#asahi · 2023-07-15 03:22
marcanyou can easily get 15+ hours of battery life on linux if you do that#asahi · 2023-07-15 03:21
marcanthis goes for any OS#asahi · 2023-07-15 03:21
marcanscree backlight swamps all other power consumers for regular desktop usage#asahi · 2023-07-15 03:21
marcanthat is *the* *one* thing that will decrease power consumption#asahi · 2023-07-15 03:21
marcankhant: lower your screen brightness#asahi · 2023-07-15 03:21
marcanit's somewhat normal for 100%->99% to be a bit fake for these things#asahi · 2023-07-15 03:20
marcanAndrewYu: the battery percentage is calculated by the battery controller, we show the same exact number as macOS#asahi · 2023-07-13 07:51
marcanI guess lina needs to look into whether kwin is doing something dumb specifically or whether there is a logical way out from that problem#asahi · 2023-07-13 07:51
marcanand this will break completely when we enable compression anyway#asahi · 2023-07-13 07:51
marcanwe can't make it not writeback because that is actually a correctness requirement for the display controller (it actually raises memory controller errors we don't currently handle but macOS does, I ran into this with m1n1)#asahi · 2023-07-13 07:50
marcanjtolio: the renderer pegging a CPU core sounds like it's trying to read out directly from the scanout framebuffer, which would be in writeback mode so stupidly slow to read out from#asahi · 2023-07-13 07:49
marcanthat sounds like running out of RAM, but I wonder if it shouldn't be#asahi · 2023-07-11 15:24
marcanreminder that m2pro anything is not supported yet and you *will* have to do a full reinstall when support is released#asahi · 2023-07-01 16:10
marcanmacOS lies in the UI and claims 8GB in the system specs because there is basically a "marketing RAM size" property. Linux does not have anything like that, it calculates RAM size based on actual RAM available to the OS, which is always less than the "round number" value.#asahi · 2023-06-29 12:59
marcanexactly :p#asahi · 2023-06-29 12:57
marcanthe Apple proprietary API for ANE (and other stuff) is CoreML. notably, it is fairly high level. chances are you will be able to do more stuff with our low level API, including non-ML stuff.#asahi · 2023-06-29 12:56
marcanI'm not aware of any particularly popular generic standard for ML accelerators, so ANE will probably get a bespoke API that various frameworks can add as a backend or whatever#asahi · 2023-06-29 12:55
marcanktz: there is no plan to support any Apple proprietary APIs. we will support open APIs like OpenGL, OpenCL, and Vulkan for the GPU.#asahi · 2023-06-26 13:43
marcani509vcb: stuff only gets hidden when resolved#asahi · 2023-06-17 11:18
marcanyes, just an nvram update#asahi · 2023-06-13 10:18
marcanmkurz: the palmdet was *horribly* broken on 16" models. if someone wants it back it's going to have to be a proper dev task to figure out how to improve it without regressing things.#asahi · 2023-06-12 07:35
marcanseb91nihel: not yet#asahi · 2023-06-11 04:36
marcanonce things are upstream it's a different story#asahi · 2023-06-11 04:36
marcanthere's no good reason to support old experiments that died#asahi · 2023-06-11 04:36
marcanrunning old kernels is not supported in general#asahi · 2023-06-11 04:36
marcanmaniacx: for anything that is not upstream (like this), I would just deprecate old behavior#asahi · 2023-06-10 13:44
marcanit doesn't do that, it just only disables the tick in that case as far as I know#asahi · 2023-06-10 13:37
marcanbut feel free to test it under other conditions#asahi · 2023-06-10 13:37
marcanbut that's idle#asahi · 2023-06-10 13:37
marcanhere it got worse by about 10% :)#asahi · 2023-06-10 13:36
marcannot sure about other workloads, but it doesn't seem helpful#asahi · 2023-06-10 13:35
marcanar: just tried nohz_full=all and if anything it made it worse I think (M2 air)#asahi · 2023-06-10 13:24
marcannohz_full=1-7 on M1/M2 would be worth trying#asahi · 2023-06-10 13:23
marcanworth trying it then I guess#asahi · 2023-06-10 13:22
marcanah but we need to specify the CPUs somehow?#asahi · 2023-06-10 13:21
marcanI haven't tested without#asahi · 2023-06-10 13:21
marcanar: CONFIG_NO_HZ_FULL=y on our kernels#asahi · 2023-06-10 13:13
marcanyou can always make better decisions about idle when you have scheduling information and know when you might need to wake up#asahi · 2023-06-10 13:13
marcanmacos doesn't even manage deep idle, it lets the CPUs decide when to do it with a heuristic, but I have no reason to believe that is actually a good thing#asahi · 2023-06-10 13:12
marcancy8aer: linux decides when to put the cores into deep sleep when idle. actually the latency numbers in the driver are BS right now, I need to do some actual measurements, so performance will probably change at some point. but the cores are already going into deep idle opportunistically.#asahi · 2023-06-10 13:10
marcanthe latter is a KDE quirk, if you don't do that it just refuses to apply the settings#asahi · 2023-06-10 13:10
marcanLinuxM1: you need to set the upper threshold to something <95 and the lower threshold to something *lower*#asahi · 2023-06-07 15:05
marcanbut cpu deep idle will help in all cases/socs#asahi · 2023-06-07 15:05
marcanover just leaving the machine idle with 0% CPU usage and screen off#asahi · 2023-06-07 15:04
marcanon base model socs, we don't, so you only get inherent power savings from a few regular devices powering down (mostly wifi I think)#asahi · 2023-06-07 15:04
marcanon pro/max we also have an extra special suspend driver that makes it save even more power in suspend#asahi · 2023-06-07 15:04
marcanis this a pro/max soc?#asahi · 2023-06-07 15:04
marcanheh yeah#asahi · 2023-06-07 15:02
marcanplease do try it though#asahi · 2023-06-07 15:02
marcanIt's better but not on par