US Progressives Push For Nvidia Antitrust Investigation (reuters.com) 42
Progressive groups and Senator Elizabeth Warren are urging the Department of Justice to investigate Nvidia for potential antitrust violations due to its dominant position in the AI chip market. The groups criticize Nvidia's bundling of software and hardware, claiming it stifles innovation and locks in customers. Reuters reports: Demand Progress and nine other groups wrote a letter (PDF) this week, opens new tab urging Department of Justice antitrust chief Jonathan Kanter to probe business practices at Nvidia, whose market value hit $3 trillion this summer on demand for chips able to run the complex models behind generative AI. The groups, which oppose monopolies and promote government oversight of tech companies, among other issues, took aim at Nvidia's bundling of software and hardware, a practice that French antitrust enforcers have flagged as they prepare to bring charges.
"This aggressively proprietary approach, which is strongly contrary to industry norms about collaboration and interoperability, acts to lock in customers and stifles innovation," the groups wrote. Nvidia has roughly 80% of the AI chip market, including the custom AI processors made by cloud computing companies like Google, Microsoft and Amazon.com. The chips made by the cloud giants are not available for sale themselves but typically rented through each platform. A spokesperson for Nvidia said: "Regulators need not be concerned, as we scrupulously adhere to all laws and ensure that NVIDIA is openly available in every cloud and on-prem for every enterprise. We'll continue to support aspiring innovators in every industry and market and are happy to provide any information regulators need."
"This aggressively proprietary approach, which is strongly contrary to industry norms about collaboration and interoperability, acts to lock in customers and stifles innovation," the groups wrote. Nvidia has roughly 80% of the AI chip market, including the custom AI processors made by cloud computing companies like Google, Microsoft and Amazon.com. The chips made by the cloud giants are not available for sale themselves but typically rented through each platform. A spokesperson for Nvidia said: "Regulators need not be concerned, as we scrupulously adhere to all laws and ensure that NVIDIA is openly available in every cloud and on-prem for every enterprise. We'll continue to support aspiring innovators in every industry and market and are happy to provide any information regulators need."
Wut (Score:5, Insightful)
Wut, how is having CUDA antitrust?
They make a gpu, they make a platform to use that gpu. Other makers scoffed at it before it was too late and somehow thats Nvidias fault?
Bundling hardware and software? Huh? So they can sell a gpu but not the drivers/api for it too?
Instead of complaining, these other companies need to just freaking innovate.
Re: (Score:3)
I feel like I understand a lot of things about computing, but for the life of me I dont understand how Intel (or other chipmakers) haven't been able to just essentially duplicate what NVidea has done with memory sharing/structure complete with their own CUDA like API that is "plug and play" for the most part.
I know thats not "easy" but seemingly well within their
Re: (Score:2)
... for the life of me I dont understand how Intel (or other chipmakers) haven't been able to just essentially duplicate what NVidea has done with memory sharing/structure ....
Patents.
Re: (Score:2)
CUDA is friggin' huge, and its optimizations are focused around the design of NVidia hardware.
NVidia just has a big head start. But others certainly could leapfrog it with more innovative approaches - ASICs, NPUs, ternary processors, even neuromorphic processors.
It's not a system of the big guy suppressing competition; it's just a system of the big guy just having a head start.
Re: (Score:2)
100% agreed. One has to manually download the CUDA Toolkit [nvidia.com] so it isn't like Nvidia is even bundling it.
AMD and Intel failed to provide a competing option with OpenCL or even a CUDA "emulator" / cross-compiler.
What has AMD and Intel done in the 17 years?? (CUDA 1.0 was released in 2007)
Re:So for starters (Score:4, Informative)
I worked for Nvidia for 13 years until I retired, 20+ years early thanks to the stock price. So I'm biased.
The employ all sorts of nasty techniques to lock down the engineers who can actually build these kind of systems. Some of it's just higher pay which is fine but there is plenty of other weird little things that are done behind the scenes. If you've ever read about how video game companies used to prevent their employees from being poached it's like that but worse
Yeah, Nvidia does lock in engineers. They pay very well. They give out stock as RSUs. They have great benefits. There have subsided meals, and good meals at that. There is no RTO requirement - just get your work done and they don't care where you are. They treat engineers as valuable assets, not fungible heads who can be replaced at will. A coworker was in a road accident and out for over a year - he received full pay and benefits while out and when he returned his job was waiting.
It's hard to get (and keep) a job there because Nvidia expects their engineers to produce quality work. If you can't do that you will be out after some chances to improve. The turnover rate in engineering is actually quite low, lower than other hardware developers like intel, AMD, Google or Apple. Mistakes are expected and tolerated (that's how you learn), not learning or repeating the same mistakes is not tolerated for long.
Nvidia believes in intellectual honesty. It's part of the culture. Decisions are based on factual evidence, not on personal aspirations or who is in charge, This fosters receptiveness to new ideas, no matter who generates them, and creates enhancements across the whole company. It's a company of 30k people but it has the excitement and fluidity of a startup. If that is lock-in then every company should do it.
And this is before we talk about things like using their market position to provide millions in free programming and consultancy services in exchange for, again, you guessed it preferential treatment of their products. In this case from software developers.
AMD, Intel and Nvidia all provide free programming advice and sometimes services for game companies. The aim is to make certain the game works as designed on each company's hardware. All card manufacturers want the game companies to be happy which means helping them as necessary. Is this millions in free programming? No it isn't. It's a visit from one or two application engineers for a short period, who will explain and demonstrate how to get the best from the hardware features the company is trying to use. If a company wants additional help, they can get it if they pay for it, and may have to put a mention of Nvidia in the product. This is true of all card manufacturers.
It doesn't help that Intel is almost certainly going to drop out of the market now that they have cut basically everything that isn't essential.
This comment is risible. Intel is fully engaged in the AI race. In fact, I believe that in a couple of years, when they have better hardware, they will be in a better position than AMD. Both Intel and Nvidia understand the extreme importance of software in AI, AMD is floundering.
Re: (Score:2, Insightful)
In a just world this would be DOA: there is ample competition in every segment of NVidia's market. They're not as good, but having poor competitors doesn't make you a "trust."
But this isn't a just world. NVidia has a lot of money. Warren and her gang want some. So here we go: tee up the lobbyists and campaign bux and the no-show non-profit gigs and all the rest.
Re:Wut (Score:5, Interesting)
> Wut, how is having CUDA antitrust?
CUDA is not the issue. When you buy up all the tier 1 and tier 2 companies making software for your competitors, then it becomes an issue.
> Bundling hardware and software? Huh? So they can sell a gpu but not the drivers/api for it too?
This isn't about CUDA or Drivers. It's about how they stifled competition by reducing choices.
> Instead of complaining, these other companies need to just freaking innovate.
This isn't other companies complaining. This is your government, which spends $B of dollars buying hardware increasingly from NVIDIA with our taxes. I pay a good 45% of my money to the government, and I'd prefer a little more of it go to teaching vs hardware.
But let me tell you some relevant stories from _my_ life... Some details may be altered to protect the innocent.
I was in the High Performance Computing (HPC) industry for a long, long time before I returned to my roots of doing graphics and user experience. Someone I know worked at a small tech company whose product offering included a Mellanox-based InfiniBand switch to sell with their clusters at a much cheaper price with a much slicker design that cut off all the unnecessary features -- saving you maybe $4K per switch and you could fit 2 switches in the space of Mellanox's 1U design - they modified the reference and removed every unnecessary feature to make a switch take up less than 1/2 the space. I thought it was a super clever design. After NVIDIA bought Mellanox, they folded that product. It was a shame, too, because that product alone could have shaved off a few percent off the prices of clusters sold to research institutions that our taxes funded. It doesn't matter to a Trillion dollar company, but to a 100 person company, that could be the whole company's viability. The way the story was told to me, their boss was given a choice: You can get discounts on the GPUs or the switches, but not both. Sounds like an illegal monopoly move to me. I always thought that my friend should make a separate business doing the switch thing, but even though he worked on the product, it wasn't his IP. He was smart and moved on to work at some big company.
In another story, NVIDIA was on the Khronos board responsible for overseeing OpenCL. OpenCL (TM Apple) was an API that enabled you to run compute loads on NVIDIA, Apple, AMD, IBM, FPGA hardware, and other platforms. NVIDIA was a big founder of this, and pushed out a very cool OpenCL 1.0 implementation that I spent 3 months optimizing a transform into a runtime for Microsoft's .NET platform and writing all kinds of tests - some contract work for some company. By OpenCL 1.1, Apple chose to move off of NVIDIA GPUs for its MacBook hardware to AMD GPUs, and NVIDIA decided to stop the engineering effort on OpenCL and put all those engineers into CUDA. Within 1 GPU generation, the OpenCL support was garbage - like only supporting a 286 instruction set when a 386 was available. Within 2 generations, it was effectively dead. That makes sense from a business perspective - products have life cycles. The company I did that work for was making an IDE for OpenCL. It was a cool product - you could run it on Mac OS, Windows, and Linux - tested against a variety of CPUs and GPUs. When NVIDIA bought up a specific compiler optimization company who had made HPC optimizations for both CUDA and OpenCL, the demand for that product went away -- almost overnight. NVIDIA knew what they were doing. They were making it harder for professional and education customers to support AMD hardware by buying out the top compiler makers and stopping manufacture of existing product lines for competing hardware. Maybe they should have had more scrutiny on this purchase from the anti-trust committee.
I did another contract job for a company doing GPU optimized H.264 encoding via some proprietary code and a newly released proprietary NVIDIA beta API. I filled in some gaps in an OpenCL and CUDA optimized algorithm and ported the interface to use NVIDIA's new encoder... Which worked great until you tried to dynamically change bitrate mid-stream to adjust for low bandwidth considerations. The bitrate wasn't dynamic as it was promised. That capability was advertised but not allowed for reasons I was told in confidence at a religious event maybe 5 years later - meeting someone high up at NVIDIA completely by accident who knew the rest of the story. Is it monopoly worthy? I can't tell you, but it was a good story about the incompetence of large companies like NVIDIA and how they can destroy opportunities when they get too big and too powerful.
People in engineering talk - especially once you get a beer or two in them. There are lots of meetups in silicon valley that I went to and got good gossip. These are _some_ of the stories relevant to this news article that I've been told. If you want to hear stories of NVIDIA being a monopoly, just invite a bunch of senior engineers and pay for a round at a High Performance Computing conference - Super Computing is a good one.
Re: (Score:2)
I don't understand why people downvoted my thoughtful reply to this thread and labeled me a troll, but I'm done with this community after 20+ years.
Re: (Score:2)
I get the same thing from time to time. While Slashdot continues to have one of the better communities for insightful discussion and moderation, the only culture of discussion has definitely shifted over the year to, at a default, use moderation to silence things you don't like.
Within the tech enthusiast sub-culture, the dominant thought is that both Nvidia and Intel are evil and if you attempt to temper that opinion with literal facts, you should expect to be downvoted, troll-modded, etc. It's not that EVE
Re: (Score:2)
Re: (Score:3)
I'm not sure I agree with the basis of this lawsuit... but I'll say my first concern is that nVidia recently stated/reiterated that the ROCm/CUDA shim violates the CUDA license. That is that there exists an open-source github project that lets CUDA be used on AMD GPUs [don't get your hopes up, ROCm still sucks]. and nVidia decided that it was too threatening. btw, this is probably uneforceable given that reverse-engineering for compatibility is legal under US law. but EULAs hardly care.
https://www.phoronix. [phoronix.com]
Re: (Score:2)
Fully agreed. And the software and hardware are "bundled" so tightly because that's needed for performance, not because of some nefarious plot. High performance inference systems in effect compile the model into a version optimized for the specific available hardware.
Re: (Score:1)
Wut, how is having CUDA antitrust?
Nvidia tries to make it illegal to reverse engineer CUDA, they put some language about it in their EULA. But the DMCA explicitly protects reverse engineering for the purpose of interoperability, so that clause is unenforceable. It constitutes a threat to sue for legally protected behavior which must be engaged in for purposes of competition, so it is anticompetitive.
Instead of complaining, these other companies need to just freaking innovate.
Instead of making legal threats, Nvidia needs to be familiar with the law and not create EULAs which are illegal to enforce.
Re: (Score:2)
Why spend money on R&D when you can spend it on lobbying.
https://www.opensecrets.org/federal-lobbying/clients/summary?id=D000000804
https://www.opensecrets.org/orgs/nvidia-corp/summary?id=D000036303
microsoft (Score:2)
Wut? how is having WINDOWS antitrust?
They make computer hardware, they make a platform to use that os. Other makers scoffed at it before it was too late and somehow that's Microsoft's fault?
Bundling hardware and software? Huh? so they can sell a gpu but not the drivers/api for it too?
Instead of complaining, these other companies need to just freaking innovate.
--
Sometimes the fool writes his own retort for you because it's moronic to forget about Microsoft.
This is not about GPUs, is about datacenter chips (Score:5, Informative)
This is not about nVIDIA's punny 40x0 Vs AMD's xt50x0 series vs Intel's Alchemist 770 gaming video cards....
This is about big boy datacenter cards, problem is, the fine reporters at reuters, Ars and other outlets call them GPUs.
This is about nVIDIA's Hopper and Blackwell N1x0 & H1x0 Datacenter FP chips, used not only for AI TRAINING but also for CFD (used among other things for F1 cars and and oil detection), and Finite Element analysis, against similar chips like AMD's MI300 and Intel's Havana and Gaudi chips (only for AI).
These chips have 64GB of HMB RAM (against the punny 8~16GB of modern gaming video cards), connected by high speed interconnects.
The fisrt problem is that nVIDIA is manufacturing constrained. They beg and plead to TSMC to give them more capacity, because they have a huge backlog, but alas, not feasible. And, there is shortage of HMB memory, so....
But, this supply shortage (not nVIDIA's fault) gives nVIDIA an oportunity to do nasty thing. For one, they seem to de-prioritize consumers that go for alternatives (Intel, AMD, or their own custom chips), they own Mellanox (one of the biggest network chip designers), so, the letter allegues, nVIDIA forces customers to use thi\eir networking solutions.
If true, these allegations are clear usage of a near monopoly position to gain unfair market advantage...
Anywho, I personaly think that this is just posturing by politicians who do not understand the intricacies of the technology at hand.
JM2C YMMV
Re: (Score:3)
The fisrt problem is that nVIDIA is manufacturing constrained. They beg and plead to TSMC to give them more capacity, because they have a huge backlog, but alas, not feasible. And, there is shortage of HMB memory, so....
The key supply constant isn't really wafers or even HBM. The key bottleneck is packaging, which is why TSMC spent a lot of money to beef up their CoWoS capabilities.
Re: (Score:2, Informative)
problem is, the fine reporters at reuters, Ars and other outlets call them GPUs.
So does NVIDIA: "NVIDIA Data Center GPU Resource Center" https://resources.nvidia.com/l... [nvidia.com]
"NVIDIA H200 Tensor Core GPU " https://www.nvidia.com/en-us/d... [nvidia.com]
Accusations of liberal overreach (Score:2)
It’s just so obvious that Nvidia is in a good position right now because they just happened to have the right product, at the right time, exqctly when the AI craze hit like a freight train. Every other chip manufacturer is coming out with a competing product. And, there’s some evidence that this craze will just burn itself out as it follows it’s natural course.
No monopoly to see
Re:Accusations of liberal overreach (Score:4, Insightful)
No monopoly to see here.
There absolutely is a de facto monopoly here. It's up to the system to determine whether they're abusing their position, but they're absolutely in that position.
Re: (Score:3)
This is not a monopoly, monopolies are artificially constructed and the government can then move to disassemble the artificial restrictions. In this case the restrictions are true supply capacity constraints and demand is greatly exceeding the available supply. The solution here is not for the government to go mess with nNVIDA, instead the government should do what it can to increase supply. At least the US has finally seen the light and has stopped driving chip makers offshore.
This similar to the problem w
Re: (Score:3, Informative)
At best, you can claim that it's not a monopoly until after the investigati
Re: (Score:2)
China mining rare earths and the US letting ours sit in the ground is not a monopoly even though China has 90% of the market. They are not artificially constraining trade. Now if China went around the world and systematically bought up all of the rare earth deposits and prevented anyone else from mining -- that would be antitrust.
nVIDIA has a very high market share due to competitor's screw ups not due to restraining trade. SpaceX is in a similar position, Space is not restraining anyone else from building
Re:Accusations of liberal overreach (Score:5, Informative)
High market share does not define a monopoly in the USA, what creates the monopoly is when the high share company interferes with the competition. Europe is different, in Europe just having a high share can get you declared a monopoly even if you aren't interfering with the competition.
That isn't correct. The US absolutely will define a monopoly based on high market share. However, the US only really cares if you use that dominant position in an anti-competitive manner. So, it usually only comes up when there are investigation into that behavior.
Re: (Score:2)
Re: (Score:2)
Nvidia was just in the right place in the right time. They're like the only car repair ship in a small town that suddenly discovered a gold deposit close by. For a small period of time, they're the only game in town and gonna be raking it in. But the problem will fix itself - more mechanics will move in and set up shop.
Re: (Score:3)
There is nothing about the definition of "monopoly" that requires it to be "artificially constructed". It literally comes from "monos" (single) + "polein" (sell) = "Single seller". Monopoly status is wherever a single supplier dominates a market. Doesn't matter a whit how they got there.
NVidia is a monopoly. Full stop. Whether it's a monopoly that is abusing its monopoly status is the issue up for debate. Most people here (myself included) seem to think "no".
Re: (Score:2)
From the article:
"But Nvidia's certainly got a big lead that could be difficult to disrupt, progressive groups fear, by manufacturing the most in-demand chips and offering superior networking technology to train AI."
This sounds suspiciously like a call to handicap the winner because "they're too good!"
Consider the fact that Nvidia has been investing their own money doing research and building products to promote and advance AI research. They've arguably built a market for a suite of products that work well
Interesting group of “progressives” /s (Score:2)
Interesting group of “progressives”:
Blue Future
Demand Progress Education Fund
Economic Security Project
Fair Vote UK
Institute for Local Self-Reliance
NextGen Competition
Open Markets Institute
P Street
Revolving Door Project
The Tech Oversight Project
Elizabeth Warren against progress - AGAIN (Score:4, Insightful)
NVIDIA spent over a decade developing their proprietary drivers and CUDA to take advantage of the similarly developed hardware.
Their hardware could not work without software, and "proprietary" means they developed it for themselves, not for the free and open
source community (FOSS) or Elizabeth Warren's failure to understand how the economy works. They made a widget. They made software to run the widget. Nobody is forced to buy ANY widget, THAT widget, or THEIR widget.
It's not unlawful "bundling" nor "anticompetitive" nor "cornering the market". It's a widget making company being successful.
This being the United States of America where I stand, you'd think elected representatives would grok that instead of doing more performative stupid shit.
But then, there's life.
Cui Bono? (Score:2)
Start digging into the background of Demand Progress and you'll find some very shady people and funding.
"industry norms"?? (Score:2)
took aim at Nvidia's bundling of software and hardware, a practice that French antitrust enforcers have flagged as they prepare to bring charges.
"This aggressively proprietary approach, which is strongly contrary to industry norms about collaboration and interoperability, acts to lock in customers and stifles innovation,
Huh. I wonder what definition they're using for "industry". Bundling hardware and software is long-running industry practice. It's both a business approach and a technical approach (allow
warren submitted put options (Score:1)
kill success (Score:1)
According to the left, apparently, any successful corporation is ipso facto a criminal.