I'm with Linus Torvalds on this—f*** NVidia.
I'm with Linus Torvalds on this—f*** NVidia.
But the only reason you're mad at Nvidia is because they made something desirable (a GPU that does NN training) but didn't do it in the way you prefer (NVLink, Linux compatibility etc).
It's not their fault AMD and others have been asleep at the wheel and handed them the whole market for free. Any company with a monopoly on the hot new thing would try to capitalize on it.
When the other GPU makers eventually catch a wakeup, the situation will finally improve.
https://www.cnbc.com/2023/12/06/meta-and-microsoft-to-buy-am...
https://www.tomshardware.com/news/amd-scores-two-big-wins-or...
https://www.tomshardware.com/tech-industry/artificial-intell...
“Oracle is set to use…”
“Elon Musk implies…”
Not one of them have actually bought anything. AMD is just muddying the waters by grouping HPC sales with AI. It’s nonsense.
Edit: I see you’ve got a startup on MI300x hardware. Now it makes sense, I guess - you need to believe. Good luck!
AMD isn't shipping in large volume yet, but MI300x instances can be spun up in preview on Azure today with availability ramping up. Microsoft has absolutely bought a lot of MI300x.
Or maybe I know things you don't. =)
You really should disclose that you’re a founder when you mention your company like that
Don't die on this hill, just do the decent thing and add a disclosure. This doesn't show you in a good light.
And yes, yours was an instance of self-marketing, even if unintentional, by name-dropping your company alongside power players like Meta and Tesla. You put yourself on people’s radar as a GPU-intensive company.
"Please don't use HN primarily for promotion. It's ok to post your own stuff part of the time, but the primary use of the site should be for curiosity."
What I posted, just the name of my company, was entirely within the bounds of common courtesy.
Of course, I do want to put myself on people's radar... I am building a GPU-intensive company and I'm making a small relevant comment on a thread.
I looked it up. Over the years, there has been endless bike shed discussion on this topic, such as:
https://news.ycombinator.com/item?id=24353959
https://news.ycombinator.com/item?id=36404027
I think that it is smart of the moderators here to simply sit on the sidelines and not try to control the discussion of this too much.
the trajectory looks good if they can just resist giving up early without seeing instant success
Sounds like I've been doing life wrong.
Intel may end up back on top.
https://www.asml.com/en/news/press-releases/2022/intel-and-a...
https://www.datacenterdynamics.com/en/news/intel-receives-it...
https://www.tomshardware.com/tech-industry/manufacturing/tsm...
EDIT: bro, you're the one that said corporations are mobile, not me. secondly, all of my points still apply to a satellite office. where the headquarters is or where it's incorporated is irrelevant. thirdly, if they still do significant business in the US then the US maintains massive leverage over them anyway. so, try again.
>EDIT: bro, you're the one that said corporations are mobile, not me
I say it and you misunderstand it. Mobile as they can open an office in any country, not move the HQ alltogethere.
It doesn't have to be a race to the bottom
The bait and switch.
So what? "Any company with a monopoly" would also pay children in company scrip to work 20 hours a day in unsafe factories unless compelled to do otherwise.
Developers valued their own development experience, features, and performance over value, efficiency, portability, openness, or freedom. The result is anyone depending on CUDA became vendor locked to Nvidia. Similar as the story with Direct3D, though thankfully there are workable solutions for D3D now implemented on top of Vulkan. Nvidia's CUDA moat may be less fordable.
It's a reimplementation of cuda on top of rocm, and it's a drop in replacement you can use with already compiled binaries. It's even faster than native rocm/hip in blender
Intel is weak, and it’s much smarter to put resources into squeezing Intel while they are on the ropes than shift too much focus to ML.
Things like Bumpgate as a major social media mindshare thing largely stem from him and his coverage. And he leaves off the part where AMD gpus were failing too (because of the same RoHS-compliant solder) and where apple kinda wanted to part ways anyway because they saw the “platform” nvidia was building and didn’t want any part of it on macOS… game knows game.
https://blog.greggant.com/posts/2021/10/13/apple-vs-nvidia-w...
That’s the problem is like the evga thing there is a distinctly contemporary take and that’s shaped by people like Charlie, and then there’s the take with the benefit of a couple years of hindsight and maybe it wasn’t exactly the way the bunnyman said it was. Maybe apple didn’t want CUDA usurping them, and maybe apple wanted to leave the way clear for the pivot to apple silicon and metal and the platform Apple wanted to build there.
The problem is the AMD fanboys are just as noxious today and - just like the evga thing - it’s gonna take years before people are willing to re-evaluate things fairly.
https://www.reddit.com/r/nvidia/comments/xgn7do/we_need_more...
https://youtu.be/vyQxNN9EF3w?t=5044
There’s a little backstory like this behind almost all of the AMD fanboy “lore” around nvidia. Crysis 2 was never a thing either - the whole point of wireframe mode is seeing the full geometry at maximum LOD, with no culling, for example, and it does not happen during actual gameplay. nvidia never “sold directly to miners”, tech media didn't actually read the RBS article they were citing and it doesn't say what they say it says. NVIDIA "recently said they're completely an AI company now" stories a few months ago was a quote from 2015. Etc. There's just a group of people who just get off on hating and they don't mind making shit up if that's what it takes.
It's essentially impossible to unwind the popular opinion of NVIDIA from this ridiculous amount of past-hate... yeah, after 30 years of blood libel from ATI fanatics they probably do have an overall negative public opinion! People know the stories are true, because NVIDIA is bad, because there's all these stories, so they must be true.
They were in the same ballpark until the past coupld of quarters. Nvidia laid the groundwork for this ages ago by developing drivers that play nice with NN.
AMD could easily have done the same at the time, and can do so now.
Even if it costs $1b-$10b to do it, AMD can afford it and it’s worth doing. They’ve got $55b in equity, and a huge $280b market cap to issue more stock into. [https:valustox.com/AMD]
They're releasing APUs with tensor cores right now.
They're finally getting their dogshit ROCm drivers fixed right now.
Yes, they're late. But they're definitely playing. This whole "asleep at the wheel" thing may have been true a year ago, but it's not true now.
[edit] Apologies, responded to the wrong comment, but I'll leave it here anyway.
AMD has a looooong history of seemingly deliberately shooting themselves in the foot every time they get ahead.
This is true. It's like blaming the company for being successful. "Stop making products that are so much better than any other product".
I'm not sure we want that?
If Android were superior, no one would care what Apple did. Everyone would simply use Android.
I think the conversation around side loading and store royalties on iOS has more to do with who owns the hardware, and what the user is permitted to do by the manufacturer.
Big-cap tech is utterly notorious for some of the worst anti-trust violations and other cartel-style behavior of any sector. This goes back as least as far as “Wintel” in the 90s and probably further that I didn’t watch up close. Suffice it to say that the Justice Department is extremely disincentivized to go after domestic economic Cinderella stories in a globally competitive world and has had to bring lawsuit after lawsuit on everything from bundling to threatening OEMs to flagrant wage fixing in print (I do have second-hand that the mass layoffs are coordinated aka Don’t Poach 2.0).
Crippling “gaming” (high margin but not famine-price gouging margin) cards, controlling supply tightly enough to prevent the market clearing at MSRP routinely, stepping at least up to and likely over the line on the GPLv2: these things or things like them have been ruled illegal before (though it’s hard to imagine that happening again).
It’s possible that tech is just a natural monopoly and we had it right with Ma Bell: innovation was high, customer prices were stable and within the means of most everyone, and investors got a solid low-beta return.
It’s easy to view the past with rose-colored glasses and the Bell Era wasn’t perfect, but IMHO this status quo is worse.
A market failure is a market failure.
There’s a big lobby on HN who want to defend or minimize or justify leaving market failures be, which is weird given they’re really bad for most people on HN, it’s a free country.
But let’s call it what it is.
They made a good product. The competition is responding too slowly. And they got lucky with this crazy (in a good way) AI boom. I don't see how this could possibly be outlawed or why you would even want to. It will resolve itself on its own given enough time.
The government as a big costumers could demand these things, even without having a low. The military should demand these things to be open for various reasons.
I think Nvida can still make plenty of money in such a world.
In cases where an interface has absurdly high value for society if its a standard, the government could also 'buy' that and open it up. Just like they do with other infrastructure.
One could make the argument that the x86 interface should be public domain as it amounts to infrastructure that most of society builds on. How such a thing would exactly work is of course up for some debate. But the concept of the government 'liberating' common interfaces makes sense from a society perspective.
How would that have stopped Nvidia from dominating the AI market?
>The government as a big costumers could demand these things, even without having a low. The military should demand these things to be open for various reasons.
I agree that governments should use APIs with multiple competing implementations (or one truly open source implementation) where possible. This could make a difference in some cases, but I doubt it would have had a big impact in this particular case as demand for GPGPU is overwhelmingly coming from the private sector.
>In cases where an interface has absurdly high value for society if its a standard, the government could also 'buy' that and open it up. Just like they do with other infrastructure.
Agreed, but is there a legal reason why AMD and others are not allowed to create a clean room implementation of CUDA? Haven't they done exactly that with ZLUDA (which they have now defunded)?
I thought not supporting CUDA was more like a failed strategic move by competitors to prevent CUDA from becoming an industry standard.
I think we agree on all the relevant principles. I just don't see how any of these principles make a big difference in this particular case.
Also, I don't see the Nvidia situation as particularly problematic. They are not too entrenched to unseat. Some of their biggest customers (themselves huge oligopolists) are shaping up to be their biggest competitors.
Most of AI processing will be inference, not training. Bringing the costs down is absolutely key for broad AI use. My bet is that the hardware margins will end up being slim.
The likely height of big-cap tech regulation was the regulated monopoly of ATT/Bell Labs/Western Digital all through the 20th century.
I prefer the outcomes in that era to the present status quo.
You can agree or disagree, and disagreeing is simple: “I prefer these outcomes because…”. I welcome such disagreement. And FWIW I’m not one of the many people who downvoted you: but they were justified in doing so because you moved a comparison of policies and their outcomes into a more abstract space of 1-bit generalizations, e.g. “talking points” vs “law/precedent”. I gave examples, it’s public record and trivially Googleable.
You didn’t fail to understand my point, you didn’t like it and took a cheap shot.
Let me (start to) Google that for you:
https://www.sec.gov/enforce/sec-enforcement-actions-fcpa-cas...
https://www.justice.gov/atr/antitrust-case-filings-alpha
https://www.csoonline.com/article/567531/the-biggest-data-br...
https://www.cnbc.com/2023/01/24/doj-files-second-antitrust-l...
https://www.channele2e.com/news/big-tech-antitrust-regulator...
https://www.nytimes.com/2019/07/23/technology/justice-depart...
... scads more law stuff and a decade ...
https://en.wikipedia.org/wiki/High-Tech_Employee_Antitrust_L...
... scads more law stuff and a decade ...
https://en.wikipedia.org/wiki/United_States_v._Microsoft_Cor....
And I never mentioned 'not citing concrete examples.', if your confused about the ask, I'll repost it: "can you write a credible argument based on known case-law/precedent/etc... "
i.e. Can you write down the actual argument that you propose to advance?
The de-facto mechanism that's occurring with market segmentation is that consumers pay less than the clearing price and enterprises pay more. And consumers benefit from this subsidy.
If you banned it, you also wouldn't change the cost of developing new generations of products, so product development just would go slower, and the market would get even lumpier and less efficient (vendors squatting newer nodes would have an advantage, but it's inefficient to launch something immediately to compete with them, etc).
Price discrimination is generally prohibited in naked form: it often if not usually induces market failures. So companies spare no expense working around it (there's a reason why ITA software people made a mint approximating NP-hard problems acceptably well and rather cheaply to compute airplane ticket prices).
"Market segmentation generally benefits consumers" is exactly the kind of Wien's Law first-order approximation that sounds compelling but doesn't fit the data.
It does not benefit consumers that a 4090 has the same amount of GDDR6 in it as a 3090, which is incidentally 2.6Gb shy of `dolphin-8x7b-v1-q4_k_m` (or whatever the NVIDIA equivalent is, I use Apple gear now because inference is the game now), while the price of that RAM dropped. The fact that an H100 or whatever the next Hopper iteration is costs more than an M-series BMW on paper says nothing about what Meta paid for 150k of them or whatever it was, and it certainly doesn't help someone who'd like a decent chat bot without sending their data to eyeball-scanner database guy. When you see a leading edge x090 for MSRP or less at Best Buy, it's a good idea to stock up if you're in the market over the last 5 years (London has been running ETH2 for what, a year and change now?). That hardly seems a subsidy to the consumer paid for by those spendthrifts in FAANG.
NVIDIA pioneered serious "GPGPU" engineering. 5-10 years ago their microarch and fab-sourcing and software and the whole show were legitimately differentiated, no one had done it. And they made a pile, and that pile seems "conflict-free". But anyone can make a fused multiply-add unit these days, EE undergraduates do it in Verilog. Intel makes PCI express cards that cost $300 and have higher memory bandwidth in the <hand-wave>decoder-only language attention model</hand-wave> use case. And their drivers are documented and work on everything.
Now did Lisa Su decide to "concentrate on the supercomputing market with the MI300XYZ" and Jensen decided to "concentrate on AI with Hopper" independently to a degree where the market is perfectly partitioned? Who knows, I certainly don't have proof one way or the other. But if someone made a call being like "I'm thinking of focusing on X but don't really see our differentiation in Y. How's Cathy?", it wouldn't be the fucking first time.
But even that isn't critical to the point: this, the here and now, isn't good for consumers, or society, or our industry, and unlike e.g. Stallman's melodrama, it fucking matters this time.
Levity aside, unfortunately this stuff is as real as Tuesday and taxes, and as serious as a heart attack used to be and an off-color tweet is now. This is the "Star Trek: TNG" v. "Blade Runner 2024" moment: for the first time in history the endless cycle of tennis courts and guillotines and eventually a new crop of nepotism-fueled classism awaiting its turn at the guillotine might end up in a spontaneously broken symmetry.
There are like 5-10 unanimous markers of human achievement that machines can trivially reproduce or exceed, and everyone acknowledges Chess and Go and Atari and fan-fiction and ImageNet and ImageNet fan-fiction are in that box.
But think about what an LLM actually does:
`argmax(P(next_token | {previous_tokens, corpus, joint_alignment_loss, epsilon}))`
They're human in a really scary sense: there's an old saying that if you can only be good at one thing, be good at lying, because then you're good at everything. LLMs beat pick-up artists in bars and tech CEOs in all-hands meetings all to hell at vaguely riffing with a ruthless, amoral, and narrow goal. If I ask Dolphin to talk like a VC? It says "path-dependent" every third sentence like it's got a Substack and a Tesla and a very respectable 3-bedroom in Los Gatos or Atherton.
If this stuff was floating around 20 years ago? We'd throw out the donor class without a second thought on seeing that a MacBook can talk like any power-broker but better.
Today? The capture is so far along that we might end up handing the reigns of "alignment" over to "effective altruists", at which point the situation is pretty static.
You think AMD doesn't want to carve out a chunk of that pie?
What Nvidia is doing is hard and the engineers with the skill to design those chips are few far between. I'm sure Nvidia's already hired most of the best in the business.
Anyone who wants to unseat Nvidia will need very talented people, who aren't cheap, and that means massive investment capital which largely doesn't exist.
The thing that is missing is AMD focusing their software engineers that develop the drivers, and making them put work into RoCM to make it usable across all cards, all driver versions, like with NVIDIA.
However, the real killer feature now is CUDA. Everyone's coding for CUDA, which AMD's hardware doesn't support, so even if they have a GPU that's on par with Nvidia, most libraries still can't make full use of it.
I still don't see the required effort put into place by Intel together with AMD in order to create an attractive alternative to CUDA.
Right now they're only only getting looks because their devices are cheaper for the hardware you get and for big projects because they're available, so you're required to have in-house experts who know how these platforms work in order to get stuff done that you know would definitely work on Nvidia.
The games Nvidia is playing with the consumer market is really annoying, but we have to thank Intel and AMD that Nvidia is in a position to do this. Microsoft is probably also at fault.
We're at a point where Apple only has to say "Oh, one more thing, you can now put Nvidia GPUs in your Mac Pro" for Intel and AMD to notice in what position they've put themselves into.
Half those teams are full of crap, a quarter will take the funds and try to do something like CUDA (but better) cough ROCm. Then another eighth to a quarter will simply not have the political clout to get the whole thing done.
Add to this that we just funded 10 teams to get 1-2 functional teams… and you see why chasing after an incumbent is hard. Even when you have near infinite money to do so.
Without going practically all-in on Zen, they'd be bankrupt.
It'd be better for them to sell low priced gaming cards that would perform poorly for non-gaming purposes and sell extremely high priced specialty cards to the people who want to use them for AI or crypto or whatever other non-gaming uses they come up with.
That'd at least keep the price of video cards low for gamers, avoid supply issues, and allow nvidia to extract massive amounts of profit from companies with no interest in video games. The only downside would be that it makes it harder for anyone who doesn't have deep pockets to get into the AI game.
That's the point. They're not trying to do gamers a favor, if it was that they'd just make more cards. What they're trying to do is market segmentation, which customers despise and resent.
Then why wouldn't they just make more of them? The excuse is supposed to be fab capacity, but the 3070 has better performance than the 4060 etc. and is built on the older process which should no longer be in short supply.
There is actually very little margins in the midrange consumer discrete GPU market. The market for discrete GPUs have been shrinking since the mid 2000s.[0] Most GPUs are sold as integrated such as SoCs and in consoles nowadays.In a shrinking market, the midrange and low-end products will cease to be profitable. Hence, Nvidia's 60/70 offerings are lackluster because they don't make much money from them. They want you to buy the 80s and 90s cards.
Furthermore, node advancements have stopped scaling $/transistor. So the transistors aren't getting cheaper, just smaller.
Lastly, Nvidia wants to allocate every last wafer they pre-purchased from TSMC to their server GPUs.
[0]https://d15shllkswkct0.cloudfront.net/wp-content/blogs.dir/1...
But those are mostly AMD, and doesn't really have anything to do with what features someone puts on their discrete gaming cards, except insofar as it implies gamers don't need the cards some amateur ML hobbyist might buy.
> In a shrinking market, the midrange and low-end products will cease to be profitable.
That's assuming the products have high independent development costs, but that isn't really the case. The low end products are essentially the high end products with fewer cores which use correspondingly less silicon -- which have higher yields because you don't need such a large area of perfect silicon or can sell a defective die as a slower part by disabling the defective section, making them profitable with a smaller margin per unit die area.
> Furthermore, node advancements have stopped scaling $/transistor. So the transistors aren't getting cheaper, just smaller.
Which implies that they can profitably continue producing almost-as-good GPUs on the older process node.
> Lastly, Nvidia wants to allocate every last wafer they pre-purchased from TSMC to their server GPUs.
Which is why the proposal is for them to make as many GPUs as the gamers could want at Samsung.
While midrange GPUs are a cut from highend GPUs, they're still significantly more expensive to manufacture than say CPUs, at a transistor to transistor level. Look at an AMD 7950x transistor count, and then an RTX 4060 transistor count. The GPU has ~50% more transistors but sell at half the price. In addition, the GPU requires RAM, a board, circuitry, and a heatsink fan. The margins simply aren't there for lowend GPUs anymore.
Previously, Nvidia and AMD can make it up through volume. But again, the market has gotten much smaller going from 60 million discrete GPUs per year sold to 30 million. That's half!
Based on your logic, AMD should feast on midrange and low end discrete GPU market because Nvidia does not have value products there. But AMD isn't feasting. You know why? Because there's also no profit there for AMD either.
Once you stop thinking like an angry gamer, these decisions start to make a lot of sense.
Customers who want petroleum are price sensitive. Therefore, petroleum exporting is a low margin business. This is why the Saudis make the profit margins they do. Wait, something's not right here.
> Look at an AMD 7950x transistor count, and then an RTX 4060 transistor count. The GPU has ~50% more transistors but sell at half the price.
You're comparing the high end CPU to the mid-range GPU. The AMD 8500G has more transistors than the RTX 4060 and costs less.
> In addition, the GPU requires RAM, a board, circuitry, and a heatsink fan.
The 8500G comes with a heatsink and fan. The 8GB of GDDR6 on the 4060 costs $27 but the 4060 costs $120 more. A printed circuit board doesn't cost $93.
> Previously, Nvidia and AMD can make it up through volume. But again, the market has gotten much smaller going from 60 million discrete GPUs per year sold to 30 million. That's half!
That's not because people stopped buying them, it's because they shifted production capacity to servers.
> Based on your logic, AMD should feast on midrange and low end discrete GPU market because Nvidia does not have value products there. But AMD isn't feasting. You know why? Because there's also no profit there for AMD either.
But they do though. You can find lower end AMD GPUs from the last two years for $125 (e.g. RX 6400) whereas the cheapest RTX 3000 or 4000 series is around twice that.
And anyway who is talking about the bottom end? The question is why they don't produce more of e.g. the RTX 3070, which is on the old Samsung 8LPP process, has fewer transistors than the RTX 4060, is faster, and is still selling for a higher price.
And anyway who is talking about the bottom end? The question is why they don't produce more of e.g. the RTX 3070, which is on the old Samsung 8LPP process, has fewer transistors than the RTX 4060, is faster, and is still selling for a higher price.
What do you think? I gave you my reasons. Why don't you take a crack at your own question? There has to be a logical business reason right? Customers who want petroleum are price sensitive. Therefore, petroleum exporting is a low margin business. This is why the Saudis make the profit margins they do. Wait, something's not right here.
One is a commodity. The other is about as high tech as it gets. Completely different economic rules that govern these products. You're comparing the high end CPU to the mid-range GPU. The AMD 8500G has more transistors than the RTX 4060 and costs less. The 8500G comes with a heatsink and fan. The 8GB of GDDR6 on the 4060 costs $27 but the 4060 it costs $120 more. A printed circuit board doesn't cost $93.
One has an entire board that needs soldering, assembled by a manufacturing line, tested with many parts, and a team of dedicated engineers optimizing drivers constantly. The other is a CPU that is machine tested, and shipped with a heatsink fan unattached. Come on now. That's not because people stopped buying them, it's because they shifted production capacity to servers.
That's not true. The discrete GPU market has been shrinking for 14 years straight with some crypto boom years here and there. See the chart I posted previously. Fewer and fewer people are buying discrete GPUs But they do though. You can find lower end AMD GPUs from the last two years for $125 (e.g. RX 6400) whereas the cheapest RTX 3000 or 4000 series is around twice that.
AMD cards do not "feast" on low-end and midrange. According to Steam charts, Nvidia still dominates midrange cards.[0] Furthermore, when I said "feast", I meant making profits. AMD does not make much profit from midrange or low end cards.The bottom line is, you keep wondering why no one is offering compelling value in the midrange area but there's a very obvious reason why: profit is not there.
Selling 30M GPUs with a huge margin is more profitable than selling 60M GPUs with a modest margin, and they can point to Bitcoin or AI as an excuse.
But also, we're talking about them crippling the cards "for gamers" so there will be cards "for gamers" -- the premise of this has to be that they're supply constrained (artificially or otherwise) because otherwise they would just make more at the evidently profitable price gamers are already paying. It can't be a lack of demand because the purpose of removing the feature is to suppress demand (and shift it to more expensive cards).
> One is a commodity. The other is about as high tech as it gets. Completely different economic rules that govern these products.
So you're saying that if a high tech product only has a limited number of suppliers then they could charge high margins even if customers are price sensitive.
> One has an entire board that needs soldering, assembled by a manufacturing line, tested, with many parts, and a team of dedicated engineers optimizing drivers constantly. The other is a CPU that is machine tested, and shipped with a heatsink fan unattached. Come on now.
GPU manufacturing is automated. The CPU heatsink isn't attached because it mounts to the system board, not because attaching it would meaningfully affect the unit price.
Driver development isn't part of the unit cost, its contribution per unit goes down when you ship more units.
You can buy an entire GPU for the price difference between the 8500G and the RTX 4060.
> That's not true. The discrete GPU market has been shrinking for 14 years straight with some crypto boom years here and there. See the chart I posted previously.
That's only because you're limiting things to discrete GPUs and customers have increasingly been purchasing GPUs in other form factors (consoles, laptops, iGPUs) which have different attachment methods but are based on the same technology.
> According to Steam charts, Nvidia still dominates midrange cards.
Steam is measuring installed base. That changes slowly, especially when prices are high.
> Furthermore, when I said "feast", I meant making profits. AMD does not make much profit from midrange or low end cards.
They make a non-zero amount of profit, which is why they do it.
Although I would modify your statement slightly:
Original: Selling 30M GPUs with a huge margin is more profitable than selling 60M GPUs with a modest margin.
Modified: Nvidia and AMD must sell at a higher ASP because the market for discrete GPUs has shrunk from 60m to 30m/year.
That's your answer! It's what I've been arguing for since my very first post. It isn't Nvidia and AMD's choice to have the market shrink in terms of raw volume. It's because many midrange gamers have largely moved onto laptops, phones, and consoles for gaming since 2010. The remaining PC gamers are willing to pay more for discrete GPUs. Hence, both Nvidia and AMD don't bother making compelling midrange GPUs.
I remember midrange GPUs that have great value such as the AMD HD 4850. I don't think those days are ever coming back.
That makes no sense as a reason not to sell more.
It also makes no sense in general because discrete GPUs aren't a distinct technology. The H100 isn't literally a bunch of RTX cards glued together, but it's approximately that and the same R&D goes into both, implying that the higher demand for this technology should allow for lower ASPs as you now have a new source of demand to spread the R&D costs into.
And the same is true for consoles and laptops. It's the same technology, it's just soldered to something instead of being in a PCIe card. Discrete GPUs aren't expensive because they can't justify the cost of the printed circuit board without selling more units, they're expensive because the market is consolidated and in the absence of more competition, Nvidia charges what they can get away with even when that is far in excess of what they would need to charge simply to remain a viable business.
> I remember midrange GPUs that have great value such as the AMD HD 4850. I don't think those days are ever coming back.
The way you get those days back is to get more competition. Support AMD and Intel and anyone else who might present a viable challenge to Nvidia so that Nvidia has to provide better value for money to keep you from switching.
You seem to think that midrange market shrunk in volume because Nvidia decided to stop offering value products, is that right?
Don’t support AMD or Intel if they have inferior products. There are now plenty of GPU makers out there including Qualcomm and Apple. Let’s not become fanboys here. R/ayymd is where you want to go if you want to become a blind AMD supporter.
I am not sure where that idea came from. It is from Samsung 8nm Fab. It never had the capacity to play with in the first place. Especially when Samsung Foundry is upgrading to chase with leading node.
There are still plenty of things being produced in fabs with older technology than that. Global Foundaries is the third largest in the world and they're offering 12nm or worse. People buy it because not everything needs a node which less than six months old and the price is right.
https://www.tomshardware.com/news/nvidia-makes-1000-profit-o...
Marginal profits need to be high when upfront costs are huge.
If you peel the layers back... isn't the real monopoly ASML?
So, as long as it treats fabs in Europe fairly, it can do whatever it wants in the rest of the world.
[1] https://en.wikipedia.org/wiki/European_Union_competition_law
So till ASML starts abusing the position they should be good legally.
But no worries: in a democracy we expect ordinary people to have some interest in the laws, after all, they are supposed to be electing people who decide on how laws should be changed (or not). So layman need to talk about laws, too.
I'm just always a bit cautious (or at least I should be). I know that eg in the US insider trading is about stealing secret from your employer; but in eg France insider trading is about having an unfair advantage over the public.
I can image that there are jurisdictions that treat monopolies by themselves as a problem. (Perhaps France, again?)
Btw, for the US herself have a look at https://fee.org/articles/the-myth-that-standard-oil-was-a-pr... to see how the prototypical case against Standard Oil wasn't really about monopoly abuse, either. At least no one really bothered proving that a monopoly was abused, they mostly just assumed it.
> Antitrust is not against monopoly, but monopoly "abuse".
So this might be true about anti-trust in the US right now. But I'm not sure whether it's true about ant-trust law in eg France?
Also, in practice this was not true about anti-trust law in the US historically: Standard Oil was smashed into pieces without anyone proving in court that consumers had been harmed, or that the monopoly had been 'abused'.
See eg https://fee.org/articles/the-myth-that-standard-oil-was-a-pr... and https://www.econlib.org/library/Enc/Antitrust.html
ASML is also part-owned by their customers. Intel, Samsung and TSMC all invested in ASML to get EUV tech over the line - see aforementioned capital requirements.
This could, theoretically, change at any moment, and Canon is trying something with their new generation of nano-print tech.
ASML is a strategic asset.
Maybe gamers can wait and the next technological leap should have priority?
People need entertainment, too, we're not machines.
If the current ML craze is really that impactful, it will happen regardless. If it can't break through it's because most of it is just a bunch of hot air.
lol
AI is goldrush and Nvidia is selling golden shovels.
They try to play it longterm though. Goldrush ends at some point, if they upset "gardeners" gamers, these may be jumping onto AMD or even Intel shovels already. That was already the case during Bitcoin goldrush.
But I wouldn't be also surprised this is a plan agreed at closed door meetings between big corporations. They may want to just kill any independent AI advantage, and force everyone to their cloud walled gardens. Future will tell.
Your AI startup has a 0.1% chance of becoming successful (going by SV definition of success, as yet another rent-seeking, privacy-abusing SaaS) while millions of gamers have a 99.9% chance of enriching their lives with entertainment. Statistically that's just the reality of this situation if you insist on being utilitarian.
We keep the GPUs, Nvidia keeps a diverse customer base, and you don't have to waste years of your life doing self-important busywork. Deal?
The steam survey suggests that Nvidia have over 90% of the GPU market, so for every design they sell somewhere near 10x the number of units, so if they sell at the same margins they get 10x the development resources per unit.
That's a lot of slack to lose to business inefficiencies, or competing with more specialized but smaller market devices. Assuming they don't just purchase such possible competitors when they pop up.
It may be that a monopoly is the "natural" end state of such tech markets. I think it's self-evident this isn't a good end state for consumers.
cutting edge hardware exists at prices that are affordable for hobbyists largely because a few key features for commercial use cases can be "artificially" turned off.
it sucks that you don't get to pay consumer prices for your highly lucrative application anymore, but the alternative looks a lot more like "tensor prices for geforce skus" than "geforce prices for tensor skus". we are already seeing this play out with crypto mining to an extent. I'd hate to see what would happen to the consumer market if AWS could just buy a bunch of RTX parts to rent out.
This has been running wild on the Internet. If anything Nvidia has arguably earned less premium in the consumer market with most of the initial R&D cost being amortised with their AI / datacenter chips.
I mean I'm sure they're overcharging for gpus in general, but the enterprise/consumer split is always tough, technically enterprise customers are subsidising the dev and research of new tech that eventually becomes consumer gpus.
Either its 20k/1k for top end of each or its idk, 7k for everyone?
They could definitely drop prices all around though and still have a buffer for r&d left over.
AMD hasn't done anything substantive in the past two years since GPT-3/GPT-3.5. Almost every research paper implements their algorithm in CUDA. AMD can't beat software with good hardware.
I'm at a startup, and we'd love to be using MI300Xs. Out stack works with it, they are amazing, our wallet is open... But we can't! We simply can't find any. It seems they are unobtanium for megacaps only.
Nvidia is in the limelight, but their product (GPU compute) is a commodity.
Once someone else has it cheaper, then it’s a race to the bottom.
The recent news was about AMD giving up on that path.
They still funded it and it was created.
They gave it to us instead of tossing it into the trash.
I don't know what you mean to imply by that.
Just because they don’t want it doesn’t mean it just vanishes or stops working.
> Shortly thereafter I got in contact with AMD and in early 2022 I have left Intel and signed a ZLUDA development contract with AMD. Once again I was asked for a far-reaching discretion: not to advertise the fact that AMD is evaluating ZLUDA and definitely not to make any commits to the public ZLUDA repo. After two years of development and some deliberation, AMD decided that there is no business case for running CUDA applications on AMD GPUs. > > One of the terms of my contract with AMD was that if AMD did not find it fit for further development, I could release it. Which brings us to today.
It's worth noting that while ZLUDA is a very cool project, it's probably not so relevant for ML. Also from the README:
> PyTorch received very little testing. ZLUDA's coverage of cuDNN APIs is very minimal (just enough to run ResNet-50) and realistically you won't get much running. > However if you are interested in trying it out you need to build it from sources with the settings below. Default PyTorch does not ship PTX and uses bundled NCCL which also builds without PTX:
PyTorch has OOTB ROCm support btw and while there are some CUDA-only libraries I'd like (FA2 for RDNA, bitsandbytes, ctranslate2, FlashInfer among others), I think sponsoring direct porting/upstreaming compatibility of the libraries probably makes more sense. Also from the ZLUDA README:
> ZLUDA offers limited support for performance libraries (cuDNN, cuBLAS, cuSPARSE, cuFFT, OptiX, NCCL).