U.S. clears way for antitrust inquiries of Nvidia, Microsoft and OpenAI
nytimes.com
nytimes.com
Hitting the only successful company in a difficult field with legal assaults is not the obvious path to success. Nvidia has nothing close to a monopoly; there is no moat and they operate in a commodity space that will behave like a normal competitive market. It just happens that their competitors were wall-to-wall complacent about how quickly the hardware world can change and they stupidly thought graphics cards were for displaying graphics. It turns out graphics cards are for building AI and crypto markets.
I don't understand. If OpenCL was so good, why did Apple create Metal instead of just using OpenCL?
OpenCL isn't bad, and I'd love to see it get to the point where it competes with CUDA as originally intended. The peaceable sentiment of "let's work together to kill the big bad demon" seems to be dead today, though. Everyone would rather sell their own CUDA-killer than work together to defeat it.
https://john.cs.olemiss.edu/heroes/papers/AMD_OpenCL_Program...
Bonus points: the rest of the software libraries intended to compete with the CUDA ecosystem are still online in the “HSA Toolkit” GitHub repo. Here’s their counterpart to the Thrust library (last updated 10 years ago):
https://github.com/HSA-Libraries/Bolt
Nvidia had multiple updates in the last year the last time I checked. That’s the problem.
Eg, opencl on Nvidia GPUs for windows is/was missing (I have a many-year GitHub issue with them + Microsoft), which matters for individuals, and Nvidia does not support OpenCL for its core convenience analytics libraries like the RAPIDS Python ecosystem, which is core to its massive data center market. We initially built for compatibility/distribution, but it didn't matter: That gap closed a lot of doors for us as a small ISV choosing what & how to build, and in turn, ultimately prevents our customers from buying AMD, Intel, etc
I'm not up to snuff on whether that qualifies as using its position anti-competitively anywhere, but it's a real market issue
If anything, the situation with OpenCL suggests AMD and friends were the ones dragging their feet. Nvidia correctly identified that being led by the OpenCL committee would lock them out of billions (trillions, if you believe the stock market) of profit and routed around the blockage rather than compromise their engineering standards.
> including how the company’s software locks customers into using its chips, as well as how Nvidia distributes those chips to customers.
The lock-in is probably a losing argument on technical merits unless NVIDIA is doing something nefarious there (which the lawsuit will potentially reveal). If nVidia is distributing chips in preferred ways to partners it wants to succeed (e.g. those partnered with MSFT which it has a very cozy and tight-knit relationship), that could be an anti trust thing.
> Microsoft structured its minority stake in OpenAI in part to avoid antitrust scrutiny, The Times has reported. Microsoft also reached a deal in March to hire most of the staff of Inflection AI, another A.I. start-up, and license its technology. Because the deal was not a standard acquisition, it may be harder for regulators to scrutinize.
So there is an argument to be made that MSFT is acutely self-aware its behavior could be viewed as engaging in antitrust. If there’s a law against structuring financial transactions, why isn’t similar behavior in the business world viewed similarly as trying to bypass the spirit of the law?
MS knows all about antitrust. They've hired a bunch of lawyers and told them to do what needs to be done to be compliant with the regulations. That shouldn't then be implied as evidence that they are guilty of antitrust activities! What are the lawyers supposed to do, structure the company to maximise the likelihood of an antitrust lawsuit being bought? Nobody should want that, regulator, company or customer.
The reason this is important is it's a sword that cuts both ways. It's always tempting to argue that whatever you'd personally like to happen is what politicians really intended, and any gap between reality and their preferred outcome is therefore a "loophole". But once you get into saying people should follow intent, not written law, others can easily argue that politicians never intended the law to be interpreted like that against them, and so therefore they are morally justified in ignoring it. It can be used against you as easily as you can use it against them.
A common example of this problem is income vs capital gains taxes. One ideological tribe is very fond of arguing that people who have income mostly from investments rather than wages are exploiting a "loophole" in tax law, but of course the reason there are different rates to begin with is exactly because politicians wanted to encourage investment. There's plenty of cases where this intent is discussed in written literature and there's no other reason to distinguish between income sources then set differing rates. There is no "loophole" and nor is the "spirit" of the law being violated. But you hear such claims all the time.
The other reason it's problematic is because you can't really know what the intent of lawmakers was. The law was their best collective effort at writing down what they wanted, as a result of numerous compromises and disagreements between different people. If they didn't write it down properly or the resulting compromise was a mess, that's on them, but a working legal system doesn't allow people to just blow off their written instructions and assume they know what was really meant.
Actually, there is. Many judges take various interpretations of the 'spirit' into account. See eg https://en.wikipedia.org/wiki/Originalism for one example.
And if the courts use some 'spirit' guide them, companies and their lawyers better try and predict what that 'spirit' recommends.
> If they didn't write it down properly or the resulting compromise was a mess, that's on them, but a working legal system doesn't allow people to just blow off their written instructions and assume they know what was really meant.
Well, then by that definition the real life legal system of the US ain't working?
These two perspectives aren't both equally valid: the courts are not allowed under any system of civics to simply do whatever the judge feels like. They are only given leeway to interpret the law when they have no other choice because the statutes are unclear. It's a last resort, and often the result will be people "getting away with it" because there's no law against what they did.
I can’t comment on the claims here, but if MSFT is skirting the intent brazenly enough, the predictable regulatory reaction is no bullying. Everyone involved is well aware of that dynamic. No need to pity MSFT.
In common law systems (US, UK), there is greater emphasis on the spirit or intent behind the law. Judges have more flexibility to interpret laws and use general principles.
Civil law systems place more importance on the letter of the law. The spirit behind the law is secondary. The written legal codes are more important.
EU law is mixture of both, but most continental European countries have civil law system (Ireland does not, Nordic countries have mix of civil and common law)
Spirit of the law isn’t perfect itself obviously because it is legitimate to point out that it becomes hard to know what the law actually is, especially since governments have gotten into legislating a lot of regulatory nuance and it’s hard to distinguish malicious compliance from good faith effort. It’s also corruptible from overzealous prosecution looking to make a name for itself by taking on unpopular entities that aren’t actually doing anything wrong.
This is precisely the opposite. "Spirit of the law" makes the rules squishy and indeterminate, providing opportunities for fancy lawyers to bend the result to their own interests.
"Letter of the law" often leads to harsh results when the law is drafted poorly, because if they wrote something dumb then you get something dumb instead of a judge rewriting the law to make people happy. But the people they're making happy are usually the powerful, so pick your poison.
It’s also important to remember that societies naturally undergo shifts over time. It’s impossible to continuously update a codified set of laws when the underlying moirés of the time have shifted; you’ll just be constantly arguing over the updates to add. Any law written perfectly today becomes imperfect simply through the passage of time. That’s why the Bible and any prescriptive religious text feels so outdated on many recommendations - it’s a snapshot in time of the values of a culture but those values change. There was even a fantastic sci-fi short story on this exact point of cultural shift [1] that’s worth a read.
The assumption here is that the rules would be complicated and provide lots of opportunities for gamesmanship. Now suppose the rule is "no company shall have more than 30% market share in any market, any that does shall be broken into no fewer than twelve independent pieces."
No loopholes, if you exceed 30% market share you get broken up. And if they find a loophole then you amend the law and take it out.
> It’s impossible to continuously update a codified set of laws when the underlying moirés of the time have shifted; you’ll just be constantly arguing over the updates to add.
That's just politics. Somehow you need a process to decide what the law should be. The output of that process is the new law. If the output sucks then get a new process. But whether people can agree on what the law should be is a separate issue than whether we should even know what the law as enacted is supposed to mean.
> That’s why the Bible and any prescriptive religious text feels so outdated on many recommendations - it’s a snapshot in time of the values of a culture but those values change.
That's fine, nobody is saying that you can't change the law if a case comes out in a bad way. But it should be the legislature rather than the courts to do it, and the new understanding shouldn't be applied to past behavior ex post facto.
Literally me? Okay, I don’t own 50% of the market—my company does.
Oh you mean my company? Okay, my company doesn’t own 50% of the market—each of my companies only control 25%.
Oh you mean me and my companies? Okay, well I only own one company and my wife owns the other one.
Oh you mean… I think you get the point…
Now move on to all the other words you used: what defines a “market”? What does “broken” mean here? What does “independent” mean here? I’m sure it’s quite clear to you—that you know it when you see it. I’m also quite sure others have different interpretations.
I agree with the spirit of your comment otherwise, but simple laws rapidly become complex laws because people are complicated and language is flawed.
But we know that the interpretation phase has a lot of wiggle room (the US constitution is a regular parade of this sort of thing - the Roe v. Wade fight or the abuse of the commerce clause for example). When they go bad these things aren't loopholes as much as they are just ignoring the written law with a polite fiction and it is up to different interest groups to work out where the power lies to get what they want. Political reality vs. the theoretical rule of law ideal.
But that is necessarily independent of what is written in the law itself. If a group has enough power to overrule the written law then what you write in the law won't be able to stop them.
"no company shall have more than 30% market share in any market, any that does shall be broken into no fewer than twelve independent pieces."
It doesn't contain the word "you".
> Okay, I don’t own 50% of the market—my company does.
Then your company would be violating the law. "Company" means a set of entities that share a common ownership.
> what defines a “market”?
A set of products or services that serve as fungible substitutes for one another.
> What does “broken” mean here?
It means they no longer share common ownership. This is also what independent means.
> I agree with the spirit of your comment otherwise, but simple laws rapidly become complex laws because people are complicated and language is flawed.
But all of those things are just their ordinary meaning. Writing them down would make the law more explicit but it doesn't make it any more complicated. The definitions aren't each a separate set of criteria that have to be complied with separately, they're just a clarification to reduce possible ambiguity.
In particular, what you're doing is resolving edge cases. But the basic law has already addressed 99% of cases, because they're not ambiguous. An independent restaurant in a major city does not have >30% market share for food because there are many, many competitors. Microsoft has >30% market share for desktop operating systems because Microsoft has ~70% market share for desktop operating systems.
And we have to distinguish between two things here. One is, you see the word "company" and the dictionary says one of the meanings is a military unit, and then Microsoft claims that they aren't a company because they aren't a military unit. But if something has two meanings and one of them doesn't make sense in context, that's not the intended meaning. Using the "spirit of the law" for this kind of resolution is inherently necessary.
The other is, the law isn't ambiguous, but the unambiguous result is undesirable and the only way for a judge to fix it is to disregard the text and make something up. They shouldn't do this.
Sometimes this would make the law more complicated, because the situation has intrinsic complexity and you have to enumerate the edge cases. Sometimes it would make the law simpler, because the existing complexity is extraneous and only an opportunity for gamesmanship. But either way it creates an evolutionary pressure for improvement.
Also, I think you’re conflating my statement which was meant to be prescriptive with a descriptive statement. I’m saying if structuring and smurfing are illegal in a financial context, MSFT doing so to skirt antitrust laws should be similarly illegal - if the laws don’t prohibit it, the laws should be updated. But I’m not a lawyer so it’s possible laws already prohibit it or MSFT violated some laws. Commenting in either direction on an information free article isn’t wise.
As for the spirit of the law vs letter of the law, that’s an ongoing debate as old as history. It’s personally weird to me to encounter letter of the law people given that the spirit of the law has a much richer history behind it to my view and seems more defensible. Some amount of rule lawyering is required, but violating the spirit of the law is a more robust legal principle that can withhold participants trying to find creative workarounds around the spirit of the law.
I don't think you're in a position to bully Microsoft. I doubt they care about your opinion.
“We packed the goods in nondescript boxes to avoid scrutiny by the border guards…”
Scrutiny isn’t something you need to avoid as a big company if you’re not fucking around. The lawyers are there to handle the scrutiny, not to hide stuff from it.
It doesn't. That is like saying that if someone structures their affairs to pay less taxes it implies that wanted to do something illegal. They had to choose some structure, they picked the safest and easiest one. In both the hypothetical tax case and the real MS case, I suspect.
The incentive is to choose structures that minimise regulatory engagement. The fact that MS is following incentives isn't a red flag. Not a green one either, it isn't interesting except as something that the NYTimes can hook insinuations on.
Structuring is explicitly illegal in the US.
This is the point that vlovich123 was making, and given you totally missed it, it seems like you might not be familiar with what the term means in this context.
Structuring your business to avoid breaking the law is good.
Structuring your business to avoid some otherwise regular audits that would reveal whether you are breaking the law is obviously suspicious. One might think that there is a reason why getting audited might result in negative consequences for your business.
But then you consider that it's used for civil asset forfeiture (i.e. law enforcement stealing money from innocent people without ever proving a real crime) and it suddenly makes sense:
https://www.washingtonpost.com/news/the-watch/wp/2014/03/24/...
Whereas with MSFT it's a different situation. There is nothing illegal or wrongful about depositing a large amount of money, there is just a heinous law that makes it illegal to do it in a common way. Whereas monopolizing a market is intended to be illegal regardless of how you do it, so there shouldn't be a way to structure things to monopolize the market without breaking the law. The way you're intended to comply with the law is by not monopolizing the market.
With banking, structuring laws mean banks need to report any transaction over $10,000.
So if I need to send you $100,000 and s split it up into 11 payments so each payment comes under the $10,000 reporting limit that breaks the law because my payment was for $100,000 but i structured it to get around the $10,000 reporting limit.
If Microsoft approaches anit trust limits if it owns more than 50% of a company and buys 49% then no strucuturing took place as they only bought 49%. Now if they bough 10% in each of 6 different shell companies that they own such that they'd control 60% of hte company, then this is illegal as it works around the limits.
But owning up to the limit is perfectly fine as it doesn't in anyway break the law or work around it. That's the very reason for having a limit, to say you can go up to this limit, but not over.
I'm guessing you don't deal in finance at all as we deal with this all the time. You'll see funds owning up to 9.9% of companies to avoid the 10% reporting threshold. Again, nothing illegal here as the government has said they are perfectly fine with funds owning up to <10% of a company.
If I think I might be doing something that antitrust regulators won't like, so I hire a lawyer to tell me what I can do, and they tell me "you can buy up to 49.9% of this company," and I buy 49.9% of the company, that's fine. It's neither illegal nor unethical. It's also not "structuring" in the technical, financial sense, but a layman might say "you structured the deal to avoid falling afoul of regulation," even though it's the wrong use of the word. I think that's what's happening in this conversation (I have no idea about what MSFT actually did or didn't do, just the direction the conversation here is going).
This depends on jurisdiction, and what judges you get.
Eg Standard Oil was slapped with anti-trust law, despite neither having a monopoly nor exploiting its market share like a monopoly. See eg https://en.wikipedia.org/wiki/Standard_Oil#Legacy_and_critic... or https://fee.org/articles/the-myth-that-standard-oil-was-a-pr...
> Some economic historians have observed that Standard Oil was in the process of losing its monopoly at the time of its breakup in 1911. Although Standard had 90 percent of American refining capacity in 1880, by 1911, that had shrunk to between 60 and 65 percent because of the expansion in capacity by competitors.
The following is for US law.
Your first sentence is not really needed because US monopoly law really doesn't care if you are actually in pole position. It's more about monopolization than it is about monopoly.
To a first approximation think of it as being about fair competition. You could have a complete 100% monopoly in some particular market but if you got that monopoly by simply outcompeting everyone else by making a better product and you were not trying to use that to expand into other markets by doing things like tying you would probably not have an antitrust problem.
I wonder why Jensen wouldn't help his cousin Lisa Su with a good advice :)
In general it isn't surprising that hardware companies like Intel and AMD fall at software - just talk to any programmer at such a hardware shop. What striking is that Jensen is able to have such a prominent and huge software development in a naturally hardware company. NVDA is the only hardware company where my programmer acquaintances don't complain about it being a hardware company.
A lot of what keeps Intel relevant is their prowess in both hardware and software, bringing properly refined products to market when it matters. Intel's failure against Nvidia had far more to do with their business strategy of CPU First missing the broadside of a barn; it's like bringing a knife to a gunfight, of course you're going to get your ass wrecked.
AMD fucking sucks at software though, no disagreement there.
The oneAPI heavily contributed to their GPU failures IMO. Choosing SYCL and promoting non-existing performance portability and pouring resources in the trying to shove CPU and GPU (and at some point FPGA!) into the same programming model while disinvesting from OpenCL was a big mistake, but it was probably the only way to please both CPU and GPU management.
And the problem is not SYCL per se but the fact that they try to make it an open standard rather then trying to provide a way to extract maximum performance from their HW. They needed to build a CUDA alternative, where, after jumping through considerable hoops, you can get to actually programming tensor cores for those who need it. Not surprisingly, Intel is not using SYCL for their GPU kernels (look at oneDNN sources). With Raja gone there may be hope, but I'm not holding my breath.
The only thing that they have that is competitive is Habana's stuff which is outside of the oneAPI.
I would not be surprised if AMD has a similar issue of infighting due selling both CPUs and GPUs.
And NVIDIA does not have this problem.
It is if success is defined as deliberate failure.
Personally, a lot of US government actions of the past few decades in hindsight do not look like they had the interests of America or Americans in mind.
Video games have used compute shaders forever, and those work just fine on other cards.
I remember doing some quite involved stuff for my thesis using OpenCL a decade ago, and it worked just fine. Nowadays OpenCL is dead for some reason...
I just don't get what is there in CUDA that makes it so special. As far as I remember, a GPGPU API consists of a shader language, a way to copy stuff to and from the GPU, a way of scheduling and synchronizing work, atomics, groupshared variables, and some interop stuff.
Is the vendor lock-in due to CUDA libs? In that case, the problem is not CUDA, but the libraries themselves. Not sure about today, but performance-portability basically didn't exist back then. You needed to do specialize your code for the GPU arch at hand. It didn't even exist between different generations of cards by the same vendor. Even if you could run CUDA code on AMD, it would be slow, so you need to do a rewrite anyways.
The vendor "lock in" is because it takes a few years for decisions to be expressed in marketable silicon and literally only Nvidia was trying to be in the market 5 years ago. I've seen a lot of AMD cards that just crashed when used for anything outside OpenGL. I had a bunch of AI related projects die back in 2019 because initialising OpenCL crashed the drivers. If you believe the official docs everything would work fine. Great card except for the fact that compute didn't work.
At the time I thought it was maybe just me. After seeing geohotz's saga trying to make tinygrad work on AMD cards and having a feel for how badly unsupported AMD hardware is by the machine learning community, it makes a lot of sense to me that it is a systemic issue and AMD didn't have any corporate sense of urgency about fixing those problems.
Maybe there is something magic in CUDA, but if there is it is probably either their memory management model or something quite technical like that. Not the API.
It's crazy, because even 10 years ago it was already obvious that machine learning was big and is only going to become more important. AlphaGo vs Lee Sedol happened in 2016. Computer vision was making big strides.
5 years ago, large language model hadn't really arrived on the scene yet, at least as impressively as today, but I think eg Google was already using machine learning for Google Translate?
But somehow Zoom overtook them during the pandemic.
The problem I had with my AMD card was that SYCL, like every other API, will involve making calls to AMD's kernel drivers and firmware that would crash the program or the computer (the crash was inevitable, but how it happened depended on circumstances).
The AMD drivers themselves are actually pretty good overall, if you want a desktop graphics card for linux I recommend AMD. Open source drivers have a noticeably higher average quality than the binary stuff Nvidia puts out. Rock solid most of the time. But for anything involving OpenCL, ROCm or friends I had a very rough experience. It didn't matter what, because the calls eventually end up going through the kernel and whatever the root problem is lives somewhere around there.
Beyond that, AMD’s implementation is broken.
Same as Vulkan Compute - SPIR-V could be cool but it’s broken on AMD hardware, and AMD institutionally opposes hitching their horse to a wagon they didn't invent themselves.
This is why people keep saying that NVIDIA isn't acting anticompetitively. They're not, it's the Steam/Valve situation where their opponents are just intent on constantly shooting themselves in the head while NVIDIA carries merrily on along getting their work done.
I have an AMD card that I try to use for AI stuff and it's an uphill battle. The most popular workloads such as Stable Diffusion or Llama more or less work. But as soon as you go into other, less mainstream workloads it starts to fall apart. Projects install CUDA versions of torch by default. To make them work you have to uninstall them manually and install the ROCM versions of torch. Then it turns out some Python dependency also uses CUDA, so you also have to find a ROCM fork for that dependency. Then there's some other dependency which only has a CUDA and CPU version and you're stuck.
Where does SYCL fit into this picture, is it a viable replacement for cross-platform GPU access?
You're pretty close, but it's one level of abstraction away from the truth: NVidia spends absolutely earth-shakingly stonking amounts of money working directly with the devs of major studios, fixing issues with their drivers as they come, and in some cases paying studios to lock-in talking to them about the bugs they find.
It's why (and how) they release new driver patches for every AAA release, and part of the reason why it triggers a redownload of the shader cache on a whole slew of Steam games every major version release.
AMD and intel don't have a chance; NVidia is inside the design loop for game development far more thoroughly than other hardware vendors.
In fact I wish Intel and AMD would do the same! It's great customer service to know that games are going to work well out of the box with Nvidia, because Nvidia spent tons of money working directly with devs on optimization and even per-game bugfixes.
As a longtime PC gamer I wish AMD released better GPUs, but they seem incapable of doing so generation after generation. Intel is getting better at the low end, so there's at least some hope there. But Nvidia is just the best in class, because they put in a huge amount of polish and work.
My understanding is this is the same approach Nvidia took with CUDA: investing a massive amount to make it fast, accurate, and broadly available, while AMD continued to push out barely-tested drivers — for example, see hotz's woes with tinygrad and finally giving up and labeling the "red" (aka AMD) tinybox driver quality as "Mediocre" on their product page, with the "green" (aka Nvidia) tinybox driver quality listed as "Great" https://tinygrad.org/ — and to this day restricts their competing library (ROCm) to specific ultra-expensive server-class GPUs.
Intel at least works very closely with Microsoft to iron out the details for Windows.
Anyone can talk, especially nowdays because of AI, ironically.
That's literally dictionary definition of idolatry.
It has a decent number of up votes, so clearly some people think it worth posting. As I said, you don't have to read it, and I don't really care if you do. I'm not sure how linking to something with a reference is worship, but ok.
You seem very angry.
What you're doing, explicitly, is placing more value on name than idea. A fallacy. I'm obviously not saying "worship of a minor deity" literally, but colloquially.
You're right, I can read it or not. Same with you, you can take valid criticism or ignore it and be all the poorer for it.
You can also continue admitting you have no rebuttal as you hyper-focus on the _diction of the conversation_ instead of the _conversation_, like how to define worship and idolatry in a casual conversation.
Generally discourse would be better if we linked to well argued and reasoned articles than have a load of blow hards that like the sound of their own voice rather too much.
If you're going to say things like "literally the dictionary definition of idolatry" when it's nothing of the sort, expect people to pick up on it. Your reasoning appears to be very much in the domain of a cheese.
I shall no longer be replying to or monitoring this thread. Take care.
And this article, especially, really put into words what I had been thinking for 2 decades:
A real question might be how Nvidia prioritizes capacity and delivery and pricing to OpenAI and Microsoft for their superior product that has no meaningful competitor. It’s not an issue that Nvidia, like standard oil, had a superior business and product. It’s an issue when their scale and the scale of their adjacent partners lead to agreements that strangle others that need access to the supply chain. For instance if Anthropic is at a disadvantage to open ai for Nvidia capacity due to agreements between them that are mutually beneficial to the exclusion of Anthropic, that’s anti competitive.
See https://en.wikipedia.org/wiki/Standard_Oil#Legacy_and_critic...
They even lost market share during the time they were alleged of having and exploiting a monopoly:
> Although Standard had 90 percent of American refining capacity in 1880, by 1911, that had shrunk to between 60 and 65 percent because of the expansion in capacity by competitors.
See https://fee.org/articles/the-myth-that-standard-oil-was-a-pr... for a more opinionated take.
Bro...
But it’s a bizarre misreading to say they were not a monopoly to say that by the time things hit the fan they had lost a lot of market share from their peak monopoly.
The specific issues that standard oil hit up against was their deals with railroads that solely benefited them by virtue of their scale and their ability to effectively dictate their own pricing for logistics. Likewise they would buy up all capacity on crucial lines for competitors and only partially use it purely to block market access for competitors. There is also a view that being incredibly efficient can reduce competition in markets adjacent to you. Success in business for an individual company isn’t the only goal of a capitalist state, but rather competition in markets is considered a goal of the state in itself - survival of the fittest implies there’s something to survive against, and an ever expanding monopoly growing into new markets and dominating them starts to retard innovation and competition systemically. You don’t have to agree with the thesis, but that’s the way the system works.
One could very easily argue this is the status-quo; Microsoft ships DirectX which they withhold from competitors, and Apple does the same with Metal. Both of those monopolize their respective markets (PC gaming and mobile gaming) yet we see zero attempts to reconcile the two.
Because of the perceived hostility between vendors, I would not be surprised to see Nvidia argue that they're already as open as they could possibly be.
Apple doesn’t have a monopoly on mobile gaming or even mobile. Windows is more of a monopoly than iOS, but it’s not a monopoly in gaming platforms by a large margin. OpenGL is a totally viable alternative. Etc.
CUDA is definitely a big part of the lock Nvidia has on their market. Other companies could with investment out class then hardware wise. But because of their API lock there’s no point.
Not really, given that Apple platforms have more or less entirely depreciated the featureset with no guarantee or expectation to keep it around.
> Other companies could with investment out class then hardware wise.
I don't even believe you. Who else can compete with Mellanox networking and custom ARM v9 cores? Even Apple doesn't really ship anything worth comparing to Nvidia's server offerings, and they're arguably best-positioned to usurp Nvidia's TSMC friendship.
When you hold a monopoly the same rules don’t apply. That’s why it’s often advantageous to allow material competition in your own market because your adjacency influence can be outsized and cause unintended consequences -even if you normally would have been allowed to do such things-. Those unintended consequences don’t require malicious intent they just have to exist and be materially a result of your monopoly. The more your maliciously collude though the worse the remedy will be for you.
Sure, until you consider the massive difference in accessible capital vs their rivals. Sure, VCs can fund startups, but Nvidia is publicly traded, which is just investment of a different form, and largely a form that you are barely accountable to. And that's on top of all the capital they directly control. That's an incredibly advantageous position to be in, and one that would be too tempting for most to resist exploiting.
Raja actually specifically called out 2012 as being the timeframe when that happened, because AMD decided the future of GPUs was iGPUs and not dGPUs, that the latter segment was going to go away entirely as iGPUs got faster and faster. Oops.
Isn't the CUDA API a fairly significant moat?
But I stress that the issue there isn't that CUDA has tutorials. It is that I expect the tutorial code not to lead to a crash or system lockup. There is a well supported path to do those basic operations.
If you look at the CUDA webpage [1] you see things like "cuRAND", "cuFFT" and "cuBLAS". That isn't a moat - those are first year software engineering topics. AMD managed to make it look like a moat by not taking the compute market seriously; and anyone else competes at highly parallel compute.
[0] As far as I can tell; doesn't exist.
It is a bit of a chicken-egg problem for me. I couldn't make OpenCL work on an AMD GPU, so I didn't manage to learn that much about OpenCL. At the time I assumed it was just me, but in hindsight I never saw an OpenCL-based approach to a compute problem that worked reliably on my machine so maybe it wasn't.
But I don't think it really matters. The algorithms in the field don't seem to be hard and I never felt like I was struggling when implementing them on the CPU without any special API at all. My issues were conceptually similar to George Hotz's famous rants where he had crashes when running the demo app in a loop. In the experimenting phase I found I couldn't run code on the GPU with any API.
I'm sure the situation improved and part of it was just me; towards the end of my time with AMD I could run stable diffusion inference and it'd work great for 10-40 minutes before the kernel paniced or whatever - so it was definitely technically possible to get a "hello world++" style thing running. But I never felt it was the APIs that were holding me back.
And openly undermined technologies and standards that aimed to do the same in an interoperable way.
This is, in my eyes, enough to hit with a litigation, but we'll see.
Isn't it one of the fundamental applications of a GPU, whether for ML compute or more commonly for 3D graphics, to multiple matrices? In what way did AMD GPUs fail to do this?
And when you try to do compute on AMD GPUs they crash your kernel if you look at them wrong.
Investigating is fine, but if the concern is the direction of AI in general then that's a job for legislation, not antitrust action. I.e. a job that Congress kind of sucks at.
Examples of Apple doing that is banning competing browsers on iOS and then pushing W3C and developers in the direction they want due to "you can't ignore us". There was a whole list of bad examples. Touch events, fighting against SPIR-V in WebGPU, fighting against adoption of Media Source Extensions (to benefit their video solutions) and etc. and etc.
They very clearly cause a ton of damage to the market by slowing down and sabotaging the progress of interoperable technology to harm competition.
And just because they don't rush to incorporate every web feature doesn't make them anti-competitive. Especially when most of the time they are right to do because either (a) they impact security or battery life or (b) they are non-standard.
Case in point SPIR-V which unless I am mistaken is exclusively controlled by Khronos.
And you should know well it's the same thing, since above listed issues are defined by the engine. Whatever label you put on top doesn't change the essence of what the problem is.
Basically, you completely missed the point.
There is far more to a browser than just the engine as we've seen with Arc.
Meanwhile, I'm happy with the direction they push W3C and developers. Without them, Google would just push the web to whatever they want (more than they already do).
Seems like EU finally started getting the point before US regulators did.
I just think that regulations in general should be applied only when necessary, and then be applied with great force.
The user experience for game consoles largely falls back on the manufacturer. If Xbox had loads of unvetted buggy games and malware swimming in the ecosystem, people might be less inclined to buy an Xbox. So Microsoft sets about establishing some control over the ecosystem. Apple's App Store was really the first time that your average Joe could download an application from the internet and not have to worry about viruses. It was a big deal that added a lot of value to the user experience.
To give an example. If CUDA wouldn't have been tied to Nvidia hardware, developers could use CUDA on any competing hardware. Being tied limits the choice. That's the essence of lock-in damage. Development tools should be development tools, not ways to control the market.
That's why there is value in something that breaks lock-in - that improves competition.
Initially I thought that AMD didn't offer a CUDA reimplementation out of NIH-syndrome (as it would be a marketing coup for NVIDIA), but then I saw that NVIDIA seem to be actively trying to shut-down independent attempts at running CUDA on non-NVIDIA hardware: https://www.techpowerup.com/319984/nvidia-cracks-down-on-cud...
Just do it. Fuck the lawyers. When you win, you toss some cash and they'll ya for it.
https://www.cnet.com/tech/tech-industry/sun-microsoft-settle...
It's been almost 25 years ago, but this was the crux of the Sun Microsystems v. Microsoft lawsuit. It's why OpenJDK exists. Even three years ago the Supreme Court ruled that you can't copyright an API.
I mean, it feels breathtakingly obvious. Not only is there historical precedents for just doing this, there're legal precedents as well.
When someone manages to overcome the cost and commoditize what lock-in was walling - it improves things. Even better if someone manages to do it while breaking lock-in itself. Sort of like writing a wrapper to run CUDA on non Nvidia GPUs (what ZLUDA is doing).
Anti-competitiveness is more leveraging the must-have nature of CUDA to shelter Nvidia from competition. E.G. an OEM can’t sell professional laptops that support CUDA if it sells AMD gaming laptops or something - just to illustrate. Or Coca Cola refusing to sell coke to a store that also sells Pepsi causing the store to not stock Pepsi.
Monopoly law isn’t meant to simply punish success and exclusivity.
I.e. a must have can be a must have without artificially added tying (lock-in) detail. Then it's not causing harm.
As far as typical gameworks examples, I am specifically thinking back to tessellation - everyone freaked about Crysis 2 (despite the fact that it didn't actually use max-LOD or no-culling during actual gameplay) or Hairworks. All of a sudden it stopped being a thing, around the Polaris/Vega era, right? You know why that was?
Because AMD finally fixed their tessellation performance. Their hardware was deficient and underperforming, and the problem went away when they fixed their hardware.
The same is true of the software today. AMD's problem isn't that NVIDIA is stomping on their fingers, it's that ROCm crashes when you run the tutorial/sample programs.
So has enough anticompetitive behavior occurred to distort the market? Unless I'm missing something, I'm just not seeing this investigation having legs.
I have been there for this history, working on ML on GPUs at NVIDIA for a few years before Jensen decided to productize my little research project, CUDNN.
CUDA won big because it made a big bet. Were it that OpenCL was ubiquitous and at feature-parity with CUDA, maybe there would be more than one player dealt-in at the table today. But everyone else folded while Nvidia ran the dealer for 10 long years.
> In pushing for CUDA, Jensen Huang aimed for the Nvidia GPUs to become a general hardware for scientific computing. CUDA was released in 2006. Around 2015, the focus of CUDA changed to neural networks.[8]
Credit to Jensen for pivoting, but I recall hearing about CUDA networks from Google tech talks in 2009 and realizing they would be huge. It wasn't anything unique to realize NNs were a huge innovation but it did take another 5 years for it to mature enough and for it to become clear that GPUs could be useful for training and whatnot. Additionally, it's important to remember that Google had a huge early lead here & worked closely with Nvidia since CUDA was much more mature than OpenCL (due to intentional sabotage or otherwise) and Nvidia's chips satisfied the compute needs of that early development.
So it was more like Google leading Nvidia to the drinking well and Nvidia eventually realizing it was potentially an untapped ocean and investing some resources. Remember, they also put resources behind cryptocurrency when that bubble was inflating. They're good at opportunistically taking advantage of those bubbles. It was also around this time period that Google realized they should start investing in dedicated accelerators with their TPUs because Nvidia could not meet their needs due to lack of focus (+ dedicated accelerators could outperform) leading to the first TPU being used internally by 2015 [2].
Pretending like Jensen is some unique visionary seeing something no one else in the industry didn't is insane. It was a confluence of factors and Jensen was adept at navigating his resources to take advantage of it. You can appreciate Nvidia's excellence here without pretending like Jensen is some kind of AI messiah.
[1] https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=1575717 [2] https://dl.acm.org/doi/pdf/10.1145/1390156.1390170
it's actually specifically not questionable because it's been covered by journalists etc.
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.”
> Around the time Huang sent the e-mail, he approached Catanzaro, Nvidia’s leading A.I. researcher, with a thought experiment. “He told me to imagine he’d marched all eight thousand of Nvidia’s employees into the parking lot,” Catanzaro said. “Then he told me I was free to select anyone from the parking lot to join my team.”
> In marketing cuda, Nvidia had sought a range of customers, including stock traders, oil prospectors, and molecular biologists. At one point, the company signed a deal with General Mills to simulate the thermal physics of cooking frozen pizza. One application that Nvidia spent little time thinking about was artificial intelligence. There didn’t seem to be much of a market.
> Despite the snub, Hinton encouraged his students to use cuda, including a Ukrainian-born protégé of his named Alex Krizhevsky, who Hinton thought was perhaps the finest programmer he’d ever met. In 2012, Krizhevsky and his research partner, Ilya Sutskever, working on a tight budget, bought two GeForce cards from Amazon. Krizhevsky then began training a visual-recognition neural network on Nvidia’s parallel-computing platform, feeding it millions of images in a single week. “He had the two G.P.U. boards whirring in his bedroom,” Hinton said. “Actually, it was his parents who paid for the quite considerable electricity costs.”
Basically Hinton thought CUDA was a better platform and kept encouraging his acolytes to utilize it with NVidia completely ignoring it. For what it’s worth ChatGPT and LLMs are not the start of the story. I remember seeing Hinton present his research in 2009 and thought it was a revolutionary step forward. It took a much longer amount of time for NVidia to notice that they should be marketing CUDA to AI. They did so around the same time that Google started building their own inference accelerators for Tensorflow. So at that point NVidia was getting multiple signals already from a big customer that there’s something here.
Yes, it's probably true that in 2009 or whatever it wasn't on NVIDIA's radar as being a major R&D priority. It was by 2014-2015, which is what I cited.
You claimed it couldn't be proven that Jensen specifically targeted AI as a priority, and it's clear that by 2015 or so he was, at a time when AMD was not even taking GPGPU itself seriously as a development priority let alone focusing on a specific niche or field like AI.
> It's questionable whether or not he anticipated this specific market ...
again: no, it is not, literally he's been focused on this specific market for a decade and we have the receipts.
if your point is that he didn't see it fifteen years ago instead of ten years ago... ok I guess you win then, discussion concluded.
They look at what you are doing with the power you actually have. Once you are the size and importance of MS, OpenAI or Nvidia, the rules change and everyone knows that.
It is definitely Jensen's fault.
I think it is a law of nature that shitty and harmful behaviour pays off, but that's not a problem, if we can correct it time-to-time. With antitrust inquires for example.
Just have a look at the size that Google, Microsoft, Meta others have got, not because they have a outstanding product (like Cuda), but because they can use their dominant market position to buy smaller businesses and monetize their products as part of a bundle (Whatsapp, Activision Blizzard acquisition and the list goes on). Or just build a new product and shove it to their customers, like with MS Teams/Threads.
The US government seems to have actually aided those US companies to keep doing their monopolist practices, with many other countries around the world fighting against them alone, like Europe.
The situation is so bad right now, that Europe itself is now distancing itself from US to potentially also fight even harder against those companies, which bring barely any jobs to Europe, but have market share in the 90s and beyond.
I'm pretty certain that companies in Europe could build its own Facebook, Instagram, Whatsapp, their own android fork etc. They don't do it because for a long time Europe has been a US pseudo-vassal.
On the other hand, Microsoft does seem to have used mere mortals to write a driver/gpu independent way to run workloads in azure (Maia/Triton).
Microsoft is truly unstoppable no matter which way this whole AI race goes.
But really, the blame has to be put on the rest of the industry. If you want to kill CUDA, you have to attack it where Nvidia won't defend; make something Open Source and cross-platform. The problem is getting everyone to sit at the same discussion table. Microsoft has a half-dozen accelerator programs in the work and benefits from the ecosystem fracture. AMD is desperate for anyone but themselves to do their work for them. Apple is hedging their bet on piecemeal acceleration while trying their hardest not to fall behind. Google is trying their hardest to distance themselves from hardware and focuses mostly on software. And Nvidia could care less what everyone does, because they're shipping the largest servers out of everyone mentioned.
> Microsoft is truly unstoppable no matter which way this whole AI race goes.
I can forsee one way they get crushed. If OpenAI keeps dealing in risky business with celebrities and intellectual property, Microsoft might be forced to divest one way or another. Without their OpenAI deal, Microsoft is in a much worse position more similar to Google. If everyone else is willing to ante-up on an Open Source CUDA alternative, then Microsoft's window closes to exploit the demand for a solution. Slim chance that everyone buries their hatchets though.
The Microsoft situation is very different however. You've got this behemoth that has basically just decided to take it's magic money tree and use it essentially buy growth through a huge number of complex deals that are designed to circumnavigate anti-trust. Sorry but "We're not buying you, we're just handing your investors a tonne of money, letting you shut down your company and then hiring all your employees" is bullshit. And also "We're investing in you <oh and here are all these terms about how we own X Y Z of your technology and you must buy A B C of our resources" is also bullshit.
You can make a credible argument that Nvidia uniquely earned the success they're seeing with the AI stuff, I don't think you can make that argument with Microsoft.
Now if you can join the dots - Nvidia has a cosy deal with MSFT to give MSFT GPUs and small companies then have to deal with MSFT because they have all the access to compute, then I think that would be a slam dunk. Leverage Microsoft's Cloud business to hoover up startups in AI.
https://web.archive.org/web/20240606160558if_/https://www.ny...
where there is smoke there is fire
All jokes aside, I know that antitrust is not just about monopoly, but if we just focus on this part, it really does not matter if one company became a monopoly fair and square by just being better/smarter/luckier than everyone else, you still probably don't want it to be a monopoly. And making sure it still has competitor is usually beneficial in the long run. A monopoly might act "fair" for a long time (although it would be hard to know since you won't have anything to compare them with), but their power allow them to turn against the consumer whenever they want with little to no consequence. As an example, you can't really complain about an inquiry against NVidia, and then complain about their pricing and/or how hard it is to get some of their card (especially the one meant for training) because they don't care about you if you don't have a good relationship with them (and a lot of $$).
All in all, pretty sure this inquiry will lead nowhere, or just a slap on the wrist, as is usually the case. It is just a signal to those company to be careful and to avoid using their dominant position on one market to push themselves unfairly in another.
-DOJ v Google
-DOJ v Facebook
-DOJ v Apple
-DOJ V Amazon
the bear needs to pick a side or even better... be a bear. be competent. bears are strong and can command a room themselves.
you have to accept responsibility you're a bear now.
And Merrick Garland is about as bi-partisan a choice as you could pick.
There is a reason Lina Khan goes after companies while her republican counterpart wont.
Thank God the supreme court will kneecap all these useless agenda-driven agencies
It is far closer to a bi-partisan choice than you are making it seem.