickChan[m]i guess it helps that the bit is documented on apple's site
#asahi-dev · 2026-06-04 09:23
svencould try to file a radar about that as well if it's confirmed and see what happens :D
#asahi-dev · 2026-06-04 09:22
NickChan[m]in the actual code it means "TSO bit exists in AIDR_EL1 but TSO bit is RAZ/WI in ACTLR_EL1"
#asahi-dev · 2026-06-04 09:21
NickChan[m]sven: I went off "CPU features: CPU advertises Apple TSO but it is broken, ignoring" in a message on matrix in the re channel
#asahi-dev · 2026-06-04 09:13
StanfordAppliedCyber[m]not to mention are paying for their products
#asahi-dev · 2026-06-04 09:13
StanfordAppliedCyber[m]apparently they really like that we will never require customer support
#asahi-dev · 2026-06-04 09:13
StanfordAppliedCyber[m]yeah i heard that a lot of their engineers actually like asahi linux, and there's been some meetings that are basically "should we put resources into supporting them"
#asahi-dev · 2026-06-04 09:11
svenand their internal process for un-restricting sysregs in iboot is apparently especially annoying
#asahi-dev · 2026-06-04 09:11
StanfordAppliedCyber[m]so just like their regular QA process :)
#asahi-dev · 2026-06-04 09:11
chadmedwe might have used coremark instead of dhrystone
#asahi-dev · 2026-06-04 09:11
svenunfortunately we are their QA and have to wait months until the bugs are fixed then :(
#asahi-dev · 2026-06-04 09:10
svenStanfordAppliedCyber[m]: yeah, i know. i got it from people at apple as well
#asahi-dev · 2026-06-04 09:10
chadmedthe actual benchmark is kind of immaterial, youre supposed to fix the pcores to 1024 and normalise everything relative to that
#asahi-dev · 2026-06-04 09:10
StanfordAppliedCyber[m]sven: this is exactly what my friend at apple said. he told me they didn't test it at all internally and apparently there were some conversations like "should we test this, we are shipping it" "sure, but after you do all your other work"
#asahi-dev · 2026-06-04 09:09
chadmedistr we didnt use dhrystone for m1 or m2
#asahi-dev · 2026-06-04 09:09
chaos_princessok, reading failure on my part, i read it and failed to notice the part where you are supposed to run dhrystone
#asahi-dev · 2026-06-04 09:09
jannaudhrystone
#asahi-dev · 2026-06-04 09:09
jannauI'm using coremark instead whrystome though since 40 year old C code is annoying
#asahi-dev · 2026-06-04 09:08
jannauDocumentation/devicetree/bindings/cpu/cpu-capacity.txt
#asahi-dev · 2026-06-04 09:07
jannauI'm currently doing it on m3 and and the perf cores are slower than the efficiency cores (both at max frequency)
#asahi-dev · 2026-06-04 09:06
chaos_princessoh, forgot to ask - so how do you even do that?
#asahi-dev · 2026-06-04 09:06
jannauyuka, chaos_princess: did you already benchmark cpu cores for capacity-dmips-mhz on m3 max?
#asahi-dev · 2026-06-04 09:06
svenand likely missed quite a bunch of stuff
#asahi-dev · 2026-06-04 09:05
svenSobek[m]: don't assume any intention for raw boot mode, it's just very untested and they probably started with "everything disabled" and just put in a few exceptions

plus ›