Nvidia is allegedly moving production of RTX 4090 GPUs to Hopper H100 GPUs

midian182

Posts: 9,765   +121
Staff member
Why it matters: It shouldn't come as too much of a surprise to see that, despite the card's very high price, most retailers' RTX 4090 inventories have been depleted, indicating we're in another situation where demand is far outweighing supply. But if a new report is accurate, there may be even fewer AD102 GPUs available as Nvidia is allocating production over to the more profitable Hopper-based H100 enterprise GPUs.

A quick look on Newegg for the RTX 4090 brings nothing but the "Out of Stock" notices that were a depressingly familiar sight during the height of the graphics card crisis. The $1,499 MSRP is a high one, but the sheer raw performance of the current Lovelace flagship has made it an appealing prospect to those who want the best of the best.

According to MyDrivers, the RTX 4090 scarcity could be exacerbated by Nvidia instructing the Taiwan Semiconductor Manufacturing Co (TSMC) to shift some production from the AD102 GPU, which powers the RTX 4090, over to the H100 compute processors. They both use the same 4N node, so switching production shouldn't be too much of a problem for TSMC.

Why would Nvidia be making such a move? As with so many things in life, much of it probably comes down to money. The RTX 4090 might have high margins, but it can't compete against the H100 GPU; the $30,000+ SXM variant features 16,896 FP32 CUDA cores, 528 Tensor cores, and 80GB of HBM3 memory connected using a 5,120-bit bus. H100-based products sell for tens of thousands of dollars, and companies often buy them in bulk for server farms and exascale computers, which means a healthier bottom line for Nvidia.

Nvidia probably needs the extra money. It had an abysmal second quarter due to the drop in gaming revenue and lower consumer demand stemming from skyrocketing inflation. It also faces a hefty bill from the "unlaunching" of the RTX 4080 12GB, especially as team green is reportedly footing most of the bill from AIB partners who must rebox and rebrand the card.

But the most significant financial hit Nvidia might be about to take could come courtesy of the US government. A few weeks ago, US officials instructed Nvidia and AMD to stop selling their high-performance AI-focused GPUs to China, which is expected to cost the former company up to $400 million.

As with all unverified reports, take this one with a healthy grain of salt. But Nvidia might believe shifting production of the AD102 to be the best course of action, especially with the AD103-based RTX 4080 16GB and the rebranded AD104-based 'RTX 4080 12GB' cards on their way to help meet some of the demand.

Thanks, Tom's Hardware

Permalink to story.

 
Love it!

So for the Supreme Leader Jensen, US$1700 is simply not a good enough offering from his peasants cult base, so he wants more!

And the rabid ones keeps giving him time and money...
 
Love it!

So for the Supreme Leader Jensen, US$1700 is simply not a good enough offering from his peasants cult base, so he wants more!

And the rabid ones keeps giving him time and money...

Gaming GPUs isnt where Nvidia makes their money, enterprise hardware and solutions are their primary sales and revenue paths.
Yeah, they're gonna take advantage of that when they can flip those units for tens of thousands of dollars to other businesses instead of waiting to sell 10+ GPUs for the same margin per unit.
Making money is the point of being a business.
 
The problem with the claim is TSMC's N4 node, and associated production facilities, aren't exclusive to Nvidia and the lead time for full die production is in the order of months, not weeks or days. So if Nvidia has made this decision now, it won't have any effect on the H100 block. Moreover, suddenly changing a production schedule will incur financial penalties - the fab agreement will require TSMC to fill those orders, no matter what.

The price of the H100 may be substantially higher than the RTX 4090, but the volume of sales is significantly lower; they're only sold to partner companies, such as HP, or in their own data/AI servers. Add-in board vendors will have placed chip orders a while back and while the production time is far shorter than it is to make and package the die, they will have clause in their contracts to ensure there is some kind of fiscal benefit if chip production falls.

I suspect the decision, if true, happened months ago and the current shortage of RTX 4090s is partly due to this. That and the fact that halo GPUs tend to sell out immediately, anyway.
 
I don't think that this is a big deal because the RTX 4090 was never supposed to be produced in huge quantities anyway. The number of people rich or dumb (or any combination of the two) enough to spend $1,600 on a frickin' video card isn't exactly huge and nVidia would rather sell out and lose a few sales than have a bunch sitting around gathering dust.
 
They're making a mythical creature out of this card. Good commercial I guess for the upcoming GPUs.
 
Gaming GPUs isnt where Nvidia makes their money, enterprise hardware and solutions are their primary sales and revenue paths.
Yeah, they're gonna take advantage of that when they can flip those units for tens of thousands of dollars to other businesses instead of waiting to sell 10+ GPUs for the same margin per unit.
Making money is the point of being a business.

Ahem:

https://investor.nvidia.com/financial-info/quarterly-results/default.aspx

Per their end of year 2022 results, gaming made up 15,868 of their $26914 (millions) in revenue. In their most recent quarter, of GPU sales, $6.7B total, 2.04B was gaming and 3.81 was data center. The quarter before, it was 8.29B, 3.75B for data center and 3.62B for gaming.

Data center is bigger, but lets not pretend that GeForce is an afterthought. Gaming GPUs make up a significant portion of Nvidia's total revenue. All this info is publicly available.
 
Ahem:

https://investor.nvidia.com/financial-info/quarterly-results/default.aspx

Per their end of year 2022 results, gaming made up 15,868 of their $26914 (millions) in revenue. In their most recent quarter, of GPU sales, $6.7B total, 2.04B was gaming and 3.81 was data center. The quarter before, it was 8.29B, 3.75B for data center and 3.62B for gaming.

Data center is bigger, but lets not pretend that GeForce is an afterthought. Gaming GPUs make up a significant portion of Nvidia's total revenue. All this info is publicly available.

Well, guess about all I can say is I suppose it serves me right for not taking the time to verify numbers.
 
I'm not really sure why this would surprise anyone. AD 102 is not meant to be a mainstream GPU, it exists only to make you covet it. It exists to part fools from their money and make those that can't afford it feel like you are getting a deal on that mid-range 40 series card that is coming next year for $500+. It serves one purpose, make the whole RTX 40 stack premium priced. In order to do that Nvidia can't be faced with a scenario where they have to drop the price in year one. It must remain expensive so that the whole stack can remain expensive. Therefore, it must be scarce. You were never really meant to own this GPU; it is meant for people with more dollars than sense.
 
I'm not really sure why this would surprise anyone. AD 102 is not meant to be a mainstream GPU, it exists only to make you covet it. It exists to part fools from their money and make those that can't afford it feel like you are getting a deal on that mid-range 40 series card that is coming next year for $500+. It serves one purpose, make the whole RTX 40 stack premium priced. In order to do that Nvidia can't be faced with a scenario where they have to drop the price in year one. It must remain expensive so that the whole stack can remain expensive. Therefore, it must be scarce. You were never really meant to own this GPU; it is meant for people with more dollars than sense.
I disagree. For only a 3k gaming PC I would pick this card (for recommended price of course).
Today, a low end pc is already 1k. Middle end which I might build next year for myself will not go lower than 2k.
The problem is that everything is getting so expensive. People have already accepted since rtx 3000 that the GPUs will be very expensive.
 
I disagree. For only a 3k gaming PC I would pick this card (for recommended price of course).
Today, a low end pc is already 1k. Middle end which I might build next year for myself will not go lower than 2k.
The problem is that everything is getting so expensive. People have already accepted since rtx 3000 that the GPUs will be very expensive.
You can easily build a high-end gaming PC for $1500.00 and even a little less if you know what matters and what doesn't. Anything above $2000 is enthusiast level. A 4090 is an enthusiast level GPU and not meant for the average PC gamer. The 4080 used to be high-end, but Nvidia seems to want to push it into the enthusiast level class or raise the price on the high-end class. We will see if that actually works. My guess is that RTX 4080 will price drop quickly, $1200 is too high and I would not be surprised at all if the RX 7800 XT is $750.00 and poses a serious challenge to that card.
 
Ahem:

https://investor.nvidia.com/financial-info/quarterly-results/default.aspx

Per their end of year 2022 results, gaming made up 15,868 of their $26914 (millions) in revenue. In their most recent quarter, of GPU sales, $6.7B total, 2.04B was gaming and 3.81 was data center. The quarter before, it was 8.29B, 3.75B for data center and 3.62B for gaming.

Data center is bigger, but lets not pretend that GeForce is an afterthought. Gaming GPUs make up a significant portion of Nvidia's total revenue. All this info is publicly available.
It's not an afterthought, but it has declined significantly. Just a couple short years ago, it was just under 50% of quarterly revenue. That's down to around 30% today, whereas Data Center has doubled in 2 years.

And, remember, revenue is not profit. While I suspect Nvidia made some good profits off GPUs over the past couple of years of crypto-mining, that is ending. Now cards are expensive to make and $1500 cards are not really in budget for most of us gamers.

With so much work being put into enterprise AI and machine learning, I can see where they might not want to focus on gaming any more. I'm sure they don't want to walk away from consumer grade GPUs but they may stick with mid-range and high-end and say good bye to the low end. Let Intel have that.
 
Back