NVIDIA Details DLSS 4 Design: A Complete AI-Driven Rendering Technology

[londiste]

There are two main ways to speed up image rendering:

1. Increase the number of transistors in the GPU chip.
2. Develop new algorithmic techniques (AI is also one of them).

So, what’s the result? More transistors mean exponentially higher costs, making it unaffordable for many users. On the other hand, improving architecture and optimizing algorithms only requires time investment, without significantly increasing manufacturing costs.

They key bit in 2 is that new techniques may have different tradeoffs, work with different/new APIs or in entirely new ways. Which all will cause friction. Ray-tracing and ML are but recent examples of the same thing. When we want to keep using the same standard FP32 compute there really is not that much on the table any more in terms of architecture.

I think you might not be completely correct on needing "only" the time investment. Time investment is cost. If you think about all the things GPU hardware vendors do - especially Nvidia with their software focus - then they spend a lot of time and money in that area.

 

[Wales]

They key bit in 2 is that new techniques may have different tradeoffs, work with different/new APIs or in entirely new ways. Which all will cause friction. Ray-tracing and ML are but recent examples of the same thing. When we want to keep using the same standard FP32 compute there really is not that much on the table any more in terms of architecture.

I think you might not be completely correct on needing "only" the time investment. Time investment is cost. If you think about all the things GPU hardware vendors do - especially Nvidia with their software focus - then they spend a lot of time and money in that area.

You're right, "time is also a cost." However, compared to the rising costs of wafer fabrication and the massive increase in GPU power consumption, the time cost that NVIDIA spreads across end users is actually more acceptable.

 

[x4it3n]

Look, I am with you that the old days of improvement were much better than what is coming out year-to-year these days.

As for higher prices, you can blame the death of Dennard scaling, the growth of AI and crypto, the maturation and mass adoption of CUDA, ever-slowing ever-increasingly costly fab nodes, the higher proportions of sales now going to deep-pocketed corporate buyers instead of retail sales, the pandemic changing habits, EVs and other products requiring more chips, higher interest rates, more expensive shipping, inflation in general, etc, etc.

But if you are implying that these companies did not care about maximizing profits back in 2005 or 2010 or 2015 but now they do care and this is why the prices are so high now but were not in the past, I have a massive AI-enhanced bridge to sell to you.

I know that AI plays a big role here, but Nvidia should be making some GAMING only GPUs (BIOS locking the GPU to Gaming only) that are cheaper, and create another tier of GPUs for people also doing 3D Rendering, Editing, etc.. The Quadro variants would remain the High-End (with a lot more VRAM and Optimized Drivers for Professional Applications/Software).

 

[igormp]

I know that AI plays a big role here, but Nvidia should be making some GAMING only GPUs (BIOS locking the GPU to Gaming only) that are cheaper, and create another tier of GPUs for people also doing 3D Rendering, Editing, etc.. The Quadro variants would remain the High-End (with a lot more VRAM and Optimized Drivers for Professional Applications/Software).

Why would they do that? The moat they acquired with their software stack was solely because people could get started on CUDA and GPGPU with a consumer card.
One of the most prominent papers that basically made multi-GPU training famous used 2xGTX 580s. Many others used 1080ti/2080tis.
Locking GPUs down to just gaming would kill Nvidia's entire reason for existing, and the gaming market is but a speck in their revenue, so dedicating R&D time for a new cheap just for that is not worth at all.

 

[Visible Noise]

I know that AI plays a big role here, but Nvidia should be making some GAMING only GPUs (BIOS locking the GPU to Gaming only) that are cheaper, and create another tier of GPUs for people also doing 3D Rendering, Editing, etc.. The Quadro variants would remain the High-End (with a lot more VRAM and Optimized Drivers for Professional Applications/Software).

Which is what AMD did. Notice they are backpedaling that decision?

 

[x4it3n]

Why would they do that? The moat they acquired with their software stack was solely because people could get started on CUDA and GPGPU with a consumer card.
One of the most prominent papers that basically made multi-GPU training famous used 2xGTX 580s. Many others used 1080ti/2080tis.
Locking GPUs down to just gaming would kill Nvidia's entire reason for existing, and the gaming market is but a speck in their revenue, so dedicating R&D time for a new cheap just for that is not worth at all.

I don't agree, most Gamers are only Gaming, not 3D Rendering, Editing, etc. so if Nvidia released some Gaming only GPUs with lower prices then they would sell even better.
If not then NVIDIA should increase GPU production and find a ways to prevent Scalping (like forbidding the resale of GPUs for the first 6 months of release).

Which is what AMD did. Notice they are backpedaling that decision?

AMD are nowhere near Nvidia... Nvidia has a 90% market share so yes AMD need to try something else, because so far CUDA has proved that it can do it all and much better. 3D Rendering, Editing, A.I. etc. is much better on Nvidia GPUs.

 

[Visible Noise]

like forbidding the resale of GPUs for the first 6 months of release

Wut?

Really?

That‘s like a steering wheel manufacturer forbidding you from selling your car.

Think about it.

 

[Caring1]

Wut?

Really?

That‘s like a steering wheel manufacturer forbidding you from selling your car.

Think about it.

Caveats can be placed on sold items such as a house, which legally have to be complied with for 12 months, then they are yours to do with as you please.

 

[TechBuyingHavoc]

I don't agree, most Gamers are only Gaming, not 3D Rendering, Editing, etc. so if Nvidia released some Gaming only GPUs with lower prices then they would sell even better.
If not then NVIDIA should increase GPU production and find a ways to prevent Scalping (like forbidding the resale of GPUs for the first 6 months of release).



AMD are nowhere near Nvidia... Nvidia has a 90% market share so yes AMD need to try something else, because so far CUDA has proved that it can do it all and much better. 3D Rendering, Editing, A.I. etc. is much better on Nvidia GPUs.

I absolutely agree with your points overall but I have to correct you on increasing GPU production. Nvidia would definitely like to make even more GPUs for AI/business markets, the bottleneck is with TSMC and how fast new fabs or production lines can be brought online. There is a limit on how much skilled labor exists for creating and validating fabs. If the world could produce more chips, it definitely would.

 

[x4it3n]

I absolutely agree with your points overall but I have to correct you on increasing GPU production. Nvidia would definitely like to make even more GPUs for AI/business markets, the bottleneck is with TSMC and how fast new fabs or production lines can be brought online. There is a limit on how much skilled labor exists for creating and validating fabs. If the world could produce more chips, it definitely would.

I agree that TSMC can't do much more but Nvidia have the money to request for more...and with new factories onpening in the US they should get more chips in the years to come. Nvidia are also investing hugely in TSMC with their cuLitho technology to improve future TSMC nodes.
But let's be honest the problem is that TSMC are the uncontested leaders and have no competition! Unfortunately, both Intel & Samsung are years behind TSMC, therefore TSMC factories are fully booked (Apple, Qualcomm, Nvidia, AMD, and even Intel for their GPUs...).

 

[TechBuyingHavoc]

I agree that TSMC can't do much more but Nvidia have the money to request for more...and with new factories onpening in the US they should get more chips in the years to come. Nvidia are also investing hugely in TSMC with their cuLitho technology to improve future TSMC nodes.
But let's be honest the problem is that TSMC are the uncontested leaders and have no competition! Unfortunately, both Intel & Samsung are years behind TSMC, therefore TSMC factories are fully booked (Apple, Qualcomm, Nvidia, AMD, and even Intel for their GPUs...).

After the tariffs that are going to cause a recession at best in the US, I don't expect any factories to come online in the expected time. More likely, a lot of investment is going to get frozen because why invest in a country that has such unstable politics and policy. No company is going to sink billions into factories because of tariffs that might disappear completely under a president from the other party or industrial subsidies that also disappear from the current president. Plus retaliatory tariffs only complicate this even more, there is going to be mass demand destruction as prices skyrocket.

Expect production problems to increase. I am glad I took Mad King Trump at his word and splurged on electronics this last winter.

 

[Rover4444]

Caveats can be placed on sold items such as a house, which legally have to be complied with for 12 months, then they are yours to do with as you please.

It's not only illegal, it's impossible to enforce. First sale doctrine means manufacturers can't control the resale of products they've already sold, and ES CPUs are sold all the time either way. The only people manufacturers can impose requirements on are dealers.

 

[x4it3n]

After the tariffs that are going to cause a recession at best in the US, I don't expect any factories to come online in the expected time. More likely, a lot of investment is going to get frozen because why invest in a country that has such unstable politics and policy. No company is going to sink billions into factories because of tariffs that might disappear completely under a president from the other party or industrial subsidies that also disappear from the current president. Plus retaliatory tariffs only complicate this even more, there is going to be mass demand destruction as prices skyrocket.

Expect production problems to increase. I am glad I took Mad King Trump at his word and splurged on electronics this last winter.

The US might go down for a while if not for a long while! They're just heading to their own loss imo, they're too arrogant and deserve a good slap! Asian countries will dominate in the future anyway (Taiwan, Korea, India, China, etc.).