https://github.com/Zagitta/gpu-unlockHas there ever been a more semitic company than nvidia?
>>110296891Yeah
>No Sneed, no Feed, no Seed, just pure Chuck's Fuck and Suck, running entirely on the fancy German car
>>110296939This is what passes for comedy among shartytroons.
>>110296943Damn those niggers are pretty funny
>>110296891I tried this and it works but you don't actually see any gains because they're limited to shit you're not gonna use
>>110296891amazing attack vector for a root kit
>>110296891I got news for you: AFAIU, all consumer electronics companies do this. it's cheaper for them to produce in single lines and "differentiate" their products based on how each product behaves according to some tests. it's called "binning", I think?
>>110297418no sir, not always. maybe all have done it some, but no, and definitely not a "top tier" product.This is the height of shame, imagine telling people that it's the best card
>>110297112Its a free 60%-100% speed increase for any CUDA workload. This isn't for fucking gaming man
>*burns you're gpu*nice unlock
Would this improve ray tracing for gaming?Also dlss uses tensor cores. This should make dlss better too right?Doesn't it use tensor cores?
>>110297640it's not for any CUDA workload, you don't get any benefit on quantized models and that's all you're gonna use, and you only get benefits if you quantize your kv cache which really dumbs down your model, you get an improvement in prefill but that's nothing on your runtimeit only really matters if you wanna train your own model or run some imagegens
>>110297619>definitely not a "top tier" product.are you retarded or just pretending? it's exactly the contrary: companies that produce top tier products are the ones that follow this process the most, since they are the ones that benefit the most from tiered consumer markets.just google binning, you fucking retard
>>110296891Yeah, Intel. It was even founded by an american, an israeli, and a hungarian. You can't get a more semitic combination than that.
>>110296891>BF16no one running stuff at home uses that.also generaly the bottleneck is the memory bandwidth anyway.
>>110297686>you don't get any benefit on quantized modelsAnon, Ampere doesn't have FP8, even if something is quantize it has to convert first to FP16 or BF16 to actually be processed. Its why quantized models are slower due to that extra step in processing them. This still speeds up the actual computation of said model.
>>110297699I run my KV cache in BF16, its about 10% faster for doing so.
>>110297713it literally says it in the repo>llama.cpp / Ollama with quantized GGUF | 0–15%, mostly at long context>INT8, FP8 and INT4 models | No changethe 15% is because of the FlashAttention and llama.cpp uses INT8 not FP16, also:>which is what nearly all AI training and unquantized inference useswhy did they have to specify unquantized if it apparently works with quantized too?
>>110297713my bad, it does use FP16, but it uses FP16 accumulation, while the speedup is for FP16 with FP32 accumulation only
>>110296891someone needs to figure out how to pwn falcon in pascal so i can modify the fan tables in my pascal cards
does this affect gaming performance at all?
>>110296891You forgot to include how there's also a mod for enabling frame gen on RTX 30/20 series.
>>110297686>it only really matters if you wanna train your own model or run some imagegensheh you clearly dont use local AI yourself once your context gets to around 100k, prefill takes as much time as token generation, being able to keep it 2x faster is better.also this helps with just everything in general for AI, yes the carda are still bandwidth bound but latency and artificial cache misses still matter to llm inference speeds
>>110298295>around 100k, prefill takes as much time as token generationsucks to suck, I guess
>GPU sounds like a jet engine when genning now.It's gonna burn my chip, isn't it?
>>110296891>just fuck with your kernel. what's the worst that could happen>checking issues>a bunch of people reporting black screens on bootyeah nah, call me when this is save and idiot-proof
>>110296891Would uncapping this reduce the card's lifespan? Considering the (absolute) state of the hardware market right now, I'm not so sure it would be worth it.
>>110299007yes, it increases power usage
>>110296891What about 50 series? Do they see a speedup too?
>>110296891They do this already for as long as the company exists.Reminder that until a few years ago, the consumer cards would try to detect if you GPU passthrough them to a VM and refuse service. Because that functionality was locked to enterprise only.
>>110299072that fucking machine killed the cat didn't it
I saw this few weeks ago and wondered how many 4090 card going to be bricked by end of year just to imagegoon faster.
>>110299198>yfw nvidia leaked this themselves to inflate prices even more
>>110297619It's called late capitalism, zoomer.ACAP, All Corpos Are Bad.And of course, AJAB.
so the question is, this was intentional? and why deliver capable gpu at these speeds anyways? why not just reduce tensor cores anyways?
>>110296891>+30-55% image generationWow.I don't know if the models I use are BF16 or whatever, I never paid attention to this detail.There is a Windows port but I suppose it won't be compatible with Windows 7 and old 475.14 drivers, will it?>>110297645>>*burns you're gpu*>>110298420>>GPU sounds like a jet engine when genning now.>It's gonna burn my chip, isn't it?Is there really a risk?Our current GPUs may be the last ones we ever have, so we better take good care of them.
>>110297418>it's cheaper for them to produce in single lines and "differentiate" their products based on how each product behaves according to some testsThis reminds me of a rumor in ~2003 when I was a teenager, I never knew if it was true or not, that AMD Durons and Intel Celerons were actually real Athlons and Pentiums with features only disabled but physically present in the chips. The same people who claimed this also claimed that all CPUs of a same generation, from 600MHz to 1GHz, were actually the same, which all overclockers knew was not true.
>>110296953>>110296939glistening gem
>>110300363its called binning and it is kind of true. you build a chip, like an intel chip and depending how it test or what works it will get binned differently.say the onboard graphics are broke you sell it as a cpu without onboard graphics, say some cores are bad, you sell it as a version with less cores
they should unlock the workstation features like mgpu too, I know its been done for some turing and older pascal cards but for ampere and newer
>>110296891it's so that your shit doesn't burn down
>>110296891Fuck. I wish I bought a 3090 instead of a 5070. Yeah it had a PCB crack repaired but come on.
>>110300476Crazy times when 3060 is more valuable than 5050
>>110299007Yes. Don't tell the retard ITTs especially OP that it overheat your GPU in a way that'll kill it many times faster. Good luck buying a new GPU with the current prices then.
>>110300601Won't it just throttle if it overheats?
>>110299007>>110300684This IS the throttle. Removing it will cause problems. Also for the poor retards who didn't plan for their uncapped GPU to require more power usage and short their system.
>>110300363>>110300385Yes binning is normal because of how making high end chips like CPUs and GPUs work, they will be randomly broken in a few spots and then you can disable that part and still have a functioning chip.They also randomly can run at higher frequencies (and more efficiently, lower voltage).It does mean it's the exact same chip, just more or less broken and better or worse clocks.This can be good for the consumer.What is bad, and what NVDA is going here is artificial segmentation, very different.>>110300363> it's cheaper for them to produce in single lines and "differentiate" their products based on how each product behaves according to some tests. Ironically they are doing less binning now. Instead more artificial segmentation, locking away features and vram.>>110300731Nope. The throttle is the power limit. This is just a artificial instruction rate limit. You'd think you'd realized because the sold-as-unlocked cards use LESS power, not more.
