#asahi-re · 2022-12-03 14:14
marcanultra is just two max dies#asahi-re · 2022-12-03 14:14
marcaneverything has one functional one per die, Ultra has two dies#asahi-re · 2022-12-03 14:14
marcaneiln: it's 16 cores per neural engine#asahi-re · 2022-12-03 09:39
svenand I think marcan tried turning them on in pmgr and that just didn’t work #asahi-re · 2022-12-03 09:38
svenIboot removes them from the template device tree iirc#asahi-re · 2022-12-03 06:40
eilnassociated clock/power node doesn't go through? #asahi-re · 2022-12-03 06:40
eilnalso what exactly is meant by "iBoot filters them out"?#asahi-re · 2022-12-03 06:40
eilnh14 firmware is definitely different tho if that signals anything#asahi-re · 2022-12-03 06:40
eilni guess m2+ lines would give the definite answer#asahi-re · 2022-12-03 06:39
eilnmy question is why they'd add more cores to ultra if 3 max models had issues with exactly that? #asahi-re · 2022-12-03 06:39
eilnseems they are pretty set on that decision#asahi-re · 2022-12-03 06:39
eilnassuming core == ne*8, that checks out#asahi-re · 2022-12-03 06:39
eilnand "16-core Neural Engine" for all else#asahi-re · 2022-12-03 06:39
eilnmarketing speak says "32-core Neural Engine" for ultra#asahi-re · 2022-12-03 06:39
eilnthat makes a lot more sense#asahi-re · 2022-12-03 06:39
eilnso "device with two neural processor circuits" == just ultra#asahi-re · 2022-12-03 06:39
eilnohh I thought max/pros had ane1 enabled#asahi-re · 2022-12-03 04:06
marcanand yes, geohot's "research" is, well... yeah what you said tracks#asahi-re · 2022-12-03 04:04
marcanit was probably intended for M1 Max too, I don't think the reason they ditched the second ANE is because it wasn't ready in time, though it could be#asahi-re · 2022-12-03 04:03
marcanmacOS has some kind of balancer/dispatcher kext for this#asahi-re · 2022-12-03 04:03
marcanconcurrency *should* be used on M1 Ultra for the ANE across each die#asahi-re · 2022-12-03 04:02
marcanmy theory is it's either broken silicon or they decided not to implement it due to power envelope reasons#asahi-re · 2022-12-03 04:01
marcanthey seem to be hard shut down#asahi-re · 2022-12-03 04:01
marcanI don't know why, but iBoot filters them out#asahi-re · 2022-12-03 04:01
marcaneiln: also you may know this already, but ane1/ane3 are fused off on Max/Ultra devices#asahi-re · 2022-12-03 01:54
amarioguyto match up the xnu "crc" with the GL CRC#asahi-re · 2022-12-03 01:53
amarioguythink i'll want to look at the gigalocker file i have very closely tomorrow#asahi-re · 2022-12-03 01:50
amarioguy(the upper two bytes i mean)#asahi-re · 2022-12-03 01:50
amarioguyon the tracer side though, the actual mailbox message has that two byte CRC at the start#asahi-re · 2022-12-03 01:50
amarioguythen it logs as "saving" the xart with a 2 byte crc#asahi-re · 2022-12-03 01:49
amarioguyhuh so during the unlock sequence after the first unlock, XNU logs the replies on the AP side with 0xffff in the upper two bytes#asahi-re · 2022-12-03 01:37
amarioguysure hopes i'm doing this mmio tracing thing right#asahi-re · 2022-12-03 01:36
amarioguy(yes i am just dumping what i find as i find it how did you know :) )#asahi-re · 2022-12-03 01:36
amarioguyokay so one of the absolute most verbose sep operations is pfk_data_unwrap and it's associated reply in the tracer#asahi-re · 2022-12-03 01:26
amarioguy(or rather right after SEP changes the lock state)#asahi-re · 2022-12-03 01:26
amarioguyLOL setting a breakpoint while the framebuffer comes on causes snow effects like old TVs#asahi-re · 2022-12-03 01:19
amarioguyalso apparently single user mode data accesses do not trigger SEP - likely because of how it's decrypted on the fly through the storage controller ephemeral key#asahi-re · 2022-12-03 00:28
amarioguyokay so the CRCs that XNU is saying the lockers have don't match up at all with the gigalocker CRCs...#asahi-re · 2022-12-03 00:03
amarioguyor maybe i have this all wrong and AP just queries the locker on demand whenever it needs to decrypt it#asahi-re · 2022-12-03 00:02
amarioguyperhaps those have smth to do with it#asahi-re · 2022-12-03 00:02
amarioguystrings on apple sep manager talk about "OOL buffers"#asahi-re · 2022-12-03 00:02
amarioguyrn i'm just trying to figure out how SEP knows where the gigalocker is because i'm not seeing any shared writes #asahi-re · 2022-12-03 00:02
amarioguyah i see#asahi-re · 2022-12-03 00:01
sventhat one just walks the pagetable to find the physical address and then just reads that #asahi-re · 2022-12-03 00:00
sventhere’s dart.ioread or something like that #asahi-re · 2022-12-02 23:56
amarioguyis a bit nervous#asahi-re · 2022-12-02 23:56
amarioguyi'm not quite sure ngl seems like the read commands in m1n1 only apply to main system memory so unless there's a translation func or smth a bit lost#asahi-re · 2022-12-02 23:56
amarioguybut i mean more when you're broken into the target and need to read the region#asahi-re · 2022-12-02 23:56
amarioguylike i mean at your command, obviously the tracer will update the view of shmem if it changes#asahi-re · 2022-12-02 23:55
amarioguysven: kind of a dumb question but...how exactly do you read a DART shmem region in the m1n1 hv debugger?