Nvidia CEO: We bet the farm on AI and no one knew it
techcrunch.com
techcrunch.com
And yet all these work and success from Nvidia was because of, if you read 90% of HN comments for the past 2 years; Luck.
They could have given up at any point in time for the past 20 years and simply not do anything CUDA or GPGPU related. Because who would want to do that when vast majority of those investment were not even bringing in much revenue. Like Intel decided to cancel Larrabee. They persevere and hit the Jackpot some 10-15 years later. But all of this was because of; Luck.
Yes. Luck plays a big part. They could have continue another 10 years and they may never find the Killer App for it. But to ignore all the investment and work for such a long time and pin it down to Luck was about as rude and as disrespectful it can be. Especially on a forum which was started by VC with the spirit of entrepreneurship.
But their Zen bet has paid off and now they're playing catch up.
Intel, OTOH, I'm not sure what their excuse is.
Intel had grown accustomed to internal mediocrity and got too big to root it out; the fact that accounting runs the company instead of someone with a vision doesn’t help. Accounting doesn’t like risky bets and Intel needed some to work out instead of cancelling everything left and right.
You're making it sound like AMD being broke at that time was some unfortunate accident due to external events, and not the result of their own blunders.
Nvidia had billions due to great products and great business decisions on their part, and AMD was broke entirely due to it's own actions, by having average products on the CPU side and making bad business decisions at the time by spending way too much money acquiring ATI, and then selling off their golden goose, the Imageon mobile GPU division to Qualcomm for pennies right at the beginning of the smartphone revolution.
It's a miracle they managed to turn things around and not end up like SGI and 3dfx, bankrupt and having their carcass devoured by Intel and Nvidia.
https://en.wikipedia.org/wiki/Advanced_Micro_Devices,_Inc._v....
That NVidia has maintained this push for the last 2 decades makes one wonder what other tricks they'll have up their sleeve.
Even something as simple as replacing TVs and monitors is a no-brainer. It's just a question of whether comfort and quality can be achieved.
For games, to the extent that VR works for them... sure, I could see that use case. But for, say, business meetings? Uh, no. Or to hang out with friends? Sure, maybe, but how would they monetize it well enough to make the investment worthwhile? To make it compelling enough that I would visit?
I'm not saying that the use cases don't exist. But either making compelling VR experiences has to be cheap[1] (which I don't see on the near horizon), or they have to add enough value that people will pay for them in some way. That strikes me as uncertain.
[1] If they get cheap, then sure, although then it will have the same problem CGI now has... there is so much of it that it has lost a lot of its magic. It's hard to be common and compelling.
People were saying the same about home computers in the '70s and it was relatively true for that period if you didn't look into the future and only considered what home computers looked like back then: janky machines built from radio shack parts by nerdy enthusiast tinkerers with electronics knowledge. Only when they were presented with prebuilt and polished products with usable apps that didn't require technical knowledge, did people see the light that home computers are the future and will become mainstream for all consumers not just enthusiasts.
Nobody can know yet how mainstream VR will look like exactly and who will get to dominate the field and dictate the product direction (Apple, Valve, Meta, etc), but for sure it will happen in the future, even if for now it's just a tinkering toy for gamers and enthusiasts with money.
You can't have a pair of glasses that is simultaneously opaque enough to give you the illusion you are in another place AND breezy enough so that your eyes won't sweat. Air is just much bigger than light, so you can't block light but allow air to pass through. So a VR headset will always be sweaty on the face. AR will never be able to create realistic opaque backgrounds.
You also can't have a holo-deck-like experience with glasses. The only reason the holo deck captures our imaginations is because it had space, smells, sounds, touch. All of those things are impossible to achieve with physical devices of any kind, or at least very close to it. The only thing that may create a holo-deck like experience would be based on brain-computer interfaces, if those are even truly possible.
Finally, to get many of the claimed benefits, even if the headsets were good enough to deliver them, you also need complex recording software that actually handles all of this. You can't have a nice realistic image of all of your colleagues in a VR meeting unless they each have a 3D camera setup, and everyone has enough bandwidth to actually receive and send all of the 3D videos, and do so with latency similar to audio traffic.
These are at least three fundamental problems that make the whole VR/AR craze just certain to fail. Again.
You don't need the exact fictional holo deck device to have a VR product that sells. 150 years ago Sci-Fi writers imagined we'd have robots in our homes doing our chores, and we do have them today, except not humanoid robots doing the washing and vacuuming by hand for us like we originally imagined, but we have fixed function dedicated robots for each task: dish washing machine, clothes washing machine, Roomba robo-vac, etc. They're a rudimentary and limited far cry from the fantasy and capable humanoid robots in sci-fi novels, yet they're ubiquitous today and sold by the millions. Same will be with VR, it will be more limited than the holo deck but it will sell at the right price/feature combo.
>You can't have a nice realistic image of all of your colleagues in a VR meeting unless they each have a 3D camera setup, and everyone has enough bandwidth to actually receive and send all of the 3D videos, and do so with latency similar to audio traffic
People also fantasized about video telephony like in star trek and yet those challenge got overcome with the introduction of 3G and camera phones and has improved ever since. Tech will also improve for VR. The iPhone already has had a 3D camera since a while now.
>These are at least three fundamental problems that make the whole VR/AR craze just certain to fail. Again.
It will fail today, but it will succeed in the future, even if you're too dead set to not see it.
But we don't have a clothes folding machine, precisely because the task is infinitely harder, from a first principles view, than anything a futurist imagined. VR/AR is the same. It's not just scaling up computer power because it's a matter of not having the physical ability to manipulate reality, energy, and matter the way we need.
Science isn't magic. How do you do haptic feedback to hands? That's an essential part of any AR system that isn't just a gimmick, and yet it's basically impossible without clunky gloves. How do you prevent damage to the human eye over extended use like a 9-5 job? The human eye did not evolve to "look at" different things that are in actuality on a screen an inch from our eyes, it really upsets the brain and the muscles that control vision, which is why VR/AR can be so tiring on your eyes. That cannot be innovated away.
Who said VR would need gloves and haptic feedback to succeed? We're talking about virtual reality here, not simulated reality. For the latter better wait for Elon's neural link or whatver we'll get that plugs our brains into the Matrix and simulates reality.
People still use mice and keyboards to interact with UIs despite having the ability to use touchscreens. Just because one thing exists, doesn't automatically mean the other dies.
>How do you prevent damage to the human eye over extended use like a 9-5 job?
Who said you need to use a VR device from 9-5 for it to succeed? You don't stare at your phone screen or tablet for 8h/day, do you? And yet you most likely have one.
> The human eye did not evolve to "look at" different things that are in actuality on a screen an inch from our eyes
Have you seen what lenses can do with light? Like move the focal point of a picture much further away? They're pretty big tech in cameras, telescopes, binoculars, and .... oh, these optical things humans wear on their faces, a couple of cm from their eyes to fix their vision issues, their called glasses I think.
How about we use those lens thingies to move the focal point of a screen that's 2cm from your eyes to 2m from your eyes? I'm telling you dude, whoever invents this tech is gonna be big.
I really liked immersive VR I experienced in research lab settings like CAVE systems, over 20 years ago. I'm probably among the people most tolerant to artifacts like frame rate jitter and lag, and most easily able to still get an immersive pop out of it. But I'm not an early adopter. I don't see the value or place for this single-purpose equipment in my life, nor my budget.
When mobile phones were just phones, I saw them getting smaller and lighter. I even fantasized about them shrinking until it was just the ear bud. I was willing to hand-wave some sci-fi UX without a display nor buttons. But, I didn't (and still don't, really) appreciate how people would care to get wireless ear buds that are still slaved to a larger device. I'd still be satisfied with wired ear phones if they were available.
I also didn't imagine today's smartphone ecosystem, even though I saw all the PDAs and other pocket computing platforms and their general purpose potential. I didn't predict the social/marketing angle that was going to make any of this seem worthwhile to average consumers. Now, I benefit from these economies of scale making the tech affordable, but I barely need it. I still prefer going to a laptop or desktop for any "real" tasks. Ironically, a smartphone ir tablet makes me feel frustrated and "mute" without a keyboard.
I understood the idea of convergence and the general purpose device. I understood the mobile/always on value and was on the early side of wire-cutting to have only a mobile phone. But, I was surprised at how rapidly the smartphone ate up digital camera and personal video camera markets. I am still surprised at how much smartphones are eating into spaces like regular laptops and PCs. And now the entire concept of a phone seems to be disappearing, with traditional voice calls being less relevant as time goes on.
So I am a fence-sitter on some kind of N-dimensional fence. I admit that a lot of tech development might occur and I probably can't guess which ones get popular. Maybe there is some VR/AR angle that will finally catch on.
But on the more general topic, I also think that all these developments above had many other passionate developers and ventures working on slightly different angles that failed in the market. As a third-party observer of decades of tech, I do also think it's a lot of luck. It is survivorship bias to fixate on how NVIDIA or Apple or anybody had some perfect strategy to ride these big waves.
You can have a fan that pull the air in or out of the inside of the glasses without any light ever coming in.
That seems easy to fix - small tubes/vents that don’t go straight through but hit a corner or big enough curve. Air can go around obstacle but light won’t.
That’s the theme of this article. Nvidia invested heavily in GPUs even though the complete business model wasn’t always clear.
Yes, they've shipped tens of millions of headsets across the world.
> It’s hard to figure out what Meta spent all that money on.
R&D on every problem for XR. Their software side is mediocre, but what Reality Labs is working on for hardware is unparalleled by any other company, even Apple.
In the case of AR/VR, I think there are two issues that make payoffs uncertain: 1) the hardware would have to get much cheaper; when will that be? 2) What is the killer app that would motivate the (large) expeditures required to produce compelling virtual environments? Games is an obvious case, but games are a very different beast than online meeting rooms -- the whole action/reward cycle of games isn't likely to work for business meetings. It isn't really clear that VR adds enough value to be compelling for non-gaming apps.
But it could easily be true that we just haven't imagined the right use case. You won't find it if you don't look.
Still I think people REALLY underestimate this space. First gaming has an outrageous TAM and VR games are GOOD. Second AR has an even bigger market, and I’ve realized that it’s not just about giving us all a HUD to walk around with, it’s also going to about having an AI assistant get information in the environment around us with a camera.
I just think it's an honorable mention because had things gone a little differently, perhaps Intel could have been king of the hill instead.
They lost the process lead and can’t figure it out for almost a decade now - we wouldn’t be having this discussion if their hardware worked out according to their plans.
One thing I never see people discuss in the context is that from the beginning NVidia drivers were closed source while Intel drivers were open source. If you don't see software as a competitive advantage that obviously limits the resources you can invest in it. Linus Torvalds said "Fuck you, NVidia" but who's fucked now?
Never once read a comment attributing Nvidia success to luck.
The only luck Nvidia has is the luck that AMD fell asleep at the wheel and couldn't get bothered to put more than 2 engineers on a CUDA competitor even when it was getting apparent that AI was worth billions (i.e. ~4-5 years ago).
“No, Nvidia just found themselves in a lucky situation.”
You have to keep in mind that for a long time concurrency was thought of as the only way to keep getting increasing performance out of CPU's, what they invested in is massive concurrency and when ML/AI hit they were well-positioned for it due to their investments in massive concurrency.
The vast majority of the comments on this article prior to the one you were replying to were attributing it solely/largely to luck.
When people say it's luck, I think they are reacting to the reality that Nvidia couldn't know, when they were doing this investing, that there was a big AI market waiting to take off. They were doing good work, but they were also very, very lucky that circumstances granted them this opportunity. There is no shame in that -- few companies achieve great success without some opportunity manifesting.
But it's a mistake to pat yourself on the back too hard, either. Without the opportunity, they'd still be making gpus with some other applications.
* This is gated by the ability to recognize those opportunities when they appear, willingness to act decisively to maximize the probability of positive outcomes and the preparedness to exploit such advantages. This tends to require mental preparedness, emotional maturity and a willingness to invest scarce resources and/or time - in advance - toward maintaining situational awareness and some excess reserve resources. Doing this is hard but these traits are learnable.
* Similarly, a portion of available conscious effort and scarce resources must be continuously expended toward being resilient to bad luck when it inevitably strikes. The net impact of misfortune can vary substantially depending on mitigation steps taken in advance. This requires accurate awareness of ambient risk factors and careful balancing of where you choose to place your limited 'air bags' and 'ounces of prevention.'
Most of these things are at least somewhat within your ability to influence, with the exception of initial conditions. At the "opening deal" of life some people are dealt better cards and some people are dealt worse cards. This is not fair, but it is what it is. The silver-lining is that, after the initial cards are dealt, it can still be a long game with many rounds. How you choose to play the cards you have in each of those rounds can lead to substantially different outcomes. Because it's a game like poker with randomness, hidden variables, subtle cues and second-order probabilities - it's easy to conclude it's almost all luck. This is unfortunate because not understanding the 'meta' of the game, or even knowing there is a meta, does make it mostly luck for some.
I think NVidia's 'good fortune' is the cumulative result of playing the meta-game effectively for a long-time and thus leading to them having the capability to maximize their outcomes when eventually finding themselves in a high-opportunity environment (aka "lucky").
They had their reasons for doing what they did and I'm sure they eventually realized they were well positioned for ML/AI, but there's no way they planned that out before ML/AI was a viable thing.
And sure, AlexNet was good but remember this is the maxwell days, tensor cores aren't even a thing yet, it was at minimum a very bold bet on the basis of "some image classifier model thing". Nobody else saw it as more than an academic toy (obviously, or they'd have jumped in too).
https://www.newyorker.com/magazine/2023/12/04/how-jensen-hua...
> Within a couple of years, every entrant in the ImageNet competition was using a neural network. By the mid-twenty-tens, neural networks trained on G.P.U.s were identifying images with ninety-six-per-cent accuracy, surpassing humans. Huang’s ten-year crusade to democratize supercomputing had succeeded. “The fact that they can solve computer vision, which is completely unstructured, leads to the question ‘What else can you teach it?’ ” Huang said to me.
> The answer seemed to be: everything. Huang concluded that neural networks would revolutionize society, and that he could use CUDA to corner the market on the necessary hardware. He announced that he was once again betting the company. “He sent out an e-mail on Friday evening saying everything is going to deep learning, and that we were no longer a graphics company,” Greg Estes, a vice-president at Nvidia, told me. “By Monday morning, we were an A.I. company. Literally, it was that fast.”
10 years out is 2014.
Use luck to be humble, use luck to empathize, use luck to build people up, but don't use luck to tear people down.
I remember talking to Jason Fried about this with 37 signals. A lot of people didn’t realize the grind that team had before things finally clicked.
It’s a common pattern to see people gloss over the insane lengths and foresite required to achieve overnight success. :)
At least judging from the upvote it does seems I have the backing from silent majority. All is not lost.
The wise decision was to keep pushing cuda and not dropping the ball on it to cut costs in the short run, and realizing the potential of using it for scientific computation early on. Then one thing led to another. When AI came around, cuda was the only mature and serious framework for the job.
They were already building GPUs mainly for gaming. Crypto then came along and swiped up a bunch of GPUs. When the self mining craze somewhat waned the AI craze started as the boundaries on its ethicality were broken down in an uncertain economic environment where corporations started racing to see who will get to the top.
The crypto craze waned just as LLMs were picking up. Any more delay with LLMs, and perhaps nvidia would have been overextended with no demand to take their inventory. There may in fact , be no nvidia in that future depending on the level of the bet.
Or take it 1 step further . if COVID hadnt happened, the crypto craze would never have occurred, and nvidia would have only been a bit-level player in the LLM craze we are now in.
I still remember nvidia ads in pc game magazines. That and 3dfx. Who knew?
Why? They were the only ones taking GPGPU seriously? AMD completely ignored it for years and Intel only became serious about GPUs very recently.
NFTs at 100K , etc.
Thats the "craze"
Before that it was still a speculative play, but hardly pervasive and available to so many.
Nvidia has been doing the hard work in preparing to succeed in this market. CUDA has been meticulously developed and maintained, creating an adhesion to their hardware that would not otherwise exist in the AI market.
It also has been willing and capable of creating lines of business hardware aimed at maximizing utility for their customers.
They also have hired and maintained a roster of the best engineers in their specialties, including the software part of the equation.
There is no part of their success that they weren't prepared to take advantage of when the opportunity presented itself. They didn't control the size of the opportunity itself, but no greatly successful company does.
Tell that to ATI/AMD, and all the now-defunct and also-ran (Matrox, 3dfx) graphics companies.
Matrox is still alive.
Iirc they found a niche in medical device imaging.
What if… you read my comment more closely, and saw it was a parenthetical example after "also-ran" and not "now-defunct".
this underplays the AMD board refusing a merger with NVIDIA (because jensen wanted to be CEO of the resulting company and hector ruiz didn't like that), and then overpaying for ATI and leaving themselves depleted of funds.
(staff writer, not blog "contributor") https://www.forbes.com/sites/briancaulfield/2012/02/22/amd-t...
Bulldozer/etc could have gone very differently if they hadn't been broke and forced to underinvest. They might not have ever been in the position of "consoles having to carry them" if they had simply done a merger-of-equals with NVIDIA, or if they had simply walked away from the graphics market afterwards instead of rushing to buy ATI.
A lot of people just can't bring themselves to admit that good decisions had a large role in putting NVIDIA where they are today, and that bad decisions had a large role in putting AMD where they were in the late 2000s/early 2010s. It wasn't just bad things happening to AMD - the bad things happened because of bad decisions, they were consequences of actions.
People overfixate on the Intel thing as being the root of all of AMD's woes, as opposed to the other 75% of the problem that they could control. Like hector ruiz was just an utter disaster all around, not just even this one thing. And even today, AMD does discounts/rebates for bulk purchases. And the best rebates come if you buy all of your hardware from them (and thus, none from your competitor) - same as Intel did, which is why the decision was eventually significantly reduced. The eventual finding was - like many things, it may or may not be anticompetitive, depending on how you use it - but volume rebates and even outright exclusivity agreements are not inherently illegal in the way AMD fans generally imply they are. Ask Pepsi.
Matrox and 3dfx were gone before they could even conceive of a general purpose GPU strategy.
[1]: The Matrox Parhelia itself being out of date tech on release to boot, lacking a Direct3D 9.0 class architecture that ATI launched two months later.
You could argue they successfully pivoted out of consumer and into b2b.
But they gave up on making their own cores.
Hence "also-rans".
What are you trying to argue here?
Case in point: Nvidia has been hosting "GPU Conferences" annually to build awareness and drive adoption of CUDA. These events surely aren't free to host but necessary to build momentum and give an edge to your custom stack.
Luck is more applicable to the cryptomining boom and bust cycles that Nvidia also profited from. Their gaming GPUs (along with AMD's) just happened to be the best available at the proof of work.
The point isn't that they didn't invest in the direction that ended up being right, it's that they didn't do it specifically with ML/AI in mind years before it was even a twinkle in google's eyes.
Well yes and no. They were certainly lucky to be at right place in the right time. But they were also consistently investing into CUDA and the AI/ML ecosystem while their competitor(s) ignored it to such an extent that NVDIA became the only real option (and deservedly so).
This is why they can effectively behave like a monopoly these days and just almost inconceivably high margins.
IIRC were on CUDA 5 or something when imagenet came out and changed the world.
They might not have imagined LLMs when they decided to invest in making their GPUs programmable but I guarantee you they extrapolated the future compute potential of vector programmable machines and decided it was not a huge risk to enable it as it is simply betting that some important application would be around to tap into it.
It was hugely experimental but does demonstrate that the idea of a GPUs being used for compute isn’t a recent one.
What people find contentious is this claim that they specifically did it for ML/AI. That's not why they invested the way they did.
https://www.acquired.fm/episodes/nvidia-the-gpu-company-1993...
https://www.acquired.fm/episodes/nvidia-the-machine-learning...
https://www.acquired.fm/episodes/nvidia-the-dawn-of-the-ai-e...
>I'm a great believer in luck. The harder I work, the more of it I seem to have.
Nvidia has been working hard(er than their competition) on the software side for almost 2 decades to be in the position they find themselves today. 16 years ago, they released CUDA for general-purpose computing on GPUs, and then 9 years ago they followed that up with cuDNN. They have a consistent pattern of making a intentional, long-term bets to diversify their market exposure and unlock new product areas while building a software ecosystem moat.
Yes, they obviously got super lucky with the cryptocurrency frenzy, but there's a reason all the miners were mostly buying Nvidia cards instead of AMD cards.
>While it is true that Nvidia cards are generally preferred by miners due to better price-to-performance, AMD GPUs such as the Radeon RX 6600 XT could still be mined on profitably until recently. […] So yes, carefully consider the condition of all used graphics cards—Nvidia or AMD.
[0] https://www.pcworld.com/article/395149/should-you-buy-a-used...
Think you're just looking at this from the angle of a gamer and not someone who's been paying attention to GPGPU compute earlier than the past 6 months.
Then this luck should have equally found AMD, who even today are struggling to pick up the ball they've been dropping for a decade now. My last PC had a Radeon, and I waited the life time of that PC assuming AMD support was just around the corner, all the while renting Nvidia cards in the cloud for any serious projects.
I've been in the ML space long enough to remember when people were just speculating about doing ML/computation on GPUs. Nvidia made that much easier and has continued to improve support and features for the past decade+ There insane success is certainly part luck, but I wouldn't be so quick to dismiss all of it as merely happenstance.
https://www.youtube.com/watch?v=WLq9zv3k5n0
The race was on, but nobody else was running.
Because they were broke and didn't have the resources to invest properly even if they wanted to.
That's not quite true -- they bet the farm on Zen, and that bet paid off. Which means that now they have the resources to also invest in AI. I'm fairly sure if they had bet the farm on AI instead of Zen or if they had tried splitting that bet they'd be bankrupt now.
Copy-pasting a comment from a discussion a little while[1] ago: CUDA was first released in 2007:
* https://en.wikipedia.org/wiki/CUDA
* https://developer.download.nvidia.com/compute/cuda/1.0/NVIDI...
Two years before the Bitcoin paper (2009):
* https://en.wikipedia.org/wiki/Bitcoin
They had a presentation called "The Era of the Personal Supercomputing" at SIGGRAPH 2007:
* https://dl.acm.org/doi/10.1145/1281500.1281647
* https://www.nvidia.com/content/events/siggraph_2007/supercom...
Ian Buck (co-?)creator of CUDA speaking in 2008:
> Ian Buck talks about his background developing Brook for GPUs at Stanford university and what paths were taken for developing a C platform for GPUs.
* https://www.youtube.com/watch?v=Cmh1EHXjJsk
> In 2003, a team of researchers led by Ian Buck unveiled Brook, the first widely adopted programming model to extend C with data-parallel constructs. Ian Buck later joined NVIDIA and led the launch of CUDA in 2006, the world's first solution for general-computing on GPUs.
* https://developer.nvidia.com/cuda-zone
* http://graphics.stanford.edu/~ianbuck/
Nvidia purposefully went after parallel computing. Specific applications (cryptocurrency, ML/AI) appeared later.
[1] https://news.ycombinator.com/item?id=38446957#unv_38447944
It is true that they got lucky several times in a big way. But CUDA was an expensive R&D for many years without clear payouts.
It didn't wane, it was decimated when ETH switched to PoS, and off of GPU mining entirely.
All of the other mined coins dropped in value as miners moved to them and dumped all their rewards, making mining those coins unprofitable as well.
It was ETH that was propping up the entire GPU mining ecosystem.
In its early days, NVidia focused on its delivery speed by having a unified software driver/integration strategy. The TNT video card was meh, the TNT2 a little better but inferior to 3Dfx... by the time they launched GeForce, 3Dfx didn't have a product ready to compete. Their driver/integration/product test strategy made that speed possible.
With that background you can understand why CUDA wasn't lucky, they repeated the same approach: combine their hardware with the software.
[1] https://www.jstor.org/stable/26861060
[2] https://en.wikipedia.org/wiki/Fifth_Generation_Computer_Syst...
Nvidia wisely recognized that only so much horsepower could be used by a conventional 3d graphics pipeline with a given screen resolution, and they needed to invest in growing future compute-heavy adjacent markets.
They invested in generalizing their GPU into a more flexible vector coprocessor for HPC and then adjacent markets. They convinced fundamental engineers and researchers in this area to come work for them.
There was deep fundamental work done by Ian Buck in 2004 on leveraging GPUs as general vector processors ( https://graphics.stanford.edu/papers/brookgpu/ ) and that leadership and deep thinking went to Nvidia, not to Intel. Intel did not have the passion from the top to care about this. They couldn't even care enough to field competitive 3D chips (and associated software), much less extend their thinking to generalize beyond it skillfully. Nvidia did.
Anyone who spent every day thinking about how to grow the vector coprocessor market would have pursued crypto and AI when they came along but Nvidia's strong engineering and profits from a leading 3D position gave them competitive advantages which they are, for now, reaping.
It felt like they weren't leaning into crypto, which surprised me. Instead it looked like they were trying to maintain gamer goodwill by not increasing consumer card costs during the boom. Of course scarcity raised secondary market princes, but Nvidia kept MSRP lower than the boom dictated.
It seemed like they were betting against crypto during the craze. And sticking to their strategy on the consumer side. So maybe that's how they stuck to a ML strategy too.
So they had all this CUDA stuff, which they must have invested in heavily because AMD showed what happens when you don't. That led to a software ecosystem for ML.
Maybe it was all luck, but a strategic choice explains some of this in hindsight.
CUDA was already digital gold in 2018. ML had moved to GPUs several years prior, and CUDA was a primary enabler of that transition.
Of course, the AI thing exploded afterwards this time so the demand dip didn't happen.
They've been selling mid-sized GPUs like the GTX 1080 at 800$ 8 years ago. A 300 mm2 GPU.
They have been rising and rising costs for a long time to milk as many dollars as they could.
People seem to forget even before the crypto craze the company had insane margins in their revenue. They weren't selling 800$ chips because those costed 600 or 500$ to produce...
Even the most expensive 4090 is hardly more than a 400$ chip, memories included to build.
Woah, only one more year and this reference will be old enough to vote.
Pushing the limits of existing hardware, it looked truly spectacular when it came out - and many would say it still does.
Crysis came out to huge fanfare due to it's impressive graphics. However even the highest end spec'd PC's had trouble running the game at a stable frame rate at high settings--iirc because the game's engine (cryengine?) was not well optimized.
Crysis performance became a benchmark for PC's, and then the joke phrase "But can it run Crysis" started, used seriously when discussing PC hardward specs, and jokingly used for unrelated hardware "This printer has 24bit color?" "But can it run Crysis?"
Then they put two and two together and started investing heavily as they saw momentum build.
There's been a decade or more of deep learning models breaking records in almost every single research field, powered (indirectly) through CUDA, cuDNN and other NVIDIA software.
AI didn't "come along" when OpenAI released ChatGPT. DNNs that have been 99% NVIDIA-focused have been beating the state-of-the-art for years and years.
Also for the record the ADA architecture (very dominant AI accelerator) was released when the stock price was like $100 (compared to the $500 now).
The chronology is this:
2007: CUDA released
2012: AlexNet (which runs on CUDA) turns heads and everyone (in the research world, not the public) starts getting really excited about AI
2012-2021: NVidia invests in CUDA while getting carried by gaming and crypto
2022-2023: AI explodes into the public consciousness and NVidia valuation balloons to absurd levels
Now, I don't think anyone can argue that NVidia was not smart to build CUDA. What no one would have predicted was how important it would turn out to be back when they started. It was more like "hey, these gaming devices are pretty good at matrix/vector math, let's build an API for people who want to do that" and then it turning out that AI is all matrix/vector math.
The point is that all of NVidia's investment in CUDA and AI was subsidized by the gaming and crypto cash cows. They didn't have to "bet the farm", they just had to keep it going as a side project and increase their investment as they saw the momentum build.
Yes, and it takes lot of effort to be in the right place.
CUDA has been heavily utilized for AI for many many years now, whole reason Nvidia is so entrenched in it is because they were they only ones taking GPGPU seriously like OpenCL (1) rose and was abandoned before we even get to your interpretation of the timeline.
(1): Easy to forget now that AMD and Apple had an common standard competitor to CUDA and completely fumbled it.
Once gen AI gets good enough to generate high quality video games, virtual worlds, etc. in real time, that will redefine gaming and entertainment.
Why wait for the next Mission Impossible movie to come out when you can experience it...as Tom Cruise...with your own storyline with all of your friends. And get a new one every day.
The curation of an experience is undervalued by many.
Just like coding isn't about mechanics. It is about understanding requirements and coming up with solutions that meet needs
Think about any good story. A million details could be different, and it would still work well.
The current expectation that movie studios or the Hollywood machine will somehow be diminished by GenAI is misguided. If anything, it will normalize $100 movie rental "experiences" you can enjoy at home with no additional hardware besides your smartphone and a cast-enabled TV, with everything rendering real-time in the cloud.
We were 2nd in the state in Illinois, and I wanted to cement our place and possibly win by selling everything to lock in gains in the final week. The person executing the trade on our team accidentally shorted it and it went up a considerable amount in the last week knocking us down quite a bit.
At that point I seriously considered dumping my savings into Nvidia. I'd be retired right now if I had done so.
Nvidia CEO: We bet the farm on AI and no one knew it https://news.ycombinator.com/item?id=37055375 (August 8, 2023 — 10 points, 4 comments)
NVIDIA will be okay though, they have the volume to get the newest nodes first. Only a few can compete with them.
I'm not an ML developer, so I'm curious. Just economies of scale?
Outside of training, using the trained network for inference seem to be changing less, so maybe a decent target for bespoke targeted hardware. And indeed it seems that custom hardware blocks are very much "catching favor" in that market - everyone seems to be adding accelerators to their SoCs.
GPUs seem to be at a happy medium, where they're often flexible enough to run techniques while still having better efficiency than a CPU.
Is this acknowledging the game market is dry?
Ray Tracing was the next frontier of visual quality. The compute was at a level where something like this was possible.
DLSS was not a "huge gambit". It was an application of AI models (many of which train and run on NVidia chips already) to the gaming space.
Making it sound like it was a "genius move" is great for promoting NVidia as being an innovator.
Personally, I didn’t think it would, but the power of a small number in the top of your screen beats facts for most people. And here we are I guess.
But there’s also the problem that AMD drivers are STILL shit. This is 20 straight years of poor drivers. Something tells me that if AMD would just get their software act together, the DLSS train may not have taken off the way it did. You do see a lot of comments that the drivers haven’t been bad for a while, but anecdotally, they definitely are.
I'm sorry but what does this even mean?
The process is not perfect and can introduce various visual artifacts, which are especially visible at lower monitor/output resolutions and the artifacts are more likely to occur when the render resolution is also much lower than the output resolution. These two also feed into each other as in general cheaper/weaker hardware (where the lower render resolutions would be used) tends to be paired with lower resolution monitors (where the artifacts would be more visible). I guess the "shitty, squashed, artifact laden" comes from there.
DLSS 3 doubles the framerate by interpolating frames in the GPU (so the game doesn't "know" about them and player input isn't handled), hence "fake frames".
The frames you’re getting are not based on gamestate, so they’re just fake placebos. This is why it’s curious to me that gamers by and large have completely accepted DLSS to the point that they’re actually completely okay with losing rasterization perf as long as a numbers in the corner of their screen is sufficiently faked.
Even in games supporting DLSS 3, there is to best of my knowledge and experience almost always a separate configuration option to turn frame generation on or off, allowing you to use DLSS solely to render non-"placebo" frames and upscale the output.
This absolutely allows you to run at a lower internal resolution, get higher "real" FPS and enjoy a reasonable upscaled output image. Sure it won't be as dramatic a difference as with frame generation, but it doesn't have the drawbacks of non-gamestate frames either. I've found DLSS for upscaling (without frame gen) fantastic frankly for getting older Nvidia parts to run new games for budget conscious gamers who don't have hundreds or thousands of dollars to drop on a new 30xx/40xx part.
Nvidia did slap “Hurr durr AI driven” buzzwords in front of theirs, but supersampling exists super far back. As far as gaming goes, supersampling was probably popularized in super old emulators, particularly zsnes.
Way more crashes. Way more weird bullshit (like audio disappearing????) that just doesn’t happen when you run an Nvidia card.
“AMD drivers are fine now” has lost all meaning. I will continue to give them a chance now and again, but it’s based on keeping competitive up, not on the drivers being good now.
They've just pivoted to AI semi recently (considering their long history) using the same hardware as always, for graphics/scientific/AI workloads. Sure they tweak vram amount/bandwidth/support additional precisions but it's all really same old same old. Just with more transistors every 2 years.
As a counter, look at MS, they have been around for more than 10 years and there is plenty of luck, hubris, good decisions, bad decisions that did not matter, because they effectively held a dominant position.
Being dominant equals to luck.
Once you are dominant though, you can stay that way for a long time.
If you’re interested in the early gpgpu work, which interestingly enough worked both on Ati( now amd ) and nvidia hardware.
I agree to disagree :-)
So from their perspective, it wouldn’t have happened at all. But I understand my comment was ambiguous.
Here is Jensen talking about building the industry 9 months before the founding of OpenAI: https://www.youtube.com/watch?v=_iBLoNG0qHk
Gpgpu dates back to the early 2000s, and it happens that deep learning as a computational problem is very well suited to GPUs.People figured this out around 2012-2013, and just like with cuda a few years before, nvidia figured out there was a huge market for them, because the underlying compute problem was well suited to their hardware.
Ai could have used different models which are not computationally friendly to gpus ( or may not even require a lot of computational power ) and nvidia wouldn’t have been so successful at ai, much like intel’s cpus are not well suited to current deep learning models.
I remember discussing with friends in the hpc community what was going to happen to nvidia - we were wondering how they could grow their compute business. And bam, massive compute needs for ai, problem solved.
"Could have not happened" means without hindsight, you could see the possibility that it might not have played out.
nVidia didn't know for a fact that deep learning architectures would have been so successful in the past couple years. GP's point is that maybe deep learning would have worked 80 years later, and not have the great breakthroughs around 2016~now.
The video from 2015 proves nothing. nVidia was deeply invested into this bet by then, and the mere fact that the video exist is evidence that they already betted heavily on AI by 2015 -- before the big results we know of today came about.