Nvidia EULA limits GeForce Data Center Usage *fork*

It would seem that since this is a restriction on the use of the drivers in a datacenter, it is similar to software licensing being different for enterprise versus home users. Ownership of the physical cards wouldn't be controlled, but a datacenter would be stuck with original drivers or maybe some kind of open-source driver package.

It's Nvidia's role as a copyright holder over the software that is being used, as in other situations a copyright holder has leeway in saying how their work can or cannot be used before authorizing a copy is legitimately held.
 
It's Nvidia's role as a copyright holder over the software that is being used, as in other situations a copyright holder has leeway in saying how their work can or cannot be used before authorizing a copy is legitimately held.
As a layperson I would say there would be some antitrust angle involved if a (hugely) dominating market player suddenly dictates you can only use their most expensive cards for professional use. That's not how copyright is supposed to be used, nor is it how a free market is supposed to work.
 
As a layperson I would say there would be some antitrust angle involved if a (hugely) dominating market player suddenly dictates you can only use their most expensive cards for professional use. That's not how copyright is supposed to be used, nor is it how a free market is supposed to work.
I'm not a legal expert, but antitrust and copyright laws would tend to work in opposite directions here.

The former is focused controlling or undoing the the actions or existence of a dominant or monopoly player, while the latter is concerned with the assignment of a monopoly with extensive authority.
There are certain exemptions to that authority, but the default case is highly favorable to the copyright holder's authority.

A somewhat gray area for individual users at least is that onerous EULA terms for a product and software in a retail box is that the EULA cannot be read or agreed to prior to purchase and opening the box.
There's less forgiveness in a professional or enterprise capacity, where the assumption shifts to the idea that the buyer is more sophisticated. If the software is itself being used as part of a commercial service like a datacenter, Nvidia could also refuse on the grounds that it does not want its copyrighted work contributing to that service or take on any assumption of warranty or support.
Even the seemingly hypocritical or arbitrary exemption of blockchain usage is part of the power that copyright allows.

The limits of the authority seem to be included in the story, where Nvidia is asserting its right to make conditions on what can be done with a reproduction of its work, and for copies created after the conditions changed.

Maybe Sakura Internet could get away with using original drivers, if the cards were bought in a retail configuration, although the obsolete code would probably make their service undesirable.

I'm not sure, but maybe they could theoretically roll out with code based on software that is more openly licensed, or reverse-engineer something. That would put the onus on Sakura Internet to make the investment in undermining Nvidia's product segmentation, and somehow provide for clients the risk management, support, and quality assurance Nvidia provides in its permitted use cases.
 
Cars have warranty for miles or years whichever fills first. If gpus were cars datacenter use case would be out of warranty fast due to hours.

Not where I live.
Car will have 3 years warranty or 60000 miles. 3 years has no limit on mileage, but if your car is 5 years old and still under 60k miles then it still is under warranty.
This year I bought Peugeot for business use and I'm planning to do 80k miles in 3 years, I will be covered from basic warranty.
 
It's Nvidia's role as a copyright holder over the software that is being used, as in other situations a copyright holder has leeway in saying how their work can or cannot be used before authorizing a copy is legitimately held.
Tenuous ground. When you buy the product, it needs to work. Part of the product is the firmware. If the firmware needs an update, the original purchaser ought to be entitled to that update, or a refund of the hardware that didn't work properly when they bought it. Selling someone a defective product and then trying to use copyright on a fix to wrest control doesn't seem a legitimate business practice.
 
Tenuous ground. When you buy the product, it needs to work. Part of the product is the firmware. If the firmware needs an update, the original purchaser ought to be entitled to that update, or a refund of the hardware that didn't work properly when they bought it. Selling someone a defective product and then trying to use copyright on a fix to wrest control doesn't seem a legitimate business practice.

But you need to consider they are differentiated by models not just firmware/software, meaning what Nvidia is doing is not tenuous or illegal; case in point each revision of CUDA has evolved Compute version/SM_xx.
In fact if we go down the route of your argument you might as well say Nvidia is breaking moral or at worst tenuous/illegal ground with the way they structured FP16 and Int8 between the various Pascal models and not all models are entitled to same functionality but while there is criticism (including from myself to some of that as I felt GV102 should had probably had full FP16 support but disabled in 1080ti) it was not illegal or much of a grey area.
However this decision regarding FP16 is more than just about margins because the top Tesla GV102 is actually more expensive than the 16GB GP100 PCIe.
The above has context when you use the word defective product/doesnt seem a legitimate business practice to describe what Nvidia is doing with the drivers.


Geforce is not Tesla or Quadro nor entitled to the same functionality and importantly not the same level of implementation; remember Tesla and Quadro are sold by business service entities that survive building such datacenters and not necessary sold directly by Nvidia in each country.
Allowing Geforce to be used in the way of Tesla/Quadro for datacenters not only affects Nvidia but also their higher level sales/service channel partners.

However like I said more detail is required on what the ToS actually entails with the Geforce change because the Pascal Titan was encouraged to be used for DL also in cluster/nodes before the Tesla GV102 launched; context was probably more along dp2a and dp4a dot product.
But if you look at Sakura and what they do, which kicked this off as DavidGraham mentioned they really were pushing against the established sales/service structure Nvidia created and knew it, tbh Nvidia would have to respond not just to protect the correct products but also their higher sales/service partners.
 
Last edited:
In fact if we go down the route of your argument you might as well say Nvidia is breaking moral or at worst tenuous/illegal ground with the way they structured FP16 and Int8 between the various Pascal models and not all models are entitled to same functionality but while there is criticism (including from myself to some of that as I felt GV102 should had probably had full FP16 support but disabled in 1080ti) it was not illegal or much of a grey area.
You ge tthe product you bought in those cases. In comparison, I recall GPUs where you could unlock the locked-off hardware and get a higher-end part for cheaper. That wasn't illegal and no amount of EULA's could prohibit users from doing whatever they wanted with their hardware. Nowadays we get fuses blown to break the hardware and limit it.

The above has context when you use the word defective product/doesnt seem a legitimate business practice to describe what Nvidia is doing with the drivers.
I don't know what they're doing with the drivers specifically, but these generally include fixes for bugs. If that's not the case and the drivers supplied with the card are faultless, then there's no need to update the FW and nVidia can provide that as an optional feature-enhancing service for those following their T&Cs.

Allowing Geforce to be used in the way of Tesla/Quadro for datacenters not only affects Nvidia but also their higher level sales/service channel partners.
That's really neither here nor there. There's a legal principle of ownership that transcends keeping some businesses in good health. If the marketplace changes and these businesses can't compete, they either about or die. The law shouldn't be tweaked to allow post-sales regulation of how products are used. In this case, nVidia needs to gimp their hardware somehow to make it less desirable versus their Tesla/Quadro lines. If they sell the same product in two packages, one cheaper than the other, what can they expect other than the cheap one being preferred?!
 
You ge tthe product you bought in those cases. In comparison, I recall GPUs where you could unlock the locked-off hardware and get a higher-end part for cheaper. That wasn't illegal and no amount of EULA's could prohibit users from doing whatever they wanted with their hardware. Nowadays we get fuses blown to break the hardware and limit it.

NVIDIA doesn't sell hardware alone, they also provide software to that hardware and support, they can choose whether a certain customer receives their software and support or not according to their conditions. It's the same principle as software applications that have a user license and an enterprise license. If a company purchased a user license to a certain software, the software owner can revoke their license and demand they use the enterprise one. Even though the product is the same in both cases.
 
Last edited:
You ge tthe product you bought in those cases. In comparison, I recall GPUs where you could unlock the locked-off hardware and get a higher-end part for cheaper. That wasn't illegal and no amount of EULA's could prohibit users from doing whatever they wanted with their hardware. Nowadays we get fuses blown to break the hardware and limit it.

I don't know what they're doing with the drivers specifically, but these generally include fixes for bugs. If that's not the case and the drivers supplied with the card are faultless, then there's no need to update the FW and nVidia can provide that as an optional feature-enhancing service for those following their T&Cs.

That's really neither here nor there. There's a legal principle of ownership that transcends keeping some businesses in good health. If the marketplace changes and these businesses can't compete, they either about or die. The law shouldn't be tweaked to allow post-sales regulation of how products are used. In this case, nVidia needs to gimp their hardware somehow to make it less desirable versus their Tesla/Quadro lines. If they sell the same product in two packages, one cheaper than the other, what can they expect other than the cheap one being preferred?!

However a corporate large tech company deliberately avoided and knowingly the sales and service structure created by Nvidia with its partners and there is already a known differentiation between Geforce and Tesla/Quadro that Sakora ignored.
From a legal standpoint Nvidia has differentiated their products; Geforce is for the retail/prosumer market while Tesla and Quadro are for the professional/datacenter/HPC/etc market, but like I said Sakura who should know better deliberately avoided this.
Ownership does not include rights/entitlement to software including firmware or even functionality/spec beyond what is defined for the product, some functions may be opened up by the Manufacturer but it is not guaranteed.
A clear example as I showed was FP16 and Int8, by your argument Nvidia should had been sued but they were not by anyone, or you may want to say Nvidia should had been sued when they enabled the Titan Pascal cards with optimised library access equivalent to Quadro for some libraries and applications as it was not accessible originally and so locked but again they were not because fits my context - this came about because of AMD using Geforce Titan rather than the Quadro card when compared to Vega FE (emphasis FE) for some professional visual applications.
What they are doing is legal and about known product differentiation between Geforce/Quadro/Tesla, and software/libraries/services including firmware are always owned/rights of the Manufacturer but even here Nvidia differentiates and has for quite a long time between Geforce and Quadro/Tesla.
As another example you will not be able to use the Geforce Pascal Titan as a Grid server-client solution that relies upon Quadro-Tesla GPUs, again what they are doing is not illegal nor suspect.

You are aware that the MI25 has HW and some functions missing from Vega FE which was sold as a prosumer GPU and no mention was ever made of the differentiation when Vega FE went on sale, so where does that leave AMD that had even less known differentiations?
And then there is the differentiations done with CPUs between consumer/prosumer/data centers-HPC.
What is missing from a lot of the narrative around this change is that Sakura is a reasonable sized tech company implementing a reasonable size data center solution that should fall under Nvidia's higher sales/service professional channel and correct product lines.
In fact it is questionable how they managed to acquire so many Titans in the 1st place; context is this should be a single project and all orders fall under/reference it.
 
Last edited:
Tenuous ground. When you buy the product, it needs to work. Part of the product is the firmware. If the firmware needs an update, the original purchaser ought to be entitled to that update, or a refund of the hardware that didn't work properly when they bought it. Selling someone a defective product and then trying to use copyright on a fix to wrest control doesn't seem a legitimate business practice.
There's a body of consumer protection law that would generally be more supportive of this position although these days there are signs it is becoming less effective. In part, the consumer is not given the chance to accept or deny EULAs prior to purchase, and because it is recognized that a consumer is unlikely to be versed in legalese and has been primarily guided by advertisement and marketing.
That's for consumers, which Nvidia's EULA change states as being the licensed case for its Geforce drivers. Even then, I think some amount of imperfection is acceptable in the end. There are still driver bugs and I've not seen a mass of successful lawsuits every time a card has issues in a game.

An enterprise that builds data centers, manages multiple employees and large-scale contracts is not an unsophisticated user. It's navigated untold numbers of regulations and contracts already, and in other instances Sakura Internet is likely using the same sorts of complex legal procedures in negotiations to its own benefit.
In a context where there is a professional or enterprise interest, there is more precedent to the idea that there can be more complicated contractual obligations or additional conditions, just like how many software packages have home, education, and professional/enterprise variations with audits or tip lines to catch companies pirating or abusing the license.
Being able to say how copyrighted software can be used is something that has been accepted for some time, and niche or technical tools and software have not had their sometimes heavy restrictions challenged.

If there's old drivers Nvidia probably cannot forbid their use since the license at the time was more permissive. Nvidia could argue that in-house or open-source software is an option, if Sakura Internet wanted to do more than just find a loophole in Nvidia's segmentation on the cheap. Downloading new drivers is asking the copyright owner for permission to receive a copy, and Nvidia is able to add conditions. The right to have a say over how one's work is used is granted to any copyright holder, so it would be an uncommon case to take that away for anyone, not just Nvidia.
Nvidia can argue that the product still works in the use cases it considers valid, and there are any number of products that have tags or provisions indicating they are only for a specific kind of use.

Antitrust could be involved if Nvidia was making moves into Sakura Internet's market, but if it's still selling Quadros and Teslas to whomever would buy them it would be a harder case to make. Nvidia could argue they could switch to AMD, and if Sakura Internet objects with something like "but AMD's software is terrible for our datacenter service", then Nvidia would argue that it sounds like good drivers are something worth paying extra for.

One possible new area where copyright might have less precedent is if Sakura Internet tried to write its own drivers and Nvidia had some kind of code-signing or secure boot measure like AMD's PSP that was made to block the usage, the back and forth on circumventing copyrighted encryption measures used to lock out otherwise normal use is not as clear.
 
You are aware that the MI25 has HW and some functions missing from Vega FE which was sold as a prosumer GPU and no mention was ever made of the differentiation when Vega FE went on sale, so where does that leave AMD that had even less known differentiations?
Well, no, I don't know anything about the hardware differentiations. ;) I'm posting because this thread talked about the EULA that was accompanying consumer cards, and pointing out that that, AFAIK, isn't legally enforceable. At any level, I don't think. If a big company wants to stack their data centre with Teslas, and then has a couple servers short and pops down the local PCWorld and gets a couple of Titans and slaps them in, if they work, how can they be policed, prevented, and/or prosecuted?

What is missing from a lot of the narrative around this change is that Sakura is a reasonable sized tech company implementing a reasonable size data center solution that should fall under Nvidia's higher sales/service professional channel and correct product lines.
These will be nVidia's business arms serving business, but if a company wants to circumvent that and buy consumer level goods, they can. I don't know of any law anywhere that can prevent that (outside of contract law). If you want to produce Hollywood movies using consumer level video editors, ther's nothing to stop to. The reason pros use professional level software at professional prices is because the features aren't there at the consumer level. If you want to shoot a movie on a consumer level camcorder, you can. If you want to write the script in Notepad, you can. There's no method to limit how someone uses a product they own which is what I'm protesting at here, there apparent move to control users post sells, based on what's been posted.

Ordinarily a company would control use of consumer level products at the enterprise level by gimping it and limiting availability. If you need 1000 Titans, you can't bulk buy them from Amazon so would be forced to buy direct from nVidia at which point they are allowed to refuse to sell you the cheaper product.

An enterprise that builds data centers, manages multiple employees and large-scale contracts is not an unsophisticated user. It's navigated untold numbers of regulations and contracts already, and in other instances Sakura Internet is likely using the same sorts of complex legal procedures in negotiations to its own benefit.
Right. Sakura should be dealing with contracts for direct supply from nVidia which would control what they can and can't do via agreement.

Being able to say how copyrighted software can be used is something that has been accepted for some time, and niche or technical tools and software have not had their sometimes heavy restrictions challenged.
Can you provide any example of where functionality or application of a software product is refused? I'm not talking licensing agreements as to who can use a piece of software, but where a piece of software states it cannot be used to create/support XYZ?

To be clear on my argument here, in case I've misunderstood what's happening, I understand that nVidia has included a license agreement with their consumer level cards saying they cannot be used in certain situations, trying to control how a product sold over the counter is used after that transaction has taken place. That's the only aspect I'm questioning the legality of. I can even recall clear evidence that you can't, when I learnt that those multibuy packs of Coke and junk food can be sold individually in stores. Even though the manufacturer labels them as 'nor for resale', they can't enforce that and have no say over what happens to the product once bought.
 
If the features on regular driver is enough for me, nvidia shouldn't be able to prevent me from using it for whatever I like, to me this point is moot from the dawn of GPGPU, I'm the user and I choose the code that runs on the card. Car analogy: Ford can't prevent me from delivery pizza, it sells me a car and that's the end of it.
 
To be clear on my argument here, in case I've misunderstood what's happening, I understand that nVidia has included a license agreement with their consumer level cards saying they cannot be used in certain situations, trying to control how a product sold over the counter is used after that transaction has taken place.
I'm the user and I choose the code that runs on the card. Car analogy: Ford can't prevent me from delivery pizza, it sells me a car and that's the end of it.
Once again, NVIDIA isn't targeting the user here, they are targeting specific enterprises who use consumer cards to build servers and then sell time on it. NVIDIA has every right to demand these enterprises to use professional grade cards instead of consumer grade ones.
The reason pros use professional level software at professional prices is because the features aren't there at the consumer level.
And NVIDIA can gimp consumer drivers to be incompatible with a lot of professional grade wrokloads or at least be less desirable, there is certainly a precedence with that with TitanXP. Obviously they are choosing not do that since it would be inconvenient to a lot of users. So they force their conditions on enterprises directly through software licensing.

Can you provide any example of where functionality or application of a software product is refused? I'm not talking licensing agreements as to who can use a piece of software, but where a piece of software states it cannot be used to create/support XYZ?
One recent example that comes to mind is VMProtect and Denuvo. At first Denuvo purchased a user license from VMProtect and used it in their newest anti tamper implementation. VMProtect caught wind of the act, and demanded Denuvo to use the enterprise license. Denuvo refused, so VMProtect revoked their license completely and then threatened to sue them. And demanded Steam to remove Denuvo games from their library as they are now using an unlicensed piece of software, essentially pirated games now. Denuvo folded in the end and all was good again.
 
Can you provide any example of where functionality or application of a software product is refused? I'm not talking licensing agreements as to who can use a piece of software, but where a piece of software states it cannot be used to create/support XYZ?
There are various levels of permissiveness in open-source, and forms that preclude/allow commercial use of code or can allow for changes not to be contributed back, or where coders can grant permission for their published code to be used but refuse to give it for organizations they disagree with (ex. military purposes). Whether an author (assuming they haven't signed rights to their publisher) wants a given book chain to carry their books, or a screen adaptation made of their book, the ways Star Trek's IP permitted for fan works with limitations, songwriters withholding permission to play their songs for certain events/causes. There are specific exemptions to this authority, possibly due to legacy reasons. In the US, artists can make song covers and with a legislated royalty they can make a cover song regardless of the song owner's permission.

Nvidia's refusing to provide the right to create a copy of their software, and software is treated as a creative work similar to the above.

To be clear on my argument here, in case I've misunderstood what's happening, I understand that nVidia has included a license agreement with their consumer level cards saying they cannot be used in certain situations, trying to control how a product sold over the counter is used after that transaction has taken place.
Nvidia's modified the conditions of their permission for the creation of a copy of their copyrighted drivers. The software license for their non-professional product drivers no longer extends to data centers. If Sakura Internet gets a copy of Nvidia's more recent drivers, and tries to use them outside of the licensing terms, it's copyright infringement since Nvidia is not providing permission for a copy of those drivers to be made in this instance.
 
One recent example that comes to mind is VMProtect and Denuvo. At first Denuvo purchased a user license from VMProtect and used it in their newest anti tamper implementation. VMProtect caught wind of the act, and demanded Denuvo to use the enterprise license. Denuvo refused, so VMProtect revoked their license completely and then threatened to sue them. And demanded Steam to remove Denuvo games from their library as they are now using an unlicensed piece of software, essentially pirated games now. Denuvo folded in the end and all was good again.

It looks like Denuvo repackaged VMProtect and resold it to end-users while Sakura is selling processing time on their own datacenter. Keeping the car analogy: Can Ford prevent me from driving for Uber?
 
And NVIDIA can gimp consumer drivers to be incompatible with a lot of professional grade wrokloads or at least be less desirable, there is certainly a precedence with that with TitanXP. Obviously they are choosing not do that since it would be inconvenient to a lot of users. So they force their conditions on enterprises directly through software licensing.
But I don't think they can. ;) That was the intention, but without contracts in place, they can't say to anyone, "don't use this hardware for this activity."

At first Denuvo purchased a user license from VMProtect and used it in their newest anti tamper implementation.
Uh, that's licensing. I explicitly asked for different examples of refused functionality...#

"Can you provide any example of where functionality or application of a software product is refused? I'm not talking licensing agreements as to who can use a piece of software, but where a piece of software states it cannot be used to create/support XYZ?"​

nVidia are free to license and control drivers (within the realms of their product working, probably), such as saying for consumer cards you have to register your product to be able to download drivers. As such, they can refuse to provide free software to anyone they choose, thus trying to force companies to use enterprise licensing.

AFAICS the issue is that nVidia, being nice to their consumers, released an ungimped high-end product, and in doing so accidentally undermined their enterprise business. Their way to remedy that is to try and control how that product is used after-market, which seems unenforceable to me. Complaining Sakura are in the wrong here is a parallel issue. Yes, they probably are (although free market economics suggests as long as they operate legally, if they can find a cheaper way to operate, that's just good business), but that doesn't make the attempt to control after-market usage right or the right way to go about remedying things. If nVidia's actions are legal and actionable, that paves the way for companies to control how we use their products, and means the potential for small print from everything..."this underwear is not to be used in conjunction with ripped jeans. These headphones are not to be used with music first recorded between the years of 1958 and 1973."
 
There are various levels of permissiveness in open-source, and forms that preclude/allow commercial use of code or can allow for changes not to be contributed back, or where coders can grant permission for their published code to be used but refuse to give it for organizations they disagree with (ex. military purposes).
Again, that's licensing, not function/application. ;) Yes, nVidia are free to refuse to give their software to whomever they want.

Nvidia's modified the conditions of their permission for the creation of a copy of their copyrighted drivers. The software license for their non-professional product drivers no longer extends to data centers. If Sakura Internet gets a copy of Nvidia's more recent drivers, and tries to use them outside of the licensing terms, it's copyright infringement since Nvidia is not providing permission for a copy of those drivers to be made in this instance.
Possibly a test case. Software rights are getting everywhere and aren't being contested enough to result in high-court rulings so everyone knows where they are. Policy makers are probably unsure what they should and shouldn't allow, having all their voting consumers on one side and all their corporate money-hats on the other. :p

I guess unlike a TV or washing machine, the drivers don't come built in but form part of the installation. Sakura would be allowed to buy the hardware, but they'd have to control it themselves because they wouldn't have rights to the software, if the EULA is valid which itself is an ongoing question mark. A ruling in nVidia's favour seems like a difficult decision though. Again, it's opening a can of worms where any time you install or update some software, the company can decide what you can and can't do with your product just by a smallprint entry. This is fine in a service (you can stop using it) but not in a bought product.
 
As long as Geforce cards are sold with CUDA being one of their USPs, I doubt it'll be easy to enforce that EULA right now for data center usage. From what I gather from local law, contracts are not allowed to have "surprising" paragraphs, i.e. stuff you don't expect. I can imagine that we're seeing the build up right now in order to change that generally accepted proposition that you can do with your cards whatever you like and they'll try to enforce it lateron for real.
 
Last edited:
I do not have any issue in general with them saying that Geforce drivers cannot be used in datacenter.
I do have an issue with them changing this after someone has already purchased the product. Could Sakura argue that their existing cards should be permitted DC use, while agreeing that future Geforce purchases cannot be used in a DC?
I do realise that you could argue "newer driver not included as part of package", but the general expectation would be that this kind of change would not happen.
 
It looks like Denuvo repackaged VMProtect and resold it to end-users while Sakura is selling processing time on their own datacenter.
Denuvo purchased the enterprise license from VMProtect. That's how they avoided the storm. And car analogies don't apply here. Again NVIDIA sells hardware + software + support, they license software and support. As such they control who can receive their license and under what circumstances.

Keeping the car analogy: Can Ford prevent me from driving for Uber?
Yes they can, if they have a lineup of cars dedicated to Uber driving and rules to prevent the use of other Ford models for that purpose. The fact that car makers don't have such rules doesn't mean they won't be able to enforce them once they have said rules. In fact they often do enforce their own rules, in 2015 they tried copyrighting car modifications. I don't know whether they succeeded or not. But the premise stands. Everyone wants their product to be categorized and sold according to their own conditions.
 
Back
Top