NVIDIA RTX 5090: Limited Availability and Restrictions on AI and Multi-GPU

[-]

PizzaCatAm@reddit

This is what monopoly activities look like.

[-]

fallingdowndizzyvr@reddit

You must be referring to the monopolistic activities by the US government. Since Nvidia has no choice in the matter. This is mandated by the government.

[-]

fallingdowndizzyvr@reddit

A government is the definition of a monopoly.

[-]

No is not, what the fuck are you talking about. That’s a laughably oversimplified take. Governments manage societies, not sell products for profit. Calling them a monopoly is like saying referees have a “monopoly on calling plays”, it’s missing the point entirely.

[-]

fallingdowndizzyvr@reddit

I see what the problem is. You don't know what monopoly means. Here. Read. Learn.

"Exclusive control by one group of the means of producing or selling a commodity or service."

https://www.wordnik.com/words/monopoly

It's good to learn a new word everyday.

What does a government do? Provide services. Exclusively.

At the risk of overtaxing your ability in one day, read this.

"

When the government allows or creates a monopoly within a market, that is in essence a government monopoly. The government is either directly or indirectly the only provider of a necessary service or product and other competition is not allowed."

https://bizfluent.com/about-5471337-government-monopoly.html

[-]

PizzaCatAm@reddit

Ah, yes, the scholarly might of wordnik.com and bizfluent.com! Truly the gold standard of economic insight, rivaling the works of Adam Smith and Milton Friedman. Bravo, my friend. I bow before your Google-fu.

Now, let me join you in the learning process. Here’s a lesson: a “monopoly,” as defined by actual economists, is not simply “exclusive control,” but an entity that maximizes profits by restricting competition and manipulating prices. Government, by contrast, operates to regulate markets, provide public goods, and ensure equitable access—things private monopolies would never bother with, as they don’t add to the bottom line. (See: Paul Samuelson’s Economics for some light reading.)

Governments provide services because no private entity would, or could, handle them universally without turning society into an auction house for basic needs. Equating public governance to a private monopoly is like arguing a fire department has a “hose monopoly” because they don’t let Amazon bid for your burning house.

But please, continue citing bizfluent as your economic Rosetta Stone. I’ll just wait here with Joseph Stiglitz and Friedrich Hayek, sipping tea and chuckling quietly.

[-]

fallingdowndizzyvr@reddit

Ah, yes, the scholarly might of wordnik.com and bizfluent.com!

Ah... yeah. As opposed to what? Reddit rando PizzaCatAm?

Here’s a lesson: a “monopoly,” as defined by actual economists,

I know you think of yourself as that, but as I just proved you aren't. So do you have any "actual economists" that back up your pizza cat claim?

(Maybe look into Paul Samuelson’s Economics for some light reading).

Then it should be easy for you to link to quotes, shouldn't it?

Governments provide services because no private entity would

Just another example of your misinformation. Governments often provide services that private entities would love to provide, except the government doesn't allow them to by maintaining their monopoly. The US postal service is a prime example of that. It's a government run monopoly.

But please, continue citing bizfluent as your economic Rosetta Stone. I’ll just wait here with Joseph Stiglitz and Friedrich Hayek, sipping tea and chuckling quietly.

Says a Reddit rando who quotes no one.

[-]

C0rn3j@reddit

This is how sanctions look like.

[-]

PizzaCatAm@reddit

Where is the solution for me, an American living in the US? That’s is neither a Trojan to their cloud or cost a kidney?

[-]

C0rn3j@reddit

Where is the solution for me, an American living in the US?

Don't buy hardware in China which will have the gimped versions - sounds like an easy one for you.

[-]

PizzaCatAm@reddit

Note that the original post didn’t mention this was the Chinese SKU, you could have said that as others.

[-]

C0rn3j@reddit

You could have read the article.

[-]

PizzaCatAm@reddit

Fuck off

[-]

SwordsAndElectrons@reddit

Where is the solution for me, an American living in the US?

Amazon? Best Buy? Newegg? Microcenter? Every retailer operating in the US, really.

If the way this card is neutered bothers you then I'm not sure why you're shopping in China.

That’s is neither a Trojan to their cloud or cost a kidney?

If what you're really trying to say is that the 5090 is too expensive, I agree with your opinion on that. It has nothing to do with this though.

[-]

Air-Glum@reddit

The solution, as others are pointing out, is simply not to buy this card which is exclusively for East Asia (China) markets and won't even be marketed to you in the US.

This isn't the main 5090, it's a region specific card that they make to handle AI tariffs. It won't affect you.

[-]

ExtremeHeat@reddit

Sure, like it won't increase demand for the 5090s.

[-]

Air-Glum@reddit

It certainly won't. At least not in the majority of the world's markets. Nobody is going to want these cards, and they certainly won't want the regular 5090 MORE just because these exist.

[-]

ArtyfacialIntelagent@reddit

This is what loose unsubstantiated rumors on the internet look like.

[-]

PizzaCatAm@reddit

I hope you are right and I end up looking like a moron 🤞

[-]

theshoutingman@reddit

I have good news.

[-]

PizzaCatAm@reddit

I’m relieved hahaha

[-]

Air-Glum@reddit

Can we maybe not cry monopoly every time a business is successful due to genuine innovation?

There are genuine monopoly powers out there and they suck. But Nvidia isn't being grossly anti-competitive about AI. They lucked into the fact that GPUs are incredibly powerful for a specific task and that ended up having wide-reaching effects, but it wasn't their GOAL 10 years ago, and they aren't restricting others from being in the space. AMD doesn't make cards that are as powerful, and nobody else has stepped up.

It was the same thing with crypto. Nvidia made the most powerful GPUs and people figured out how to use that power for something outside of gaming. It wasn't Nvidia's goal, and it sure wasn't anything that couldn't have been prevented if other companies stepped up.

I'm not saying Nvidia can do no wrong or is awesome, but so far it's been a few years and they hit the jackpot on a revolutionary technology. They're milking that for all it's worth, as any business owner would, but for the part they arent there because of anti-competitive practices or bullying others out of the space. They're just the best at it right now. And there's a huge difference between being a leader in a brand new field because you are the best product in the field, and being a giant monopoly that has stymied innovation and held everything back.

At least give them a few years to truly show their dark colors before we start calling them monopoly monster. Otherwise it comes off disingenuous.

[-]

SwordsAndElectrons@reddit

it wasn't their GOAL

That's just silly. I suppose CUDA just popped into being with no investment of development resources?

I agree with most of what you say about Nvidia not really engaging in abusive monopolistic behavior.

However, they didn't just luck into this situation by having the highest performing gaming card, nor was trying to have the highest performing gaming card why they designed their hardware to have the compute performance that it does or why they invested in a superior general-purpose compute API.

Now I'm no antitrust expert and have no idea whether their proprietary API meets any legal standard for being anticompetitive, but lets not pretend that everything out there would run on just any ol' GPU. CUDA forms a moat that prevents many workloads from being able to do that, which keeps their price premium higher than performance alone would allow. People that would consider using a competitor need to balance it against not just the raw performance difference, but also the investment in redeveloping and/or compatibility with their toolchains.

I don't think Nvidia is some evil empire in all this, but you paint them as far to passive. They didn't have a development focus on compute for the last couple decades just for funsies.

[-]

Air-Glum@reddit

I apologize, I wasn't trying to imply it was pure luck or that they were passive. CUDA is critical for where they're at and they put a ton of work and risk into it. My point is that CUDA has been around for a LONG time, and nobody else has really stepped up to meet it in capabilities.

Weirdly, the closest thing to Nvidia from a hobbyist standpoint for running AI and LLMs locally on a decent budget is Apple, who also poured a bunch of research and money into developing their M series chips, which are awful convenient for running AI stuff. A bunch of other people are now trying for similar designs, like AMD's AI chip announced at CES.

My point is that being the first to a field and being a leader in it is not, in itself, monopolistic. Sometimes you make an invention and it takes other people time to catch up. Doing anticompetitive stuff to prevent other people from catching up is obviously shitty, but tech inertia is a BIG thing. It took Apple a long time to shift away from Intel and x86 even when they decided they were going to do it, and it paid off. Nvidia has poured a bunch of time and resources into getting themselves where they are. They didn't KNOW that AI was going to be what it was, but they made smart calls that put them in a position to be the leaders in it, and they invested heavily in AI research once it started getting off the ground.

Other businesses can catch up, but it's going to take them time and investment as well. AMD can't immediately shift gears due to that inertia, but that isn't Nvidia inherently being a monopoly.

[-]

Charming_Squirrel_13@reddit

I'd add that Nvidia also took a big risk with CUDA almost 20 years ago. There were real concerns that they were spending so much money developing products no one was asking for. Even after CUDA's release, not a lot of developers took them up on the offer. They played the long game and it worked out really well for them.

[-]

Holyragumuffin@reddit

Innovation and monopoly can be simultaneously true.

[-]

m3kw@reddit

This is what protecting gamers that wants it for gaming, otherwise they’d be all be in data centres

[-]

PizzaCatAm@reddit

So where is the non gaming solution that is not a Trojan horse for their cloud services?

[-]

adel_b@reddit

maybe they are trying to protect the main customers, the gamers... this would be awesome if true, I don't recall other companies doing it

[-]

Fluboxer@reddit

[-]

adel_b@reddit

the subject is rtx

[-]

norbertus@reddit

NVIDIA's main customers are not gamers, not even close.

Gaming made NVIDIA about $3 billion last year, compared to $26 billion for data centers.

Even if they're expensive, the gaming cards are a bargain compared to server hardware with similar specs.

But the gaming hardware is nerfed because NVIDIA doesn't want to risk losing business on the margins for their real bread and butter: servers.

[-]

davew111@reddit

Main customers? nVidia didn't get a 3.3 trillion market cap because of gamers.

[-]

frivolousfidget@reddit

Main customers? Have you seen the size of datacenter orders?

[-]

NotRandomseer@reddit

For the rtx cards. That's the consumer line after all

[-]

adel_b@reddit

for pc card graphic?

[-]

Neex@reddit

This is a government imposed sanction dude.

[-]

Accomplished_Mode170@reddit

Seems specific to the 5090D sku from the article; boycott worthy if true of other SKUs

What we know, according to the leaked frequencies of an RTX 5090D that has been seen in the Asian country, is that the card can reach up to 3 GHz, but it will lower its frequency to 2.6 GHz in its GPU (Boost + ASIC, equals that of the RTX 4090), while its memory will drop dramatically to 14 Gbps to give a bandwidth of 896.1 GB/s.

[-]

moldyjellybean@reddit

They’ve been doing this for a long time. 12 years ago I could pass through any AMD GPU to my home lab GPU using vt-d or iommu in a hypervisor.

Nvidia blocked that for their consumer GPU SKU. They did this with their GPUs during the mining boom. Getting tired of this anti consumer bullshit. I pay you money to buy it, I own it I should be able to run it as I please.

[-]

David_Delaune@reddit

Nvidia blocked that for their consumer GPU SKU.

When was the last time you tried this? I just recently setup Hyper-V with gpu passthrough (Windows Server 2022) and it was working just fine. There are instructions over on MSDN to get it setup.

I ended up disabling it because the host operating system loses access to the gpu when you assign it to a hyper-v guest.

[-]

moldyjellybean@reddit

I know they changed their stance for this sometime around 2022? But it was pretty anti consumer for a long time when I was doing it with esx 4.0 and AMD probably 10+ years ago.

[-]

AmericanNewt8@reddit

For 4090 they nixed a fuse so fp16 tensor accumulate would be limited.

[-]

Psychopompe@reddit

What's this story?

[-]

DeltaSqueezer@reddit

Yup, they nerfed it a bit for AI. I'm pretty sure they'll nerf the 5090 in some way to distance it from non-consumer product lines.

[-]

Psychopompe@reddit

Do you have a link? I failed to find more details about that.

[-]

PizzaCatAm@reddit

Awesome, I hope this is the case.

[-]

SuperChewbacca@reddit

I agree, the article was confusing, especially because it was translated. Are they just talking about the 5090D?

Does all the existing open source training, and inference software work the same and get the additional performance for the regular 5090?

[-]

ItsAMeUsernamio@reddit

They were showing Flux benchmarks on stage so it should.

4090 also had a 4090D model in China with similar restrictions. Went to a computer market in Hong Kong and all the stores had those on display and no 4090s.

[-]

RedditDiedLongAgo@reddit

the Asian country

🇹🇼

[-]

frivolousfidget@reddit

If so it is more likely like this to comply with regulations on AI.

[-]

jms4607@reddit

My plan is hope 4090 price drops, then implement multi-4090 p2p with hotz modified driver

[-]

nasenbohrer@reddit

p2p?

[-]

cr0wburn@reddit

The 5090D not the normal 5090, right?

[-]

-PANORAMIX-@reddit

Yep, misleading post

[-]

nvidiot@reddit

Yeah, that's a China-exclusive GPU to work around USA's current embargo on GPUs with certain level of AI computation level. It's got a whole bunch of restrictions to specifically gimp its AI performance so it can be sold in China.

[-]

Inevitable_Fan8194@reddit

Well, that's… petty. Especially since the main effect will probably be the emergence of a Chinese competitor for Nvidia. What do they think? That they are going to say "oh no, too bad, we won't do AI, then"?

[-]

cyberdork@reddit

China has no technology which could make chips beyond what's used in the RTX 30xx generations. Even the technology used to make their best chips needs to be bought in, since they don't have any own advanced DUV tool maker.
If they want to create a competitor to Nvidia they first need to create competitors for multiple of the leading edge semicon tool makers. One day they might be able to make chips on the level of the 40xx series, but by that time Nvidia is already at the 150xx series.

[-]

nicolas_06@reddit

Nvidia has no choice this...

[-]

Inevitable_Fan8194@reddit

Oh yeah, I'm not blaming them. When I said "what do they think?", I was referring to lawmakers.

[-]

ASYMT0TIC@reddit

There is a point to this. No one on earth can touch than TSMC at chips right now, and I believe Samsung are the closest ones in second place. Both of them are US allies. China is still a few years behind, and as a result their AI chips can't be as power efficient. The US has been holding on to this card for just the right time to use it, and the time to use it is during the critical point in arms race toward the greatest super weapon the world has ever known.

Of course they know that this will only add fire to China's efforts to reach parity with TSMC, and that they will get there eventually. But right now, the only concern is getting to AGI faster than the adversary, as even if the winner gets there only half a year sooner it might as well be a century depending on how it all plays out.

[-]

MizantropaMiskretulo@reddit

Another thing to note, even with the most efficient GPUs, large data centers require immense power.

China can spin up new nuclear power plants much faster and cheaper than anywhere else in the world...

[-]

DifficultyFit1895@reddit

I’m surprised people are not talking about this more in terms of the hardware. The current technology and all near-term prospects of improved technology are incredibly energy inefficient. We have to imagine some breakthrough will occur to make the processors able to do more with less energy. We know it’s physically possible because we have over 8 billion examples here running on about 20 watts.

[-]

nicolas_06@reddit

I don't believe 6 month make a difference and AGI will not come from hardware but software and design. The engineers and researchers will make many changes, like hundred or thousands with a few being critical that will lead to AGI eventually.

Also depending AGI may be many years from today and for all we know by then, they may manage to design better hardware and decide to restrict its export too.

[-]

Blankaccount111@reddit

AGI will not come from hardware

LLM's only exist because of hardware advancements. They have been around in theories and papers since the 1970's but the hardware was not available to make it feasible.

[-]

nicolas_06@reddit

The hardware legal in China is not much slower. It the same order of magnitude perf not 100X slower. At that game it doesn't matter much.

But difference in software architecture like transformer really change the game.

[-]

Aischylos@reddit

It spawned out of hardware, but even with worse hardware, China is still pumping out models like deepseek and QwQ.

[-]

ASYMT0TIC@reddit

We don't know, and that's why it's important. Maybe it won't make a difference, or maybe an ASI tasked with gaming global propaganda and influence could collapse a foreign regime in half a year.

[-]

ElectronSpiderwort@reddit

I'm curious about "the greatest superweapon" statement. How is AGI/ASI that, rather than a doomsday scenario for all of us? Heck, we're not using the brains we currently have to process obvious facts and well regarded conclusions; how is an AGI going to be a strategic win?

[-]

ASYMT0TIC@reddit

Brains are lazy by design because they are energy constrained in nature. ASI isn't. You answered your own question.

BTW it can be both a superweapon and a doomsday scenario.

[-]

Charming_Squirrel_13@reddit

The secretary of commerce said exactly this. the goal at this point is to put up roadblocks to slow down China's AI progress while we continue to extend our lead. If NVDA ends up with new competitors, I think that would simply be seen as the price to pay in this emerging Cold War.

A similar technique was used against the Soviet Union. You force your adversary to devote resources toward catching up, rather than them using those resources to leap frog you.

As an NVDA investor, I'm annoyed by these restrictions, but I'm not surprised in the slightest that there's a line in the sand drawn here by the us government.

[-]

TakuyaTeng@reddit

They don't think and it's a problem. They'll stifle any form of innovation or competition on behalf of those that have the money to bribe them. They'll do it to the detriment of the US but they don't care because they can line their pockets.

[-]

sedition666@reddit

It is a nice sentiment but China has been trying to make CPU competitors for years now and are nowhere close. And the machines to even make the most advanced chips are all western made.

[-]

SwordsAndElectrons@reddit

Especially since the main effect will probably be the emergence of a Chinese competitor for Nvidia. What do they think? That they are going to say "oh no, too bad, we won't do AI, then"?

Sure, but developing such a competitor will slow them down, so this is better than doing nothing in the eyes of those that believe something should be done.

Regardless of whather I agree, if it results in a new competitor in the GPU space, or even just the AI space, and that causes the price of Nvidia cards to come back down from the stratosphere... I really can't say that bothers me much.

[-]

Inevitable_Fan8194@reddit

Sure, but developing such a competitor will slow them down

Yeah, that was my thought as well, but that's kind of backward too if this was they're idea. If the recent history of Chinese industry taught us anything, it's that they don't really care to be the first on the market. Ultimately, they win the markets because of low cost.

[-]

chase_yolo@reddit

It's the short term shareholder benefit. On to the next one after a competitor arrives

[-]

UGH-ThatsAJackdaw@reddit

When you find a chip fab capable of turning out 2nm devices with a decent yield, lmk. TSMC seems unlikely to shelf all their existing agreements with Nvidia and AMD to start making copies of those same chips reverse-engineered by the State.

[-]

sludgybeast@reddit

Maybe im dumb but why would AI laws in the US affect what is allowed to be sold internally in China?

[-]

Big-Profit-1612@reddit

Export controls.

[-]

Mickenfox@reddit

Because the US government can punish Nvidia based on what they do in China.

[-]

beryugyo619@reddit

Well if it crosses Taiwan strait it's not "internally" sold

[-]

TheDailySpank@reddit

NVidia is a US based company.

[-]

toadbike@reddit

Exporters get to decide what they export.

[-]

SithLordRising@reddit

Well that sucks

[-]

pc_g33k@reddit

Even for the non-D version?

[-]

wickedsoloist@reddit

Then i’m not buying. Lol. Look at the low iq company. We are the ones paying their salaries and they are gatekeeping the product? The full power? Fuck them. Fucking NVIDIA. They think themselves are important and best company in the world. Which they not.

[-]

foxgirlmoon@reddit

You are not the ones paying Nvidia's salaries lmao.

You are but a single drop in the bucket.

Nvidia does not care about the average consumer, not anymore. They get their monies from selling big server farm devices, not consumer stuff.

[-]

Hunting-Succcubus@reddit

but nvidia should not forget their roots.

[-]

wickedsoloist@reddit

https://s201.q4cdn.com/141608511/files/doc_financials/2024/Q4FY24/Rev_by_Mkt_Qtrly_Trend_Q424.pdf

You say so? Datacenter revenues are not going to perform like this forever. Datacenters are not buying those chips to replace every 2-3 years.

[-]

Big-Profit-1612@reddit

Datacenter hardware is EOL in 5 years.

[-]

ortegaalfredo@reddit

>Then i’m not buying. Lol.

This is exactly what they want. This is a product for gamers. They want you to buy their professional line.

[-]

wickedsoloist@reddit

I'm not going to pay tens of thousands dollars to their fking RTX 6000 or H100 chips. Also I'm not interested with AI training or crypto mining. I just hate to be limited. I pay the product so I own it. I can do whatever I want with it. Company cant decide what i can do and what i cant.

[-]