News Edward Snowden slams Nvidia's RTX 50-series 'F-tier value,' whistleblows on lackluster VRAM capacity

[Peksha]

GDDR6 and GDDR6X

forgot about it, sorry

 

[bit_user]

I'm not familiar enough of the roadblocks which might exist between PAM3/4, but Ada supported PAM2/4 at the same time.

People don't call it PAM2 - its just a binary (i.e. NRZ) signal.

The complexity of supporting both GDDR6X and GDDR7 extends beyond the analog/digital conversions. GDDR7 also uses a different subchannel structure, with 4x 8-bit per chip, instead of 2x 16-bit. The PAM3 signalling also has some interesting implications for the burst structure, as detailed in the TPU link @Peksha shared.

Worst case, you might end up having to nearly double the size of the memory controllers, which already take up a decent amount of die space, as you can see in this post. That's die space they can't devote to CUDA cores or other useful stuff. So, I think they'd rather just support GDDR7 and hopefully the market will drive down prices and increase its availability, before long.

That's just my take. I'm obviously not a chip designer, so maybe I'm wrong about this and we will see some chips supporting both. Certainly, as the width of the DRAM interface shrinks, the overhead of supporting both would become less, so it might be more likely to happen in the lower-end dies.

 

[mac_angel]

im getting tired of people blaming gpu manufacturers for not having "enough" vram when the problem is that game devs are shoving bloatware rt features into games. ALSO vram isnt the only factor that makes a gpu "good"

that isnt exactly true...

View: https://www.youtube.com/watch?v=iGUejCAupSg

20 series was still a massive leap in game performance + they upped the vram on their budget cards on that series.

It was an increase, but nowhere near enough to justify buying a new card. The difference between the 10 and the 20 was minimal. And nothing compared to the 20 to 30, or the 30 to 40. It's the same thing for the 50 series. Yea, minimal increase in performance, but you're paying the extra for the "new and shiny" thing, which happens to be mostly based around "fake frames".
I stand by what I said. I still believe that NVidia purposely gimps their own cards because they have no competition. They don't use any of the chips' full potential so they can release refreshes, as well as continue with "planned obsolescence". The 10 series was a great generation, and NVidia HATED it. Because no one was buying every generation.

 

[V8VENOM]

No it's you misunderstanding the documentation and not understanding kernel mode vs user mode.

No ... the kernel/user mode isn't relevant to my context. You seem to be focused on mapping of physical memory address space, I'm not. I saying you can direct allocation to VRAM with DX12, the OS will allow it.

determining the meta-parameters of an AI model

I think the term is "supervised learning". The problems we ran into was more cycles actually reduced accuracy regardless of how good the data/source was ... we needed to establish a matrix to identify when patterns and relationship (training) was producing negative results.

True, nVidia will not fail, I can only hope they learn.

 

[bit_user]

The problems we ran into was more cycles actually reduced accuracy regardless of how good the data/source was ... we needed to establish a matrix to identify when patterns and relationship (training) was producing negative results.

Sounds like you're talking about overfitting. That happens when the size of your training set is too small for the model.

 

[williamjeremiah]

Meh, give it a few months, and there'll be a 5080ti with at least 20GB of vram, selling at the price of the AIB 5080's, and the 5080 FE will drop by $200. (not that it'll make much difference, 'cause tariffs)

 

[Rob1C]

Samsung said that the new 24GB GDDR7 wouldn't be available until early 2025, so they still have a bit of time.

As for Micron it will be a couple of years:

With a 512 bit bus I know who I'm blaming for not putting more than 32GB on the card; but now we can have 48 (in our dreams).

 

[bit_user]

As for Micron it will be a couple of years:

Thanks for the chart, but you seem to have misread it. Micron clearly shows GDDR7 at 16-24Gb (32 Gbps) landing in 2024-2025. The 24GB+ that comes in 2026 is 36 Gbps.

FWIW, the fastest any of the released Nvidia RTX 5000 GPUs are running their memory is just at 30 Gbps. So, Micron should definitely be an option for them.

With a 512 bit bus I know who I'm blaming for not putting more than 32GB on the card; but now we can have 48 (in our dreams).

I don't get this point. First, Snowden was complaining about the 5080 having only 16 GB. I've only heard a few AI folks complaining about the 5090's 32 GB, not gamers. Second, to increase either of these would've required doubling the GDDR7 chips (or widening the RTX 5080's memory interface), or else waiting for the 24 Gb chips to launch.

 

[iocedmyself]

You presume that's even an option, but Micron is the sole supplier of GDDR6X (which is proprietary - not a JEDEC standard, I might add) and only offers it in 16 Gb capacity.

​

https://www.micron.com/products/memory/graphics-memory/gddr6x​

​

With GDDR7 now shipping, it's very unlikely Micron will introduce any higher-capacity GDDR6X dies, at this point.


Performance usually follows the improvement of whatever is the bottleneck. The 30% improvement suggests that the CUDA cores were the main bottleneck in the RTX 4090. So, that's why gaming only improved by the amount of CUDA cores they added (with an additional penalty for the clock speed decrease).

I think the RTX 5090's memory bandwidth & capacity went up so much purely for AI.


The article said a packaged die probably costs about $350:

​

https://www.tomshardware.com/pc-com...ts-die-could-cost-as-little-as-usd290-to-make​

​

You forgot the GDDR7 memory. More importantly, you're ignoring the engineering costs, which were estimated for Blackwell at $9 Billion, according to one estimate I read. There's also software and support. Drivers, technologies like DLSS, and Nvidia's extensive range of libraries for game developers cost a lot of money to develop and maintain.

After accounting for all their costs, Nvidia needs to make some profit, as do their AIB board partners, distributors, and finally retailers. So, even if the hardware costs like $550 or $600 to produce, I don't consider a MSRP of $2k too surprising. Could they sell it for less? Sure. But, in the absence of real competition at that level, why would they?

If we consider the situation of a large retail chain buying directly from an AIB partner, and each makes a 10% profit, then that puts their price at $1,652.89. If $250 of that is for memory, PCB, cooler, VRM, assembly, etc. then that puts Nvidia's price for the ASIC at about $1400. If Nvidia is making a 75% profit margin on these, which is the company's overall net profit, then it puts their cost at $800, after all engineering, software, R&D, and support are added in and gives them $600 of profit per 3rd party board that's sold at MSRP (which I know is a bit of a fiction, right now). So, they're taking just 30% of that overall $2k board price as profit. If their take went to zero, then the final cost of the board would only drop to $1270.50. So, don't look at that $550 or $600 figure and think It means Nvidia is making like $1400 or $1450 on each AIB card sold! Obviously, they make more on their Founders Edition, but I'm just talking about their MSRP, here.

Morally, do I think they should increase the price of their flagship gaming cards so much faster than inflation? No, I'm not happy with this situation, either. Were it not for the AI boom, I think these GPUs would neither be so expensive nor quite so powerful.


I don't really have much sympathy for people who expect to upgrade their graphics card, every generation. That's neither normal nor cost-efficient.

If someone had a 5-7 year old graphics card, my advice would be either to go for the RTX 5070 Ti or wait for at least the Supers to launch in a year or so. But, it depends a lot on their budget and what their old card is. The 5070 Ti is just what's most interesting to me, personally.


Given your account was registered in 2006, I have to ask if you simply weren't paying attention during the launch of the GTX 500, the GTX 700, or the GTX 900 series. All of those generations used the same node as their predecessor. And, in the case of RTX 2000, they were virtually the same. The fact of Nvidia reusing the microarchitecture of a prior generation also isn't new.

This is not a "sky is falling" moment, like you're trying to paint it. There are better nodes that Nvidia could use, but they're expensive and in high-demand right now. Given their relatively large dies, client GPUs aren't typically on leading-edge nodes. So, this ended up being a very incremental generation. That's how it goes, especially in the wake of RTX 4000 - which was above average in overall generational performance gains, if a bit on the pricey side.

BTW, I actually have a theory that Nvidia renamed some of the lower-end RTX 4000 models shortly before launch, in light of the weak GPU market at the time. That would explain the lackluster generational gains in some of the tiers.


Intel got through their quad-core era, finally. I'm not sure if there's a good analogy, here.

Ok first off GDRR6X works out to about $5 per GB, where GDDR7 is estimated to be $7-$8 per GB....it's negligible price increase compared to the overall cost of the card.

AIB partners have already come out and stated that selling cards for MSRP would be akin to charity. Which is why AIB cards started out listing for 20% - 50% more, and thanks to the manufactured scarcity, prices have now been raised to be between 60-80% over MSRP. "Partners" aren't getting much of a discount to begin with, AND they're competing with Nvidia's in house FE cards....which is why EVGA announced they were bowing out of the GPU market when the 3000 series first launched....5 years ago. Most renowned and respected partner in the industry gave up, and the situation has gotten much, much worse since.

Lets take a look back through history at Nvidia's 80 class cards.
780 2013 $649 = 24% uplift over previous gen
980 2014 $549 = 38% uplift
1080 2016 $649 =51% uplift
2080 2018 $699 =39% uplift
3080 2020 $699 =63% uplift
4080 2022 $1199=49% uplift
5080 2025 $ 999 =15% uplift

Except the 4080 isn't really the comparison point for the 5080 is it? The $999 4080 super is, and that performance uplift is 8.3%

It doesn't really matter how much money they sunk into this "new" architecture when the architecture is effectively bad. 33% increase to cuda cores and 80% increase in memory bandwidth, for less than 30% performance gain on the halo product....AND RT performance has regressed compared to the 4000 series, showing 10-15% losses in some titles.....while requiring 30% more power. It doesn't merit the imaginary 33% price hike, let alone the 60-80% that's actually being forced on consumers because Nvidia is charging AIBs 90% the MSRP of the card.

Also remember that the 4090, was still a cut down die....the full AD102 die had 18432 cuda cores.....meaning the 5090 is only a 15% increase in cuda cores over the previous generation.

You think people should hold onto their gpu's for 5-7 years? Well the 5000 series is basically a slightly tweaked 4000 series, so it's already 3 years old out the gate, at substantially higher pricing with memory configurations that were already dismal FIVE years ago. 8 gigs hasn't been enough VRAM for several years, and 12 gigs is getting to be the minimum even for 1080p. Even more so as DLSS and FG increase VRAM usage....and both those technologies have been viewed as default requirements for the past three years. So how is it feasible to expect people to hold onto cards that were bad value and bad performance when they launched, for seven years?

You can't look at any single issue in a vacuum....it's the combination of everything, and there is no defense.

 

[bit_user]

thanks to the manufactured scarcity, prices have now been raised to be between 60-80% over MSRP.

There's no way this degree of scarcity is working to Nvidia's advantage. They have an incentive to slightly undersupply the market, but if you do the math, there's no way they come out ahead with this ridiculous extent of a shortage.

Lets take a look back through history at Nvidia's 80 class cards.
780 2013 $649 = 24% uplift over previous gen
980 2014 $549 = 38% uplift
1080 2016 $649 =51% uplift
2080 2018 $699 =39% uplift
3080 2020 $699 =63% uplift
4080 2022 $1199=49% uplift
5080 2025 $ 999 =15% uplift

There's your answer. After two generations of above-average gains, they run out of gas. I've said it before: the seeds of dissatisfaction in the RTX 5000's were planted when Nvidia went so big in the RTX 4000 generation. Not only did it make RTX 4000 expensive, which made everyone angry, but it left them less room for the RTX 5000 to improve on it, without advancing to a newer node.

Except the 4080 isn't really the comparison point for the 5080 is it? The $999 4080 super is, and that performance uplift is 8.3%

If you're not going to compare the launch models from one gen to the next, then you should look at all of the supers, not just the one that supports your argument. However, I think the better thing is comparing one generation's launch models to another's.

Also remember that the 4090, was still a cut down die....the full AD102 die had 18432 cuda cores.....meaning the 5090 is only a 15% increase in cuda cores over the previous generation.

Again, that's not a fair comparison. Yield issues meant they couldn't sell fully functioning dies to everyone, so they reserved them for workstation cards. You have to compare what they actually sell, not imaginary products that would necessarily drive different economics.

You think people should hold onto their gpu's for 5-7 years?

No, but it's usually not cost-effective to upgrade every generation. So, I expect most normal people to upgrade every 2-3 generations. So, more like 3-5 years.

Well the 5000 series is basically a slightly tweaked 4000 series, so it's already 3 years old out the gate,

You can say that about several of their previous generations. The product is the thing people are buying. The price, performance, and power are what count - not what's under the hood.

8 gigs hasn't been enough VRAM for several years, and 12 gigs is getting to be the minimum even for 1080p. Even more so as DLSS and FG increase VRAM usage....and both those technologies have been viewed as default requirements for the past three years.

I think the main use cases for DLSS and FG are using a higher resolution monitor, which is also going to demand more memory. Certainly, when you upgrade your monitor, it's reasonable to upgrade graphics cards.

You can't look at any single issue in a vacuum....it's the combination of everything, and there is no defense.

This isn't about what I think Nvidia should have done. I never defended the decision to go with 16 GB in the RTX 5080. I just pointed out that they sort of brought this issue on themselves, by launching it without 24 Gb GDDR7 for whatever reason. I'm just explaining their decision space. It seems to me that you're so upset by their decisions that you're not looking at it from their perspective and trying to think about the constraints they're dealing with.

 

[thestryker]

There's your answer. After two generations of above-average gains, they run out of gas. I've said it before: the seeds of dissatisfaction in the RTX 5000's were planted when Nvidia went so big in the RTX 4000 generation. Not only did it make RTX 4000 expensive, which made everyone angry, but it left them less room for the RTX 5000 to improve on it, without advancing to a newer node.

We can call it what it really is here: nvidia maintaining high margins. They didn't run out of gas so much as there's no market reason for them to either sell GPUs for less or make them bigger at the same price. There's no technical reason for them not putting another 10-15% cores at each tier below the 5090.

I do certainly agree the problem is entirely of nvidia's own making due to the decisions on both pricing and die size with the 40 series. Of course at the same time for that performance level what else is anyone going to buy? AMD won't have anything that can compete this generation and the soonest Intel could possibly would be a big Celestial GPU which is undoubtedly over a year out still. Even if Intel did release something that could compete in 2026 nvidia could just bump the VRAM to 24GB and raise the power limit to 400W to gain another 10-15% performance over the current 5080 without changing the die at all.

 

[bit_user]

We can call it what it really is here: nvidia maintaining high margins. They didn't run out of gas so much as there's no market reason for them to either sell GPUs for less or make them bigger at the same price.

My comment was targeted just at the generational gains, not pricing. Nvidia really can't make the RTX 5090 much bigger. That pretty much trickles down the lineup.

There's no technical reason for them not putting another 10-15% cores at each tier below the 5090.

One thing I've noticed about Nvidia is that they're pretty keen on maintaining similar perf/$, within a generation. The 5080's MSRP is exactly half of the 5090's and so is the number of its CUDA cores. In fact the die size of the GB103 is actually a bit more than half as big as the GB102.

So, the RTX 5080 gets roughly half the cores and half the memory of the RTX 5090, but slightly higher core & memory clocks. All at half the MSRP. And I expect that it will tend to sell closer to that MSRP, as well. After the various launch issues get sorted out, that is.

 

[thestryker]

One thing I've noticed about Nvidia is that they're pretty keen on maintaining similar perf/$, within a generation.

Yes and no, because the scaling never used to be set by the halo tier product but rather whatever was below it. This price scaling off the halo tier product started with the 40 series and they're just carrying it forward because it means higher margins.

 

[bit_user]

Yes and no, because the scaling never used to be set by the halo tier product but rather whatever was below it This price scaling off the halo tier product started with the 40 series and they're just carrying it forward because it means higher margins.

In the RTX 4000 series, the RTX 4070 Ti and RTX 4090 both had a better perf/$ ratio than the RTX 4080 (i.e. according to list prices). To remain in line with the other two, the RTX 4080 should've cost $1050, not $1200.

 

[thestryker]

In the RTX 4000 series, the RTX 4070 Ti and RTX 4090 both had a better perf/$ ratio than the RTX 4080 (i.e. according to list prices). To remain in line with the other two, the RTX 4080 should've cost $1050, not $1200.

Curious what data you're looking at which doesn't display close to linear performance scaling between those when using the 4090 as baseline. I think it was the 4070 which was slightly better than the rest of the 40 series (when compared to itself not necessarily 30 series). If you're just looking at FPS/$ that should always improve as you go down the stack.

1080p
4080 is about 89% of a 4090
4070 Ti is about 79% of a 4090

1440p
4080 is about 85% of a 4090
4070 Ti is about 70% of a 4090

4k
4080 is about 79% of a 4090
4070 Ti is about 62% of a 4090

 

[bit_user]

Curious what data you're looking at which doesn't display close to linear performance scaling between those when using the 4090 as baseline.

FLOPS/$. Your data is heavily skewed by games with low occupancy. Too much of the 4090's capacity is going to waste. Maybe try with RT and maxed out settings, if you're going to use averages. Either that, or use raster games that achieve high occupancy on the 4090.

 

[thestryker]

FLOPS/$. Your data is heavily skewed by games with low occupancy. Too much of the 4090's capacity is going to waste. Maybe try with RT and maxed out settings, if you're going to use averages. Either that, or use raster games that achieve high occupancy on the 4090.

Then you're using a conditionally accurate metric which doesn't apply to general usage. That's certainly a valid way to measure, but not what I'd use.

 

[bit_user]

Then you're using a conditionally accurate metric which doesn't apply to general usage. That's certainly a valid way to measure, but not what I'd use.

It works for AI, which is largely driving their pricing and I think usually achieves good occupancy.

 

[V8VENOM]

There's no way this degree of scarcity is working to Nvidia's advantage. They have an incentive to slightly undersupply the market, but if you do the math, there's no way they come out ahead with this ridiculous extent of a shortage.

The advantage is beating AMD 9070 XT to market ... nVidia were worried. Getting even a few loud customers to tell the world how great the 5090 is would stop people from buying AMD alternatives. Go visit the nVidia forums, you'll find a slew of pissed of potential customers saying ONLY the 5090 will do and no one else can compete. nVidia should be worried, a 9070XT that is on par or a little better than a 5080 for less than $900.

What I'm confused about is why are all the AMD 7900 XTX and 7900 XT also out of stock? And why are the AMD 9800X3D CPUs all out of stock? Don't really understand this ... (this happened long before Tariffs).

One thing I've noticed about Nvidia is that they're pretty keen on maintaining similar perf/$,

Not sure about that, go back to 2014 and the Titan Z at $3000 ... nVidia were testing the consumer waters. That's when they tasted the lure of huge profit ... 4-way SLI with two Titan Z ... remember SLI, buying 2 GPUs at $1200 each for $2400. But, SLI had too many issues and people stopped buying two or more GPUs so nVidia needed to come up with ways to recover the drop in revenue by increasing single GPU prices.

When a partner card cost $2900 and nVidia FE card is $2000 and they both perform within 3% of each other ... you see the problem? The ONLY way to keep partners viable or interested is to limit supply so consumers are forced to buy from anyone that has one in stock ... no one would spend $900 or more on a GPU to get an extra 3% performance ... they would spend it to get the card now.

I have a suspicion AMD have something up their sleeve for their next gen GPUs ... no one announces they're getting out of the top tier GPU market ... it just doesn't make business sense to "tell the world" ... unless ... you want to mislead your competitors.

 

[Ogotai]

When a partner card cost $2900 and nVidia FE card is $2000

too bad both prices are out of reach for alot of people, specially considering those are probably US prices...
a 590 starts at $3265 and tops out at $4200 here... the 5080 ? $1700 to $2250.. both are out of reach by most of those i know...

IF amd were to release a card that nipping at the heels of a 5080 across the board, even in RT in some games and it was priced here for less then $1000, id guess, they would be getting the amd option....

 

[bit_user]

Not sure about that, go back to 2014 and the Titan Z at $3000 ... nVidia were testing the consumer waters.

Oh, the Titan cards were on their own planet. They never offered a comparable value to the cards below.

When a partner card cost $2900 and nVidia FE card is $2000 and they both perform within 3% of each other ... you see the problem?

Except I doubt the partner cards would be selling for that much, if not for the scarcity. Their cost structure is the same as Nvidia's. The only difference is that they need to tack their own profit margins on the card and it's not nearly that much.

Just wait and see where their price settles out, if/when the scarcity is resolved. That's pretty much the only way we can know for sure. That said, ASUS GPU prices are often quite crazy. I have long wonder who buys their GPUs and why. So, I'd say you're already biasing the argument, by using ASUS as your example.

 

[V8VENOM]

So, I'd say you're already biasing the argument, by using ASUS as your example.

But I wasn't using ASUS as my example? I didn't even mention a specific partner? They all seem to be doing the price increase over MSRP, not just ASUS. But that's sorta my point, those partner cards would NOT be selling for that much over nVidia MSRP if it weren't for nVidia keeping supply short on purpose. And yes, supply will be held short for many months ... current suggestions by retailers like BestBuy is that 5090 supply may never be sufficient for the entire 2025 and into 2026.

For a retailer to suggest, means it is absolutely supply manipulation by nVidia so as to give partners the ability to make more profit. Of course that profit isn't realized quickly but does it matter for yearly financials? Does that matter for partners? Higher profit margins but slower revenue over time ... the business model seems to favor this approach.

If you don't think this is nVidia manipulation, then I'll need to see real proof that nVidia is struggling with making these 5090 GPUs? Not from nVidia's control media outlets, but from actual real data (data that nVidia know isn't available to the public). But more importantly, how nVidia suddenly "don't struggle with supply" when demand shifts? In 2020 Q3 alone, nVidia was shipping over 8 Million GPUs ... in 2025 they could only provide about 1000 world wide?

This was all part off nVidia's plan to manipulate profit margins and I'm sure partners were onboard from day one post EVGA departure. There is no solid real reason to stop making 4090's entirely ... completely different production line that has/had zero impact in what could or couldn't be manufactured for 5090s.

If nVidia lost all it's partners who followed the same route as EVGA, it would not be an "advantage" for nVidia. EVGA left because nVidia didn't cut them sufficient margin to do a better card that demonstrated real results (more than 3% difference).

What is surprising, is that folks like Steve at Gamer Nexus haven't dug into this manipulative business model. I guess technically the business model doesn't violate any US laws and with the current administration, nVidia is unlikely to be challenged by congress. Consumers should consider funding/sending lobbyist to congress and start the ball rolling ... it'll be a heck of a lot cheaper than spending $5000-$7000 for a GPU and/or waiting for nVidia to decide when/if to increase supply. Probably at least 1-2M potential customers, so $10 each you have funded a $20M legal team to take nVidia on.

If we "the customers" do nothing, this situation is only going to keep getting worse and worse.

 

[thestryker]

But I wasn't using ASUS as my example? I didn't even mention a specific partner? They all seem to be doing the price increase over MSRP, not just ASUS. But that's sorta my point, those partner cards would NOT be selling for that much over nVidia MSRP if it weren't for nVidia keeping supply short on purpose. And yes, supply will be held short for many months ... current suggestions by retailers like BestBuy is that 5090 supply may never be sufficient for the entire 2025 and into 2026.

AIBs rarely had more than one 4090 SKU at MSRP throughout the lifetime of the card (if that one was even available) so it really doesn't have much to do with supply. It has a lot more to do with the margins nvidia is making and since they effectively control the video card market AIBs don't have a choice but to go with it.

This is Asus official pricing information (I've seen no other AIB give any confirmations of pricing): https://videocardz.com/newz/asus-re...x-5090-5080-up-to-3099-for-5090-rog-astral-lc

 

[V8VENOM]

Doesn’t really matter, if an AIB card is $900 more than FE card and only produces 3% gain (if any), the ONLY way people will buy an AIB card is if FE supply is limited. This is how AIB with the help of nVidia are increasing margin’s.

If you had a choice between getting an AIB card for $2900 now, or wait 12 months for an FE card for $2000 … what would you do? Supply is being “used/artifically created” to manage margins … either way nVidia make their money (thru customers or AIB, doesn’t matter to them).