琥珀青葉@KohakuLab

Graduate student in Taiwan. Leader of KohakuLab. Researcher/Dev in ComfyOrg

[琥珀青葉@KohakuLab @KBlueleaf]

TT-VidT is accepted at #NeurIPS2026 🎉

A temporal-decoupled methodology in video pretraining to provide real "motion" info extracted from the video and only requires small scale training to reach high quality results.

Huge thanks to all collaborators, @NVIDIAAI and @spellbrush

[琥珀青葉@KohakuLab @KBlueleaf]

I have 2x2080ti-22G (modded) and I found 2080ti also have similar unlock stuff from community,
I'm trying to make it become initramfs module (easier to recover than the UEFI module design) and see if it works.

UwUb

[琥珀青葉@KohakuLab @KBlueleaf]

Confirmed that 3090 unlock is successful
with 360watt power wall, I got 75T -> 135T (measured in Pytorch)
Which is insane

I used https://t.co/a9UZiHD4Gu with commit a9b0c75

[琥珀青葉@KohakuLab @KBlueleaf]

@bobziyangding UH such nerf is WELL KNOWN for the whole compute community
We just not really know how nvidia exactly nerf it before

And it also turned out 3090/4090/5090 need fully different solution to resolve such nerf, only 30 (or 20) can just write a register to fix it.

[琥珀青葉@KohakuLab @KBlueleaf]

Confirmed that 3090 unlock is successful
with 360watt power wall, I got 75T -> 135T (measured in Pytorch)
Which is insane

I used https://t.co/a9UZiHD4Gu with commit a9b0c75

[琥珀青葉@KohakuLab @KBlueleaf]

@Epic_Bob1 the repo i paste is linux only
works for 30 series, incude 3070ti

[琥珀青葉@KohakuLab @KBlueleaf]

@CDerinbogaz This won't increase your token throughput and I never claim token throughput, T here means TFlops not Token

[琥珀青葉@KohakuLab @KBlueleaf]

@BellaDabddest tensor core is not related to gaming

[琥珀青葉@KohakuLab @KBlueleaf]

@GustavSincero 40: yes
30: not sure, unless have similar path as 40 series

[琥珀青葉@KohakuLab @KBlueleaf]

@roach_eth Yes

[琥珀青葉@KohakuLab @KBlueleaf]

@Askinitezu Should

[琥珀青葉@KohakuLab @KBlueleaf]

@Maestro_Pacelli Ask GPT, ngl
let GPT to check your machine, check those github repo
and it will know how to do things safely

[琥珀青葉@KohakuLab @KBlueleaf]

@Char1es4k 因為這東西是靠exploit去改GPU private reg的 (4090那個則是hack firmware)

[琥珀青葉@KohakuLab @KBlueleaf]

@NEOAethyr yes, check the thread some ppl have posted a repo for 4090

[琥珀青葉@KohakuLab @KBlueleaf]

@Andrei_Teaches it have same nerf, but how to unlock is still unclear

[琥珀青葉@KohakuLab @KBlueleaf]

@SprayCactus all 30 series will work

[琥珀青葉@KohakuLab @KBlueleaf]

@pale_avenger all 30 series will work

[琥珀青葉@KohakuLab @KBlueleaf]

@scrubz49 no, at least not now

[琥珀青葉@KohakuLab @KBlueleaf]

@VegasMorph not measurable difference.
as the total power are capped and actually you have exactly same "speed" with fp16 accmulation

[琥珀青葉@KohakuLab @KBlueleaf]

@euribem 30/40 should all works, 20 also, 50 need different solution

[琥珀青葉@KohakuLab @KBlueleaf]

@cont3mplador no, at least not now

[琥珀青葉@KohakuLab @KBlueleaf]

@Major5965 the lock/nerf is a thing
the method for unlock is not, 5090 need fundamentally different path
I'm still checking if 5090 firmware utilize some efuse to avoid unlock, if efuse only for vbios/firmware version then it MAY BE possible but still unclear

[琥珀青葉@KohakuLab @KBlueleaf]

@HyperS815 A100 never got nerfed bruh, NV nerfed consumer grade level card to sell pro level card