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]

@BackWindowz Ask AI agent, ngl
like GPT or Gemini Argon
let them run in some agent in your machine
BTW for windows method I recommend you to prepare a usb disk or so

[琥珀青葉@KohakuLab @KBlueleaf]

Now thanks to community resources like gpu-unlock-ada or gpu-unlock-4080s. I successfully unlocked my 4090 on "Windows" system.

All the details are released in https://t.co/UOfas5elK9

Plz check it out!
https://t.co/i4exHAkipR

[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]

@haor233_re 當時一張27500rmb買的 現在一張55000起價
我買了四張 樂

[琥珀青葉@KohakuLab @KBlueleaf]

@GuillermoCR_ merged, thx for contribution UwUb

[琥珀青葉@KohakuLab @KBlueleaf]

@kdpisda it can be my msi afterburner's config not applied correctly XD

[琥珀青葉@KohakuLab @KBlueleaf]

@nomoreAcetoplay 是說5090的mxfp8/nvfp4一開始就沒鎖
Fp16/bf16有鎖不過還不知道怎麼解鎖
但是就算有鎖還是比4090快，因為帶寬兩倍
4090的實際使用性能被帶寬影響了

[琥珀青葉@KohakuLab @KBlueleaf]

@itizjuann @coffee_sal you won't get faster on decode bcuz decode is purely memory bandwidth bound
you will get speed up in prefill but no more than 50% bcuz pytorch matmul is optimal case already

[琥珀青葉@KohakuLab @KBlueleaf]

@trenchplanet how can you unlock nerfed tensor core in a GPU have no tensor core

[琥珀青葉@KohakuLab @KBlueleaf]

It is interesting that, ppl don't discuss the risk about use exploit to patch things inside GPU firmware may be dangerous but claiming "NV nerf your performance for stability" or so.

BRUH plz check every detail before you discuss anything, NV only nerf "some" specific dtype combination's performance.

1. Tensor core's most heat comes from multiplier, accumulator is very small compare to other part inside tensor core
2. in consumer card, FP16 matmul FP16 accumulation can run in full speed, but FP32 accumulation can only run in half speed. Which means your card can run multiplier in full speed
3. in 50-series, Mxfp8 and NVfp4 can run FP32 accumulation in full speed. which means your accumulator can run in full speed. (those 2 dtype even have more multiplier and need to handle the block scale lol)
4. so there is NO reason such nerf is designed for stability (the difference between 3090/3090ti can be about quality or stablity, some flaw in chip make NV need to block some SM and result in weaker version). And Usually the raw power of pro card is even worse than consumer card (like the SM's frequency, vector FP32 speed, if you compare p6k and 5090 you will find this)

The nerf is fully designed for spearate the market of pro card and consumer grade card. and use same design (with firmware cap) is actually cheaper in tape out.

The only risk of such "unlock" is if you don't prepare proper fallback entry inside your grub or UEFI, you may lost your OS. That's all.

The risk is even smaller than use apt to install nvidia driver.(X)

[琥珀青葉@KohakuLab @KBlueleaf]

@sadlemonjuice yes, exactly
actually not really for "windows" but "for any UEFI OS"
bcuz it run EFI application before any os to patch the register value
Then the EFI application start the OS you want to run, which means if you want, you can say this works for FreeBSD

[琥珀青葉@KohakuLab @KBlueleaf]

Latest news on 4090: fully reproduced the EFI applicaiton path and provided more details in my repo:

https://t.co/rNU4xZc2bC

[琥珀青葉@KohakuLab @KBlueleaf]

Now thanks to community resources like gpu-unlock-ada or gpu-unlock-4080s. I successfully unlocked my 4090 on "Windows" system.

All the details are released in https://t.co/UOfas5elK9

Plz check it out!
https://t.co/i4exHAkipR

[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]

@unixpickle TL;DR:
1. the whole nerf is not about stability, bcuz some specific combination originally have no nerf means the hardware are fine
2. the nerf is fully firmware itself half the issue rate, not about any precision, heat, frequency change

[琥珀青葉@KohakuLab @KBlueleaf]

@unixpickle https://t.co/QfOWJ8SpWW
check my post plz

[琥珀青葉@KohakuLab @KBlueleaf]

It is interesting that, ppl don't discuss the risk about use exploit to patch things inside GPU firmware may be dangerous but claiming "NV nerf your performance for stability" or so.

BRUH plz check every detail before you discuss anything, NV only nerf "some" specific dtype combination's performance.

1. Tensor core's most heat comes from multiplier, accumulator is very small compare to other part inside tensor core
2. in consumer card, FP16 matmul FP16 accumulation can run in full speed, but FP32 accumulation can only run in half speed. Which means your card can run multiplier in full speed
3. in 50-series, Mxfp8 and NVfp4 can run FP32 accumulation in full speed. which means your accumulator can run in full speed. (those 2 dtype even have more multiplier and need to handle the block scale lol)
4. so there is NO reason such nerf is designed for stability (the difference between 3090/3090ti can be about quality or stablity, some flaw in chip make NV need to block some SM and result in weaker version). And Usually the raw power of pro card is even worse than consumer card (like the SM's frequency, vector FP32 speed, if you compare p6k and 5090 you will find this)

The nerf is fully designed for spearate the market of pro card and consumer grade card. and use same design (with firmware cap) is actually cheaper in tape out.

The only risk of such "unlock" is if you don't prepare proper fallback entry inside your grub or UEFI, you may lost your OS. That's all.

The risk is even smaller than use apt to install nvidia driver.(X)

[琥珀青葉@KohakuLab @KBlueleaf]

@auryn_macmillan 30series and 40 series both have some leaked vbios that support larger vram size so those china ppl just put more vram onto their card

[琥珀青葉@KohakuLab @KBlueleaf]

Now thanks to community resources like gpu-unlock-ada or gpu-unlock-4080s. I successfully unlocked my 4090 on "Windows" system.

All the details are released in https://t.co/UOfas5elK9

Plz check it out!
https://t.co/i4exHAkipR

[琥珀青葉@KohakuLab @KBlueleaf]

But since 5090 have no-nerf mxfp8/nvfp4 and 2x bandwidth, my benchmark shows after this unlock trend, the ordering is still 5090 > 4090 > 3090 > 2080ti.

Which is interesting.

[琥珀青葉@KohakuLab @KBlueleaf]

In the unlock I achieved +80% speed on mmapeak (my modification, which is more accurate, old mmapeak have all zero input which make result unclear)

and +50% speed on pytorch. which means 4090 now nearly catch up AIO version 5090 in fp16/bf16.

[琥珀青葉@KohakuLab @KBlueleaf]

@G36_maid 沒事，我當時也沒上交大
https://t.co/FTrW8JozsH

然後我現在github 7.5k star
HF 3.5M download
Pypi 6.5M download

UwUb

[琥珀青葉@KohakuLab @KBlueleaf]

@GodResourceInc No overclocking