News Nvidia Prepping RTX 4080 With New Silicon, Report Claims

[watzupken]

I actually don't believe cost is the main driver for the high price of these GPUs. Consider the fact that the AD104 for example, the die size of the chip is about the same as the size of the Navi 31 GCD @ about 300mm2. Yet to consumers, they cost about the same when Nvidia wanted to launch the AD104 as a RTX 4080 12GB. The Navi 31 GPUs are clearly more complex even with the chiplet design, and with higher BOM to produce each card.

 

[bit_user]

Moore's law says 2x transistor density in 18months. I know that is no longer feasible. But transistor density is largely what determines performance and cost per transistor. With each generation the cost of each transistor should go down.

That said, yes wafer cost are going up. However this merely slows down the price decrease per transistor per generation.

What's funny about that is how Eben Upton said in Toms' launch coverage of the Pi 4 that they moved from 40 nm to 28 nm, because that node had the lowest cost per transistor. And, at the time (though this was back in 2019 and a lot has changed), he expected the next Pi would also be made on 28 nm.

Now, perhaps he was accounting for not only the per-wafer costs, but also the various up-front and engineering costs, amortized over the lifetime (or at least the first year's worth) of production they expected to do. Otherwise, something seem very broken.

Looking at transistor count and cost per transistor based on wafer node, the margin Nvidia and amd are pulling are insane compared to what it used to be.

What figures are you using, for that? I think it's probably more telling to look at the gross margins in their quarterly financial reporting, since that captures all of the engineering, support, and marketing costs. Those bills all have to get paid somehow, and these companies' income is nearly 100% goods-based, not services (though Intel looks to be challenging that notion).

Basically the fps/dollar has stagnated.

That's not the fault of the hardware. If you look at GFLOPS/$, there's a strong generational rate of improvement. If fps/$ isn't going up, that says more about the games and what resolutions people use.

Year-Mo	Graphics Card	GFLOPS	MSRP	GFLOPS/$
2013-11​	GTX 780 Ti	5046​	$699​	7.22​
2015-06​	GTX 980 Ti	6054​	$649​	9.33​
2017-03​	GXT 1080 Ti	11340​	$699​	16.22​
2018-09​	RTX 2080 Ti	13448​	$999​	13.46​
2020-09​	RTX 3090	35580​	$1,499​	23.74​
2022-10​	RTX 4090	82600​	$1,599​	51.66​

Okay, so there's a regression in the RTX 2080 Ti, but that's because they only went from 16 nm to 12 nm, and burned most of their additional silicon budget on ray tracing and tensor cores, while simultaneously increasing the price. A lot of people were snapping up GTX 1080 Ti's, at the time, in an open rejection of what a poor value it represented.

BTW, the small jump between the 780 Ti and 980 Ti is largely due to the fact that they were both made on 28 nm. Maybe, if Nvidia could've charged more for it, the 980 Ti would've been made on 20 nm (but, it's also possible there simply was no GPU-suited 20 nm node, at the time).

it was obvious before the release of the cards crypto upheld these high prices. How AMD and NVIDIA didn't see the collapse in demand at these prices is beyond me.

Ever heard of the game musical chairs? It's a little like that. You know the music is going to stop, but you don't know when. And, until it does, you have to keep pace with the others.

Six years back $750 would have gotten you a #1 card. $550 #2

Yeah, it's not fun to see prices increase faster than inflation.

It happened with CPUs, as well, once the core-count race got going. I still lament that you have to buy a lot more cores than you may want or need, just to get a desktop CPU with the highest turbo limits.

 

[bit_user]

I've calculated a number of ROI to justify projects. I know what goes into an engineering project in terms of soft and hard cost and how that amortizes out.

But, "drivers" constitutes an increasingly large portion of the budget, and it's an ongoing cost for the hardware's entire life. That might differ somewhat, from projects you've planned.

I put the terms in quotes, since it encompasses quite a large amount of software activity, ranging from helping partners optimize their games, maintaining Nvidia's own first-party libraries, and optimizing their drivers on the back-end, for new AAA title launches.

And you act as if it is our responsibility to support overly aggressive growth with their staff increases. I got news for you: It's not our responsibility to support a company if they decide to grow too fast.

It's not a responsibility, but it is a reality. Engineering costs money, and maybe they're milking gamers to fund their forays into self-driving and robotics, but that's their call and you can choose to either pay the price or go with an alternative.

A majority of gamers are just getting priced out. Even decent 1080p entry level is inaccessible for most now.

Really? What do you define as "decent 1080p"?

A GTX 1660 Super is going for $270, new. Its specs are almost identical to my GTX 980 Ti. Is that no longer even a 1080p card?

 

[umeng2002_2]

Every press release and product pricing decision over the past 4 years or so has indicated that nVidia is tired of selling silicon to consumers. They would rather just sell data-center products.

Yes, even with them trying to get AIBs out of the market with their FE cards. They slowly ween AIBs out of making GPUs. Then nVidia just keeps increasing the price of FE cards until no one wants them. Then they just stop offering them as gaming products. No AIB companies to suddenly make mad. They still need those AIBs to be OEM manufacturers for their data-center-only products.

There is a war against general computing for consumers/ normal people. If everything moves back into data centers, it can all be controlled easier and charged more money for.

"Don't want to spend $2500 for an RTX 5080? Just rent time on our Geforce Now service. Only $100 a month." Next crypto boom, nVidia has all of the GPUs in their cloud servers. They make all the money mining when gamers aren't actually gaming on the hardware.

nVidia's software is so polished because it's spilled over from their industrial solutions. AMD just makes hardware and lets other companies' software engineers' fight it out to see how to use the things.

 

[bit_user]

Every press release and product pricing decision over the past 4 years or so has indicated that nVidia is tired of selling silicon to consumers. They would rather just sell data-center products.

Some of the new tech in their GPUs doesn't really translate to either cloud gaming or other computing applications. So, that argues they're not trying to get completely out of the consumer graphics card business, or else why would they be investing it in?

Yes, even with them trying to get AIBs out of the market with their FE cards. ... They still need those AIBs to be OEM manufacturers for their data-center-only products.

That makes no sense. If they can make their own FE cards, they certainly don't need AIBs to make data center cards. Anyway, cards are the easy part (relatively speaking).

There is a war against general computing for consumers/ normal people. If everything moves back into data centers, it can all be controlled easier and charged more money for.

5 years ago, you might've had a stronger case. At least, on the CPU end of things. However, we've seen an unprecedented amount of innovation and advancement pour into consumer platforms. In that time, we've gone from 4-thread to 16-thread CPUs being the norm, with up to 32-thread CPUs also coming to the mainstream. DDR5, NVMe, and PCIe 5.0 also have come to the desktop, with CXL seeming to be in the works.

I think the real issue is your sense of entitlement that the industry should make faster and more capacious stuff each year, but that prices should never increase.

At a more abstract level, when thinking about the breakdown of Moore's Law & friends, it makes sense that it wouldn't be a brick wall, but rather a somewhat asymptotic price curve as we push closer and closer to the limits. I think we're definitely seeing that, if you look at wafer pricing.

nVidia's software is so polished because it's spilled over from their industrial solutions. AMD just makes hardware and lets other companies' software engineers' fight it out to see how to use the things.

I don't think there's a lot of truth to that, on either front. For Nvidia's part, AI, cloud, and HPC are all Linux-based. There's probably not much overlap between their Linux and Windows drivers, at that level. The overlap should mainly be in the userspace portions, where the bulk of CUDA, Vulkan, Direct3D, and OpenGL get implemented - all APIs not used for "industrial" solutions. Not only that, but Nvidia makes libraries and tools specifically for game developers that also have no place in "industrial" applications.

Regarding AMD, they do all the same things as Nvidia, software-wise. Maybe their game oriented libraries and tools aren't as good - I'm really in no position to say. Apart from that, they have to implement all the same APIs and features as Nvidia, and even went so far as to create a clone of CUDA which they call HIP.

 

[spongiemaster]

I've calculated a number of ROI to justify projects. I know what goes into an engineering project in terms of soft and hard cost and how that amortizes out.

And you act as if it is our responsibility to support overly aggressive growth with their staff increases. I got news for you: It's not our responsibility to support a company if they decide to grow too fast.

Yet somehow their margins keep growing. Having more engineers does not seem to be hurting them too much.

I'm calling you on it Spongie: Hogwash.

BTW:
980Ti = $650 on launch.
1080Ti = $700 on launch
Each was a #1 rank.

A majority of gamers are just getting priced out. Even decent 1080p entry level is inaccessible for most now.

What does any of this have to do with the cost per transistor being an awful metric to judge video card prices? Just sounds like a sour grapes rant.

 

[digitalgriffin]

But, "drivers" constitutes an increasingly large portion of the budget, and it's an ongoing cost for the hardware's entire life. That might differ somewhat, from projects you've planned.

I put the terms in quotes, since it encompasses quite a large amount of software activity, ranging from helping partners optimize their games, maintaining Nvidia's own first-party libraries, and optimizing their drivers on the back-end, for new AAA title launches.


It's not a responsibility, but it is a reality. Engineering costs money, and maybe they're milking gamers to fund their forays into self-driving and robotics, but that's their call and you can choose to either pay the price or go with an alternative.


Really? What do you define as "decent 1080p"?

A GTX 1660 Super is going for $270, new. Its specs are almost identical to my GTX 980 Ti. Is that no longer even a 1080p card?

1660 super is the equivalent to a 960 from a few years ago in terms of aging. It was fine when it was released. But try playing a cyberpunk, flight sim, or the new Harry Potter for fortnight using the pat st unreal engine.

Back when the 480/1060 came out you could turn up the fidelity to high on 99% of the games and still be highly playable at 1080p. The 1660 is becoming a eSports card, not a full on gaming card.

My transistor count argument is based on the fact more transistors typically typed better performance. If I can pack twice as many transistors into a given area then there should be a reasonable uplift in performance without a huge bump in price.

With the exception of things like tensor cores and rt cores, the pipeline for rendering hasn't changed much. (I am impressed ooe is in use as the t generates a lot of heat and is best used when there's steady work in the pipe) But it's not like we are moving to a whole new arch where we render volumetrics instead of triangles. A lot of these cores drivers use the same HLSL (they have to) which translates into low level at run time. It's not that complicated a tech.

 

[bit_user]

My transistor count argument is based on the fact more transistors typically typed better performance. If I can pack twice as many transistors into a given area then there should be a reasonable uplift in performance without a huge bump in price.

If you hold the die size roughly constant and wafers cost more on new process nodes, then obviously those dies are going to cost more!

With the exception of things like tensor cores and rt cores, the pipeline for rendering hasn't changed much.

Tensor cores and ray tracing do consume a lot of die space and engineering resources (both hardware & software). Also, there are new shader types, such as "mesh" shaders.

Introduction to Turing Mesh Shaders | NVIDIA Technical Blog

Turing introduces a new programmable geometric shading pipeline, mesh shaders,, enabling threads to cooperatively generate compact meshes on the chip.

developer.nvidia.com

A lot of these cores drivers use the same HLSL (they have to) which translates into low level at run time. It's not that complicated a tech.

• New GPUs have new micro-architectures, which means different paths and additional backends in the shader compiler. They also have different bugs that need workarounds.
• We also don't have visibility into changes made to various fixed-function units, but those surely happen and need to be supported.
• At a macro architecture level, you surely have scheduling & resource-management code specific to different GPUs.
• A big part of "driver" maintenance also involves optimizing the shaders used in AAA titles, so they run better on all supported GPU models.

A few years ago, Jensen Huang was quoted saying "Our software layer is, therefore, about 30% of the overall cost". I'm sure it's only gone up, since then.

​

​

View: https://www.reddit.com/r/hardware/comments/jj71td/nvidia_a_software_company_that_produces_chips/​

​

 

[digitalgriffin]

If you hold the die size roughly constant and wafers cost more on new process nodes, then obviously those dies are going to cost more!


Tensor cores and ray tracing do consume a lot of die space and engineering resources (both hardware & software). Also, there are new shader types, such as "mesh" shaders.

​

Introduction to Turing Mesh Shaders | NVIDIA Technical Blog

Turing introduces a new programmable geometric shading pipeline, mesh shaders,, enabling threads to cooperatively generate compact meshes on the chip.

developer.nvidia.com

​

• New GPUs have new micro-architectures, which means different paths and additional backends in the shader compiler. They also have different bugs that need workarounds.
• We also don't have visibility into changes made to various fixed-function units, but those surely happen and need to be supported.
• At a macro architecture level, you surely have scheduling & resource-management code specific to different GPUs.
• A big part of "driver" maintenance also involves optimizing the shaders used in AAA titles, so they run better on all supported GPU models.

A few years ago, Jensen Huang was quoted saying "Our software layer is, therefore, about 30% of the overall cost". I'm sure it's only gone up, since then.

​

​

View: https://www.reddit.com/r/hardware/comments/jj71td/nvidia_a_software_company_that_produces_chips/​

​

Yes a more advanced node will cost more for the same size chip. Twice the transistor density though doesn't mean twice the chip cost for the same die area.

Margins are at all time highs. They just got greedy. And now they are paying the price for that greed.

 

[Brubbler]

This will not translate into the consumer saving a single dime. We'll have to wait many months to see any reasonable price drop on the 4000 series, if we ever see one at all. A decision will be made only when an inch of dust settles on all of the boxes and even then it'll only be a meagre drop that wont inspire builders. This gen is a lost cause.

 

[bit_user]

This will not translate into the consumer saving a single dime.

Pretty much. There could be heavier discounting on the older products, though. But, I think they'll protect the margins on the new products until the bitter end. Even there, what's likely to get addressed are outliers like the RTX 4080, rather than across-the-board cuts.

This gen is a lost cause.

Yes, for two reasons:

1. If input costs have gone up, then the only way you really address that is through design changes that are too drastic to implement mid-gen. They would only affect future products still in the development pipeline.
2. The more volume they lose, the more precious the margins are on units they're still selling. You have to be really confident that you're going to recoup any reductions in margin via more sales, or else it turns out to be a losing proposition. Time-limited discounts are one way to test this.