GeoHotz is writing software for us, we will finally beat CUDA! Fuck ROCm!
https://geohot.github.io//blog/jekyll/update/2023/05/24/the-tiny-corp-raised-5M.html
AMDChads!
GeoHotz is writing software for us, we will finally beat CUDA! Fuck ROCm!
https://geohot.github.io//blog/jekyll/update/2023/05/24/the-tiny-corp-raised-5M.html
Damn, sounds fun. I would like to work on something grand and 100% technical like this instead of the shit that I'm doing on my job. Too bad I don't live in the US and my knowledge is probably not that good for this.
>Mercury Balance
lmao
what drives someone to deposit 5 million dollars at some shitty pajeet tech powered fintech startup
>t. poorfag
it's been amazing watching this clown fail upwards
Benefits of being in the tribe.
>GeoHotz is writing software for us
he's not writing shit. he's a well known compuslive lying grifter that has never produced anything of value or importance since he stole the exploit for bsd and packaged it to break into ios and playstation.
this motherfucker was using stack overflow to copy paste code during livestreams and he was forced to leave twitter after he was exposed as a talentless hack. he has no credibility.
everything he touches turns to shit. ask him about his "100% self driving" scam that he tried to pull a couple of years ago. did a whole presentation for it but had no technology to show and used mockup pictures, same pictures on his website, to promote it and to try and get funding.
>reputable hacker
you don't know shit about this abysmal failure.
this is accurate. and yes, there are irc logs.
>100% self driving
That's probably his best work. You should see his comma robot or his cryptocurrency. Now he is saying because human brain is equal to 100 Nvidia GPUs, we will have AGI in 10 years.
>Now he is saying because human brain is equal to 100 Nvidia GPUs, we will have AGI in 10 years.
thats a stupid take that we have pointed out to him so many times but his brain probably is running at a single thread
https://aiimpacts.org/brain-performance-in-flops/
the worst estimates puts our brains at zetta flops
the best estimates are ronnaflops
remind us when the first dgx with epyc was launched anon
please do
First he fixed Twitter and now he will kill Nvidia. Can anyone stop this man?
i dont think he did fix twitter he went there and said hed starting nuking microservices and made the site shit for a couple months kek
I want to believe.
yeah just how he fixed twitter in his few weeks he was there, just like he claimed.
while this guy is a scammer we could all learn from him how tons of overconfidence can still bring you somewhere in life.
id rather be anon than another faggy tool like him or you
life isnt about working
If it is that easy why hasn't AMD done it? ROI easily in 100x-1000x range.
amd had not been taking ai seriously until recently that was obvious by the state of rocm, it hardly supported any cards
Because they're working on feature parity with CUDA, not a subset of PyTorch features. ROCm is much broader in scope than this and involves supporting much more than the tiny handful of AI models that aren't designed to run on PyTorch/Tensorflow/JAX/etc.
Because AMD, strategically, doesn't care. They want supercomputer customers. Everything else is incidental, which is why they've been completely fine with rocm breaking or never working in the first place on so many of their cards, even some professional cards. Because no big datacenter science customer ever bought a crapton of them.
It's better than the Canadian alternative.
>dude we're going to change the world bro
>okay but can you give us your complete business plan
>we need milestones, no, more and concretely achievable milestones so we can value your offering in its entirety
>okay but we need evidence you're serious - can you get some funding from a canadian bank and the government
>as a canadian bank, i am unable to support actions that would be overly disruptive to one of our major clients
>hi i'm from the government can you fill in all these forms
>wow that's risky can you tone it down and fit it into our checkboxes? we'll give you free tax rebates, if we approve
>remember to have your forms in on time for approval, exactly in the form we want to see them done up
So you either surrender your awesome world-changing potential idea to become essentially an extension of a government jobs program, or you invest in real estate.
Even Canada's most world-changingest charity ended up being little more than a real estate business with subtly partisan marketing to kids tacked onto it.
>Because AMD, strategically, doesn't care. They want supercomputer customers. Everything else is incidental, which is why they've been completely fine with rocm breaking or never working in the first place on so many of their cards, even some professional cards. Because no big datacenter science customer ever bought a crapton of them.
they don't buy them cuz no work
>They want supercomputer customers
No, that's all they can get. There's no difference between a supercomputer and a modern datacenter. AWS configured some EC2 instances and placed 42 on TOP500 for shits and giggles. Governments don't want to pay for a fleet of nvidia's extremely expensive superPODs, so they go for the bottom bidder. When you're a crony for some government official selling toilets from mexico to the US government it's a nice deal but not when you're a chipmaker. Everyone else turned their noses up at supercomputer contracts.
>thats all they can get
first of all lets clear something you are mentally challenged now that we all agreed on that let me give you a rundown
the datacenter and generally supercomputer market is extremely conservative they do not change their computers in like 5-10 years they either built on top of it or upgrade them
the ORIGINAL plan for frontier exascale was with intel using raptor lakes and being at 1 exaflops with 150MW of power
the amd system currently is at 1.2 exaflops at 21MW of power...
it is set to replace both titan (amd opteron+nvidia 2013) and titan 2(intel+nvidia 2018) since its literally 6 times faster than both of them
the el capitan a 2 exascale system that was originally being built by intel was a 221MW power supercomputer
with amd its a 2.2 exascale system using only 40MW of power...
and its set to replace ats-2 a supercomputer that is barely 5 years old based on nvidia
and if the rumors are true amd also won the JU competition of europe (1 exascale and 5 other supercomputers ranginf from 20pflops up to 500 pflops)
you know why they keep choosing amd?
not because they are cheap they arent
but in the grand scale of things their hardware is supported by a trully open source framework
which neither nvidia or intel can possibly give nvidia will never make cuda open source
the very same thing is happening in datacenters too amd literally RAPED intel in the datacenter segment for q1 2023 they lost 95% of their income qTq to amd
its really simple amd's offering right now with MI300 and epyc along with miopen being totally open source is simply superior to...well anything
>but in the grand scale of things their hardware is supported by a trully open source framework
>totally open source
obese amd fanboys that shit in the street and use the hands of homeless children as toilet paper never cease to surprise and amaze.
im guessing you are old enough to drink but not old enough to remember the state of cuda before 2014
aka it was shit and it took nvidia and LWR engineers to rewrite the damn thing for over 2 years in order to be in a somewhat decent state to be used on titan
same thing literally will happen with miopen like it or not open source will always be superior anon
>you know why they keep choosing amd?
>not because they are cheap they arent
It's pretty obvious that's exactly why. AMD's data center business posted $1.3 billion in revenue including Epyc, Xilinx FPGAs, and instinct. It's obvious that no one's buying instinct. Epyc is the main seller but Xilinx was posting up to $1B in revenue quarterly before the merger. At a meager $1.3 billion in revenue I doubt their GPUs are doing more than like $100 million a quarter, if that. Once the dumb contracts Raja signed AMD up for on his way out expire they'll probably kill instinct altogether. It's just going to increasingly become a cost center for R&D.
>nobody buys MI
are you ok anon ?epyc servers ONLY and always come with MI cards........................
Probably the most retarded thing I've read in this thread. Putting aside the fact most server integrators and big cloud providers will offer EPYC with nvidia GPUs, nvidia is one of AMD's largest partners. At this time superPODs only ship with EPYC CPUs. That will change in the future as nvidia focuses on their own CPUs though, e.g. Grace. No one's buying MI cards. No one wants MI cards.
amd has dropped nvidia from their hardware manufacturers list since 2020 anon
stop living in the past
The first DGX servers to use EPYC didn't even ship until 2021 tard lol
You are retarded. AMD would love regular computing costumer but their shitty ass drivers and their train wreck of an SDK makes it impossible to compete unless it's for gigantic compute that will get a specific workload and can be babysat and tuned by an army of dev and scientists
The proof is that they literally have no presence in datacenters (hint: super computers aren't data center clients lol). Try to find a single public or even private/custom cloud that offers AMD cards about the years old mi25. Hint, you will find an azure preview that started a year ago and went nowhere presumably because no one gave a shit lmao
And even if you go through all those hoops AMD offers nothing over nvidia. At best the hardware may cost a little less.
Exactly. People seem to think that corporate compute users care that much about sticking to Nvidia. But they don't, yes the enterprise shit they sell is expensive but not prohibitively so and AMD has very comparable pricing for its enterprise or MI line. Why would anyone put themselves through the pain of rewriting librairies or even use barely tested ports that are after thoughts compared to their Nvidia equivalent? That's why I laugh everyone says that x popular library now supports rocm or whatever, like have fun beta testing stuff forever in your production environment if you ever try to do that.
Sure, if you want to run SD at home it might work well enough, but I couldn't even imagine having a datacenter and messing around building a service around AMD cards. Nvidia gives you a complete, very well tested suite that makes it a breeze (nvlink, vgpu..), which makes any compute advantage irrelevant for AMD. The only possible users are again, well funded labs that run very very specific workloads and have total control over the whole stack. To the point where patching broken stuff is viable.
It's sad because it makes it harder for regular users to get hold of truly powerful compute cards since the lack of usage upstream (from the user's pov) makes everything janky to use.
Non huge companies still will likely just continue to use PyTorch if their data scientists are used to it.
Rooting for tinygrad but the PyTorch/huggingface ecosystem is insanely popular.
Isn't this guy a meme? I genuinely don't know anything about him.
>t.
Don't know any of his newer accomplishments but he did accomplish to hack the PS3 and was able to install CFW on it.
Sony even went against him in court, though nothing came out of it so Sony went and paid him to not touch any Playstation forever.
proof the he hacked the PS3
fake and gay
He didn't hack the PS3. He watched fail0verflow's presentation on it, then did it and drummed up some press over it for himself. The guy is just a grifter which is why he's not really known for having done much of anything except press releases.
i hacked my ps3 as well, how come i'm not famous
Because you didn't put any work into reversing the firmware, decrypting it and make it possible to load unsigned code in a time where this was simply not publically possible.
You think you deserve as much credit for using a screwdriver to change an outlet as the ones who used a screwdriver to build your house?
>reversing the firmware, decrypting it and make it possible to load unsigned code in a time where this was simply not publically possible.
goyhotz literally didn't do any of that either
>[Citation needed]
look up who fail0verflow is, kid
He didn't hack the PS3. He simply released the keys based on an exploit found by fail0verflow. Very few people know what really happened.
I think the IRC logs are still around though, right? I remember one of the actual hackers wrote an article about what actually happened years later but it obviously didn't matter at that point. Does anyone have a link? Or am I just hallucinating it from the IRC messages
He managed to bypass hypervisor.
He singlehandedly fixed Twitter by adding a close button to the "Sign Up/Login to view more tweets by this user" dialog.
Does that sound like a "meme" to you?
Didn't think so.
He was hired to fix the search feature though.
The funniest thing is that it's standard for those sites to have a login wall after a couple of scrolls, so there probably was a research done to show that putting up the wall was beneficiary in some way and he just went and undid it.
I still Like it, because I just close any site that does this. I did discover nitter before the "fix, though, making it useless for me
And then the cunts removed it.
Last I checked it was still there.
I can stare at it for hours. It's a technological marvel.
They didn't, he just never managed to properly remove it.
It does sound like a meme and it did fix twitter, we live in a meme reality.
He became famous after he jailbreaked his iPhone as a teenager and hasn't really done anything noteworthy since.
israelitevidia is finished
so is this dude actually a tech wiz or just someone who is good at marketing?
he hasn't successfully marketed anything
he has successfully hacked and created tech that nobody actually uses
"marketing" as in self-promotion
he's self-driving car start up is profitable, retard.
>self driving cars are a scam
- George Hotz
ok? i mean it doesn't change the fact that he it's still profitable
that implies that he's scamming his customers
NTA but what's wrong with that? Scamming sheep is an excellent business model.
>NTA
kys
KYS ESL.
Also, his comma.ai isn't profitable. It'll likely shut down.
>Also, his comma.ai isn't profitable. It'll likely shut down.
it is profitable and they aren't shutting down soon. they recently got bigger offices
>they recently got bigger offices
i am not part of the convo, but a startup getting bigger offices is by no means sign that they are doing well. A ton of times it is a sign that they are burning money. Or that the landlord is related to the ceo or somebody
they are profitable. one of the few startups that are actually profitable with an actual working product
KYS ESL
literally who
more like a drama queen than a tech wiz
i hate VC startup business naggers so much
how does this keep happening
>dude we're going to change the world bro
>we're android for cars bro (oh we failed but that's normal startups fail haha)
>now we're android, but for compute, bro
rip amd
>It will work for a bit, but then it will start to hurt.
Speaking from experience?
>But how do you make money doing that?
>We don’t. We raised $5M from an amazing set of investors who are aligned with real value creation.
>We make money selling computers for more than they cost to make. (preorder a tinybox today).
I have no idea what a tinybox is, i suppose its that rack he presented on twitch but I don't see how this is supposed to work either way. People can already buy computers on the internet where is the value being created here?
>cute af guy
>braindead söy retard
every fucking time
hasn't done anything since the PS3 days. guy's a joke looking to stay relevant
He just became top 10 ctf player in 2014 right after the ps3. https://ctftime.org/team/5162
Pretty much this, he just follows the hype rain around
>tinygrad is based on Karpathy's micrograd
>jumped on the self driving train, delivered nothing beyond what's already been done
>jumped on ML hardware train, delivered nothing
>jumped on the twitter train, couldnt last 12 weeks
I tried watching his AI hardware stream and halfway in he just gives up and says that hardware is too hard and slow because you have to actually test your stuff and make sure that it works which is too much work for him, then he spends the rest of the stream lamenting how google wont sell their ML hardware to him
Basically his whole shtick is taking something popular someone has already done, writing some wrapper code around it and calling it a day
Attention span of a goldfish
like a true open-source scholar!
I still use my comma 2. just werks , but after hours of re-flashing and troubleshooting thanks to an update and obscure hidden info in github issues, again, true open-source experience
Why not just use rusticl or vulkan? Pretty sure the AMD carda are good for gaming and the mesa drivers deliver the promised performance.
would be very cool to see my 7900 xt be useful for this shit
>The human brain has about 20 PFLOPS of compute
Dropped. The brain is not a computer, you can't say how many FLOPS it has, just like you can't say what the frame rate of the human eye is. It just doesn't make sense on a fundamental level.
why not just fix ROCm then? It runs fine for me though I'm sure there's tons of shit to make better
i mean amd has frontier and el capital now...
tools and documentation are on the way one way or another for miopen
>https://geohot.github.io//blog/jekyll/update/2023/05/24/the-tiny-corp-raised-5M.html
Can someone explain to me what he meant by that as if I were a 5 year old, or a golden retriever?
>Can someone explain to me what he meant by that as if I were a 5 year old, or a golden retriever?
He wants to compete with Nvidia on the AI market.
He will start by doing his own AI framework that can fully utilize RDNA3 cards.
Then much later he will make his own hardware.
>With the way AI is going, we risk large entities controlling the majority of the compute in the world. I do not want “I think there’s a world market for maybe five computers.” to ever be the world we live in
>The goal of the tiny corp is: “to commoditize the petaflop”
>I don’t want to live in a world of closed AI running in a cloud you’ve never seen, I want everyone to have an AI that they own, both training and inference. I want compute to be available from 50 different companies all competing to drive the price to zero. And I want an open source framework to run cutting edge AI on any one of those 50 chips as seamlessly as Linux supports 50 network cards
Absolutely fucking delusional to think that OpenAI can buy out nvidia let alone fabs.
>he used a technology invented by AI to send his message back in time
>we kill OpenAI as he instructed
>the message will never get sent
>a bunch of zero cost movement operations
Physically impossible. At best you can re-index on-the-fly, but some indices will be less efficient than others when it comes time to load from VRAM into registers and groupshared; this is a cost. He repeats this claim with different, extra nonsensical phrasing on the tinygrad page:
>MovementOps operate on one tensor and move the data around, copy-free
It will be a huge flop.
Even if he succedes on doing the good software he is talking.
The next step will be to convince all those researchers who are already used to work with Cuda, to learn how to code with another stuff.
I was wondering where the hackers went, in particular Geohot came to mind, and finally a reputable hacker like Geohot is working on it (but he's not at his prime, so we can only cross our fingers now.) Hopefully his involvement brings more people into the show
>a reputable hacker like Geohot
really showing how little you know
what a piece of shit. he is literally a scammer and so many people are going to fall for it because he sounds so confident.
>geofraud
>Melon Musk reference
way to out yourself as a grifter, lmao