Home Computer graphics NVIDIA CEO Declares Moore’s Law Dead, Details GeForce RTX 40 Series Cost Challenges

NVIDIA CEO Declares Moore’s Law Dead, Details GeForce RTX 40 Series Cost Challenges


Moore’s Law has been at the heart of technological advancements since it was first asserted in 1965. Gordon Moore invented the Golden Rule which observes and predicts that the number of transistors in a chip doubles approximately every two years. This is largely accomplished by making transistors smaller and smaller, but ultimately the physics of semiconductor manufacturing, like anything else, has its limits.

The 2015 International Technology Roadmap pinned 2021 as the year when miniaturization would reach its limit. While further progress can undoubtedly be made, it is becoming economically impossible to do so. Now that we’re in 2022, it’s debatable whether we’ve quite reached that point. For its part, Intel thinks it has the means to extension of Moore’s law in the era of Angstrom, using techniques like the stacked fork foil transistor.

NVIDIA isn’t as optimistic about breaking Moore’s Law limits, but has other tricks up its sleeve. We had the opportunity to ask NVIDIA Founder and CEO Jensen Huang about his thoughts on process technology following GTC 2022 and the launch of the Ada Lovelace GeForce 40 series graphics cards.

NVIDIA CEO Reflects on Process Technology

Dave Altavilla, our editor, asked, “What was the impact of switching from Samsung 8N to TSMC 4N? It seems that TSMC 4N was a major advantage for you [NVIDIA]. I just wanted to understand how the process plays into Ada Lovelace’s success, because she’s an awesome chip.

Huang’s response was that the gains were surprisingly modest. “Generationally processed from 8N ​​to 4N, this process gain was probably around 15%. But unfortunately the cost increases by more than 15%,” he explained.

NVIDIA Architecture Ada Lovelace

Although he did not specify how much the cost had increased, he compared Moore’s Law to a density of transistors versus the cost curve which historically had increased. He said: “But for the very first time, starting at about, I think about seven nanometers, yeah, about seven nanometers, the curve actually went down. He didn’t flatten out. He refused. The CEO of NVIDIA then said it bluntly, noting that “Moore’s law is dead”.

As process technology has approached physical limits, the complexity of production has increased dramatically. Jensen explained that the development of his chip is done in a series of steps, with about one step completed in the factory per day. Current cycle times have ballooned to around four months from start to finish.

“It’s not because TSMC is trying to capture more profit. It’s just not true,” he assured, “it’s a lot of steps. is more expensive, it takes longer in the fabs.

“So you have to find other ways to do it.”

Huang explains that, “the way we solved it, Dave, with Ada is architecture.” Ada combines several architectures: CUDA, Tensor and RT. While traditional advancements have been made through CUDA for rendering and rasterization, Jensen noted that Tensor and AI are the “giant lever” for Ada.
NVIDIA DLSS 3 pipeline

“I would say that DLSS produces a better looking pixel than GroundTruth, than the rendered pixel we calculated in the past,” Huang explained, claiming objectivity. He went on to say, “And the reason is that DLSS learned that pixel, not by calculation, it learned from a 16K image. Not a 4K resolution image, a 16K resolution image. And so the pixels he learned from were beautiful. There were no pixels. It’s close to raw. And so we had the AI ​​learn what colors to put in there, and it makes for a nicer color.

Interestingly, NVIDIA’s technology progress with DLSS 3 also marks, in a way, a reduced need for pure GPU rendering, at least when the graphics requirements are high. The CEO claimed that artificial intelligence does better. “It’s subtle enough that in my opinion it’s better than GroundTruth,” he said, “It’s definitely different from GroundTruth, but it’s better than GroundTruth.”

NVIDIA DLSS 3 optical flow acceleration

To accomplish this, graphics rendering transcended a set of operations performed on individually siled machines around the world. “Ada is generating infographics, not just right now. There’s a supercomputer in the background that was training the neural network,” Huang explains, “And so, in many ways, Ada is Ada plus the supercomputer that we have. And this is the future. It’s the future, and that’s why the infographic is so amazing.

Tensor cores gain prominence, and so does AI

Huang went on to note that “…the combination between the supercomputer for priority training and the coding of pixel knowledge in the neural network, and then the processing of the neural network, which is structurally very good for the processing of the tensor core , we’re effectively increasing the throughput of our GPUs by a factor of 8. Well, that more than overcame the weakness of Moore’s Law.

Artificial intelligence arrived just in time

Talk about DLSS 3 specifically, Huang said, “Without the CPU doing anything, we generated an additional frame. It doesn’t predict the future yet, but it’s near and it predicts an extra frame without the CPU ever having to touch it. If these images are accurate and as beautiful as the traditional render, if not better, we’re excited to see the progress this will bring in the future.
NVIDIA DLSS 3 reconstructs 7 of the 8 displayed pixels

“Well, we just doubled the frame rate, but the neural network does all the work. And so, I think we have to overcome the weakness that we are at the end of Moore’s Law, not in giving up, but coming up with much smarter techniques, and thank goodness artificial intelligence arrived just in time.

Where 3D Graphics Go From Here

Some have resisted the advent of neural networks and a shift from the “ground truth” of raw GPU rendering in general. To that end, we might ask what the purpose of a graphics card is: isn’t it to generate as much eye candy as possible?
Future performance gains with DLSS 3 technology

Yes, the way this has been accomplished historically is by rendering but this is not the only way to go. Our desires alone cannot beat physics. Whereas AI scaling techniques had a shaky start with peculiar artifacts and quirks, technology advanced and improved rapidly and dramatically. And of course, we can’t claim that traditional rendering has been out of trouble That is.

The unfortunate consequence, as discussed in the process technology discussion, is that costs have increased. Hitting the Moore’s Law wall hurts, but AI could probably provide the path to lower-cost solutions, eventually. For those who expect or depend on “pure” rendering and rasterization, the way forward will likely only become more expensive.