#
#asahi · 2026-08-03 12:57
fl0_idgot it
#asahi · 2026-08-03 12:57
chaos_princesssince the decoder is what takes most of the time
#asahi · 2026-08-03 12:57
chaos_princessand on macos, it does slightly reduce the runtime, but like a few percent.
#asahi · 2026-08-03 12:57
fl0_idright
#asahi · 2026-08-03 12:57
fl0_idsupposedly even davinci resolve uses it. but on my version I couldn't find this switch
#asahi · 2026-08-03 12:56
chaos_princessfl0_id: it is more complex than that, the thing above only runs the encoder portion
#asahi · 2026-08-03 12:56
fl0_idin the exaample, eiln ran sth like whisper, which I guess could work. but not sure if it's worth it over gpu
#asahi · 2026-08-03 12:56
iyesyea on macos i use it a lot, because i use macos for creative and productivity tasks like image processing and photo/video editing, and those apps have good NPU support. I am not a fan of generative stuff or anything that replaces human creativity, but to solve technical problems, like AI denoise, masking, etc. it is great!
#asahi · 2026-08-03 12:52
chaos_princessand if you do local stuff - you are probably running a quant, and the npu is not flexible enough to do dequantization
#asahi · 2026-08-03 12:52
iyesokay so i wouldnt benefit from the npu anyway even if there was a good driver and api
#asahi · 2026-08-03 12:51
chaos_princessthe llm runtime is usually dominated by the token generation phase, which the npu is inherently terrible at
#asahi · 2026-08-03 12:50
iyessee if it can help me automate refactoring code or generating test cases. i am very new to this AI stuff, havent used it much, but i want to explore it
#asahi · 2026-08-03 12:49
iyesyeah on linux i am mostly interested in trying to run local LLMs for assistive tasks
#asahi · 2026-08-03 12:49
chaos_princessdepends on how exactly their model is structured
#asahi · 2026-08-03 12:49
iyesive wondered even in macOS, for proprietary productivity/creative applications like Adobe Lightroom that give you the option to run AI models on the GPU or the NPU, which is better. i am not sure...
#asahi · 2026-08-03 12:48
chaos_princessit is suited for serious ml stuff, just not the kind of ml stuff that everyone is currently crazy about
#asahi · 2026-08-03 12:47
iyesit vaguely matches my understanding of the purpose of this hardware. they put it in the chip so that random little ML tasks that are not very computationally intensive can run efficiently without consuming much power. it's not suited for serious ML stuff
#asahi · 2026-08-03 12:46
iyesi read the messages you linked
#asahi · 2026-08-03 12:46
iyesokay
#asahi · 2026-08-03 12:43
chaos_princessdon't bother, it is basically a proof of concept, not a real driver
#asahi · 2026-08-03 12:42
iyes(there is no prebuilt package or anything)
#asahi · 2026-08-03 12:42
iyesand does it compile with the latest kernels? if i want to install it on current up to date fedora asahi remix, the way to do that is to clone the repo and build it?
#asahi · 2026-08-03 12:41
chaos_princesshttps://oftc.catirclogs.org/asahi/2026-08-02#35560832;
#asahi · 2026-08-03 12:41
iyescan it run any real applications? or only some test cases?
#asahi · 2026-08-03 12:40
fl0_idthere's like a benchmark for whisper here https://github.com/ggml-org/whisper.cpp/pull/1021
#asahi · 2026-08-03 12:40
elisaadojust not for the things you think it works for
#asahi · 2026-08-03 12:40
fl0_idiyes it's here. https://github.com/eiln/ane It doesn't really have userspace afaik
#asahi · 2026-08-03 12:40
elisaadoit's not, it works
#asahi · 2026-08-03 12:36
iyeswhat happened to the WIP neural engine driver? is it abandoned?
#asahi · 2026-08-03 11:44
chadmedgive the installer a go now
#asahi · 2026-08-03 11:42
Dysnostalgia[m]now more space
#asahi · 2026-08-03 11:42
Dysnostalgia[m]uploaded an image: (175KiB) < https://matrix.org/oftc/media/v1/media/download/AQKucmrzoH1N2nK1cFfiQGCyazzUw_XNz-5uardtBix5P7VtHbTpbb7ei0W6eG_7W75ZOjgRjz6KXxTKVsoZzyFCegD4eTMAAG1hdHJpeC5vcmcvY2NyRlZKVmdWYXlXTU1Ud2p1UG1JUWx1 >
#asahi · 2026-08-03 11:41
Dysnostalgia[m]https://discussions.apple.com/thread/255341336?sortBy=rank
#asahi · 2026-08-03 11:41
Dysnostalgia[m]purged time machine backups
#asahi · 2026-08-03 11:35
Dysnostalgia[m]uploaded an image: (155KiB) < https://matrix.org/oftc/media/v1/media/download/AZo31oLrVUzahDSYsl1EZ_X3adAqDZPX0FIsYPEVpHLy8WUIALsAj7JhEF6ueD8_SeqhdV0Dw7J25xxEktiPFutCegD4FnTgAG1hdHJpeC5vcmcvZ2hvSUZGclpHaGN5U2ZNYUZ1dk9YSU5t >
#asahi · 2026-08-03 11:34
Dysnostalgia[m]uploaded an image: (106KiB) < https://matrix.org/oftc/media/v1/media/download/AVhztKj_-BLMB3WGUZtC_hORP6uBPJSM8UBR4QTTLtoPs5KHdPQHrjvd-e_nv9Ahy70_aZ3vTkAh90jkqn5UoN1CegD4B7HQAG1hdHJpeC5vcmcvbVFWVVFvVFVMZFBQdEN5SUJrdlVncHd4 >
#asahi · 2026-08-03 11:34
Dysnostalgia[m]disk utility shows more space filled
#asahi · 2026-08-03 11:32
Dysnostalgia[m]yes
#asahi · 2026-08-03 11:31
chadmeddo you have time machine enabled?
#asahi · 2026-08-03 11:29
Dysnostalgia[m]is this not enough space
#asahi · 2026-08-03 11:29
Dysnostalgia[m]uploaded an image: (38KiB) < https://matrix.org/oftc/media/v1/media/download/Af-PVhlzaHeky1j6Se24NWAkRUbQkN2GZuqMr69TiJKNdMKct-hhskZEiV3ahwtH4dEjrehULB4IOp01xrbrL3FCegD3vE6gAG1hdHJpeC5vcmcvem9iQXRUTmtYTkZDdkN0ZkdUSVBJdkVk >
#asahi · 2026-08-03 11:28
Dysnostalgia[m]uploaded an image: (55KiB) < https://matrix.org/oftc/media/v1/media/download/AVYTbMzabLI5qBh9CYkdFMvkfOwBv4RgKaVvI-bp8FDMpgIzryiHi1eX3-NIuTX1eX9cj2UcMe298jdugH-Br7hCegD3q_KwAG1hdHJpeC5vcmcvV1dGRUpqR0FpS3RncnJlRll5ZUVOTFh2 >
#asahi · 2026-08-03 08:43
elisaadothanks LarstiQ :)
#asahi · 2026-08-03 04:53
LarstiQelisaado: kampeer ze!
#asahi · 2026-08-02 22:34
namespaceBuonanotte a tutto #ilpub <º))))><
#asahi · 2026-08-02 21:11
fl0_idelisaado seems perfect to poke at things haha if not too hot for technology
#asahi · 2026-08-02 17:45
v1g1lantsuch an awesome project ty if anybody reads this
#asahi · 2026-08-02 16:16
elisaadolmk if I can help, this summer holiday is kinda boring me to death
#asahi · 2026-08-02 16:12
chaos_princess¯\_(ツ)_/¯
#asahi · 2026-08-02 16:10
elisaadotime to poke at some DCP registers on macOS?

plus ›