Tensor Pinning. gpu-unlock Debut! Will It Increase your TensorCore Performance by up to 50-100%. Yes!
We take a look at Tensor Pinning or squeezing the most speed out of Tensor Core disabled 2080, 3080, 3090, 4080 GPUS!
A rapid discovery has been found in Nvidia GPU's that looks to increase tensor core performance dramatically once a unlock has been applied. How was this happening? It looks like it was an economic decision where they would combine GPU runs, then simply disable half of some capabilities on some cards before marking them as 'consumer' and kicking them out the door. The higher end cards would then be configured for full capability, and sold to the hyper-scalers and commercial outfits with a different marking on them. Nobody was the wiser. From an economic standpoint this made a lot of sense, as these runs cost millions of dollars each, and to get as much purpose out of a single fab run as you can - it would be most feasible. The consumer would happily receive a publicly 'stripped' variant GPU which had some functions disabled, and a second 'everything enabled' variant that would get sold for more to the deep pockets. Everybody was happy. From Nvidia's end it protected their high-end orders from getting watered down but also saved them from the cost of two fab runs. Nobody really looked back then because there wasn't a world-wide screaming demand, like there is today for a near infinite demand for inference. That all fell when gpu-unlock debuted:
How The Unlock Initially Was Found:
In Linux it discovered by shifting the value two small registers, and ensuring it was done before the driver could disable it.
Register | Original example | Unlocked value |
|---|---|---|
|
|
|
|
|
|
The unlock recipes soon followed - and varied for each card, and as each one was discovered people would update the git, as of right now the following cards look to have had their unlocks found.
If you want to dive in here is a quick-start guide:
Comparable gits also sprung up with various instructions. This is atypically always the norm now, everything has 5+ copies now even Tower Defense.
The 2080 / 4080 model lockouts then fell.
The numbers were very significant, anyone using any type of Tensor core had a very vested interest. 55.9 TFLOPs - 111.0 TFOPS on FP16. That's a double!
A confirmed unlock on the 4080 fell here:
The git was moving fast, and forks showed the progress.
woodsonl
Thus far the 5090 unlock has not fallen however is under intense research to see if it is possible. The potential to possibly double fp-16 / fp-32 inference awaits if they can crack it.
Here is a github that looks to be updating by the hour. Atypically with this many people looking and sharing someone finds a way and then tells everyone else. If polymarket had a book I'd give it less than a week.
Zhaojp-Frank