Asked why Nvidias latest 40-Serie graphics cards cost just one thousand dollars, CEO Jensen Huang said that Moores Law is dead. He explained that the days of constant rising costs are over, with the technology is continuing to slowed and the pandemic shortage pushed things further.
But don’t worry too much. The advancements of AI and gaming will combine to propel mankind’s ambitious dreams, like the metaverse.
Huang spoke at the Nvidias virtual meeting last week.
In 1965, Moores Law assumed that the number of chips on that chip would double every couple of years. Upon the release of a metronome, a year on a chip would double or costs would halve.
It kept on the road for decades, mostly because of advances in manufacturing. But as the limits of physics reach a minimum, those advances aren’t used for granted anymore. Intel is investing heavily to stop the law. But Huang said that smart chip design must take over, which is why the company moved from a new architecture to the latest generation of graphics chips. The new 40 series graphics chips will give a great performance for PC games as soon as we get ready to hit the global grind.
Nvidia’s Omniverse Cloud is a cloud-based computing device.
It is more important than ever to keep current advances in performance and power efficiency going, as did a second attempt to build the metaverse, the 3D universe of virtual worlds all connected, like in novels such as Snow CrashandReady Player One. Nvidia has built a variety of tests and simulation tools that would be able to simulate this metaverse.
Those are not real metaverses unless real-time and can accommodate many more people than allowing access to 3D spaces today. Nvidia plans to use the Omniverse to create a digital twin of the Earth in a supercomputing simulation dubbed Earth 2, so that the climate is predicted for decades to come.
This means we should get the metaverse for free, and you should need all the power to make this sound even better. And he noted that AI, built up by the graphics oriented forward by gaming, will allow developers to auto-populate their metaverse worlds with 3D content interesting. In other words, in order to improve the game, the AI, and the chips will also be driving forward. To me, it seems like a new law is a little bit a bit new.
Here’s the transcript from the newspaper that we wrote. We transcribed the entire newspaper that we were attended and withdrew to me and a number of other journalists.
Jensen Huang of Nvidia says Moores law is dead.
Q: How big can the SaaS business be?
Huang: That’s hard, isn’t it? That’s the true answer. This depends on the service we offer. Maybe another way to go is to be one. That GTC, we announced new chips, new SDKs and new cloud services. I had selected two of them. One of them is large language models. If you hadn’t got the chance to look into the effectiveness of language-based models, and the implications of this approach on AI, please do so. It’s important stuff.
Large language models are difficult to learn. The applications are very diverse. It has learned very much human knowledge, therefore it can recognize patterns but also contains lots of encoded human knowledge. That has a human memory, must you? In one way it encoded a lot of our knowledge and skills. If you want to adapt something that it was never trained to do, it was never trained to answer questions or to explain the story or to publish a breaking news article. It was never trained to do that. A student can learn this skill with a few more shots of learning.
This basic idea of fine tuning and adapting for new skills, or what’s called zero or few shots of learning, has huge implications in many fields. You see such a high amount of funding in digital biology. Large language models have learned the language of the structure of proteins and the language of chemistry. So we put up this model. How large can it be? My sense is that every company in any country speaking every language has probably tens of different skills that their company could adapt, that our large language models would perform. That isn’t really a big chance, but it’s likely one of the biggest software opportunities of all time. That’s because automation of intelligence is one of the largest ever opportunities.
The other opportunity we spoke about was the cloud cloud. Remember what OmniVerse is like. OmniVerse has a number of features. It’s first a characteristic of a dog. It can be stored. It can composite physical information, 3D information over multiple layers or what is called schemas. It can describe geometry, textures and materials. The properties of weight and weight. Connectivity. Who is the supplier? How much is the cost? What does that mean? What is the supply chain? I would be surprised to see that behavior, kinematic behaviors it would be AI behaviors.
Nvidia’s Omniverse Avatar Cloud Engine is the latest model for the Nvidia GAV.
OmniVerse stores data first. It’s about connecting multiple agents. The agents are people. They’re robots. They can be autonomous systems. The third thing that’s done is to postpone this other world. As for simulation, it’s a way of saying a simulation engine. OmniVerse is basically three things. This is a new type of storage unit. This is a new type of connecting platform. It’s a new type of computer platform. You can write a project with OmniVerse. You can use OmniVerse to connect other applications. For example, we showed many examples, with Adobe being connected to AutoDesk and integrating multiple other applications. We’re bringing them together. You might be connecting people. You could connect. You may be going to be using robots. They can get on with agents.
The best way to think about what we did with OmniVerse is to think about it almost like the easiest way to monetize something that is certainly like a database. It’s a cloud-based database. This database is only in 3D. This database has multiple users. These are two applications we had. The first is the large language model, the other is the OmniVerse, basically a database engine that will be in the cloud. I think those two announcements are good. I’ll get plenty of opportunities to talk about it again and again. But these two SaaS platforms will be very long-term platforms for our company. Well keep them run in many clouds and so forth.
Q: Nvidia said it would reduce GPU sales through to Q4. Do you mean business or financial finance? Can you confirm the lower selling will last a few more quarters?
Huang: It depends on the Q4 which ends in January. It’s off at a month. I can tell you that because we guide only one-quarter at a time, we have big differences, much lower prices than those selling out of the market. A significantly lower price than what is out of the bargain market. I hope that the channel will be tuned for a re-launch to make for a great start for Ada by Q4. With a small dose of action, start shipping Ada first. As for the vast majority of Ada, then it’s scheduled to launch next quarter. I can’t predict the future too much, but a bit too much expectancy and current thinking are so that what we see at the marketplace, what we know in the channel and what we do are ours, our next we will have a decent 4 or 4 Q4 for Ada.
Q: What do you think about the progress of the metaverse, especially a real-time metaverse, that would be more responsive to the internet to the moment of us? If its coming along maybe slower than those in the past, what are some things that could make it happen faster? Wouldn’t Nvidia consider the option to invest to make this happen faster?
Huang: We have to do a few things, to make the metaverse, a real-time metaverse, realisable. First of all, as you know, the metaverse is created by users. That translates to the ability of AI. And in the future, it is extremely likely that best describe some characteristics of houses or cities or neighborhoods similar to those like this, like Toronto or New York City, and that creates a new city for us. If we don’t like this, we could give you additional prompts or we can just hit enter before it generates one we want to start from. And then, from that world to it, modify it.
The AI for creating virtual worlds is being realized as we talk about it. You know that, the core of that is precisely the technology I was talking about just a second ago, called large language models. You’d better learn from all the creations of humanity, and you’d better imagine 3D worlds. Because of words through a large language model, will come out someday, triangles, geometry, textures and materials. Using that, we would alter it. No of this is pre-relative, and all of the simulations of light need to be done in real time. That’s the reason why the latest technologies that were developing for RTX based on simple narrow renderings are very important. It is a brutal force. Well, we need the help of AI to do that. We just demonstrated Ada with DLSS3 and the results are pretty insanely impressive.
The first part is generating worlds. The second is simulating a world. And then the third part is to remarry that, the thing you’ve mentioned earlier, about interactivity. We have to get on to the speed of light. One new type of data center must be installed worldwide. I talked about it at GTC and called it a GDN. Since Akamai was given CDN, I think there’s a new world for GDN, a graphics distribution network. We demonstrated how strong this is, through the improvement of our GeForce Now network. We have that in 100 regions around the world. By doing that we can have computer graphics, which is almost instantaneous. We’ve shown that on planetary scale, we can generate interactive graphics up to 10 milliseconds, essentially interactive.
In a Lowes shop with the Magic Leap 2 headset, a sales representative uses the Magic Leap headset.
And then the second part of this is how do raytrapping in an augmented way, or as a VR-like way. We also demonstrated that recently. The pieces are getting together. Besides the engine itself, a database engine known as OmniVerse Nucleus, the worlds that are either built by humans or augmented by AI, the universes made of a robot through the simulation and rendering, and then graphics, GDNs around the world, all the pieces were putting together were coming together. Until you saw us at GTC and working with a really good company called ReMap. With the help of their CEO, we created a new design studio, so they could publish a photo-configurator all the way to the world, literally using the press of the button. In instant, we published an interactive simulation of cars in every corner of the world. I think the pieces are going together. Now that Ada is in production, we must get Ada positioned under the fire of the public clouds of the world, positioned in companies around the world and continue to build our distributed GDNs. The software will be there. The computing infrastructure will be here. We were pretty close.
Q: Given the inventory issue and physical supply chain issues we see that with OmniVerse cloud youre moving into SaaS. You have GeForce now. Do you think a point, where youre supplying the card as a service, instead of disdistributing it anymore?
Huang: I don’t think so. There are customers who like to own them. The customer likes renting. You’re renting or letting us out or you are able to do it for you. Businesses are the same way. Do you like any capex or opex? Startups prefer to have things in an opex format. Large, established companies should have capex. You can rent any kind of property you choose, don’t you? If you use it all the time and load it, you’d rather just use it and have it. Some people prefer outsourcing the manufacturing.
Remember that AI will become a factory. It’s going to be the biggest factory of the future. You know a factory is made up of raw materials, and that something comes out of it. It will be easier for the factories to bring data, and in the future what has to come is the intelligence, the models. The transformation of the energy will be. Even though some people hate to sell their factory, others prefer to own this. It depends on the business model you’re in.
We’ll probably continue to build computers with HP and Dell and the OEMs around the world. Now continue to provide cloud infrastructure by the CSPs. It’s very nice to know that Nvidia is a full stack accelerator. One more way to say this, I said it, twice, but an acceleration computing company needs to be complete. The reason for that is because there’s not a magic thing you put into the computer and it doesn’t matter what kind of application it is, it runs 100 times faster. Accelerated computing involves understanding the application, the domain of the application, and re-factoring the entire stack so it is much faster.
And so accelerated computing, over the course of the last 25 years we started with computer graphics, went into computing and AI, then data analysis. You saw us recently in graph analytics. Over the years we have stepped into so many disciplines that it seems that Nvidia architecture accelerates everything, but that isn’t true. We’re going to speed. We just accelerate 3,000 things. These 3,000 things are all accelerated under a single architecture, so it looks like things get faster if you bring the Nvidia chip into your system. Because we threw a single domain with the other two. We had a hard time.
We had the discipline to live with the same architecture in order to be the only one with an extended time frame. Since the software stack we accelerated was the end result of the new chips we created, for example Hopper. If you develop a new software that’s in line with our architecture, it will run on an entire installed base of 300, 400 million chips. Because of the discipline that has continued for more than a decade, it’s definitely this magical chip that accelerates computing. What good do is push this platform out in the best possible way to the world so that people can develop an application for it. Maybe there are new quantum algorithmes that we can develop for him, so that it is ready for cryptography in 10 or 20 years. Starting with new search optimizations. The development of new automated detection methods, digital fingerprinting. The platform should be there so people can use it.
There are 3 different domains where you’ll see us do more. The reason why it’s difficult to do more is because it makes it difficult for me to do that if I did it once myself, not only would I understand, but we can open the drawers so that the rest of us can understand it. Let’s give you an example. Obviously you’ve seen us now take computer graphics to the OmniVerse. We built our own engine, our own system. We were able to get it right. This is because we wanted to learn the best techniques to do real-time raytracing on a large data scale, using AI and brute force to route tracing. Without OmniVerse, we would have never developed that skill. I don’t think any game developer would do it. We pushed the frontier for that reason, so we can open up RTX, and RTX DI and RTX GI and DLSS, and we can get that into everyone else’s applications.
A Nvidias Earth-2 simulation will model the climate change.
Drive is the second area you saw us do this. We built a autonomous car system that runs across the street so I can learn how to build robotics from end to end. We can build a technology that is so much more flexible that we can create robotics in one piece. We have built Drive now. We opened up all of the pieces. Our synthetic data can be used to generate. They can use our simulators and so on. You can use our computational unit.
The third area is big language models. I built one of the world’s largest models in the world, earliest, almost before anyone else did. It’s called Megatron 530B. It is still one of the most sophisticated languages in the world. We can say that as a service, so we can understand how it works.
And then, then of course, so that we understand how to build a platform in orbit for metaverse applications that will be primarily focused on industrial metaverse applications, and then we can do it. They must build a database-engine. We built the OmniVerse Nucleus that was well placed into the cloud. In some applications, one might think that one can make a contribution, in which it is very hard. You must look at a data center scale and an entire stack scale across the planet. But otherwise, keep the platforms completely open.
Q: I want to ask you more about China export control restrictions. Based on the information about the requirements for the licences at this moment, do you anticipate that all your future products beyond Hopper would be affected by those standards, based on performance and interconnect standards? And if so, do you have plans for China to market specific products that will continue to conform to the rules, but that would incorporate new features as you develop them?
Huang: Hopper is not, first of all, a product. Hopper is an architecture. Ampere is not a product. Ampere is an architecture. Notice that Ampere has A10, A10G, A100, A40, A30 and so on. Where’s Ampere – there are, gosh, how many versions? Probably 15 or 20 o’clock. She’s in the same way. Hopper’s will be used to make many purchases. These restrictions specify a specific combination of computing and chip to chip interconnection. It is very specific to this. To keep it within this specification, the space under this specification is vast for us and for the customers. In fact the vast majority of our customers aren’t affected by the specification.
We expect that for the USA and China, a lot of products can be manufactured, without a limit and without a license. However, if a customer wants to use the limitations specified by the limitations or beyond, we should go get a license for that. You could argue that the goal isn’t to stop a business. The goal is to know who it is that would need the skills at this limit and give the U.S. the opportunity to decide whether this level of technology can be available to others.
Q: I recently spoke to a big software developer about AI and the metaverse. We talked to a bit about how AI can help develop virtual worlds and games. Sure, asset creation is a problem, but also pathfinding is more often than not. Regarding the automotive industry, these technologies may be very relevant. You have situational awareness like that. Could you give us an idea of how this could develop in the future?
Huang: As you’re viewing the keynote, you’ll notice there were a lot of different areas where we demonstrated pathfindering. When you think about our auto-driving vehicle, you’ll see three things happening. The sensors are coming into the computer. We can sense the environment using deep learning. We could see and then reconstruct the environment. The reconstruction doesn’t have to be due to the fidelity of the present generation, it has to know its surroundings and important features, and where a barrier is found and where it’s likely that obstacles will be found in the near future. There is perception, and then second, world model creation. Developing the world model, you must see where everything else is around it, what map tells you, where you are inside the world, and reconstruct the object from that from the map and from that from those of that person. Some people call it localization and mapping for robotics.
Two Isaac robots arrive in the warehouse of Omniverse.
The third part is for path planning, planning and control. Planning and control are route planning, which has AI, and then path planning, which is about wayfinding. The wayfinding has a rhyming with places where you should go and where the obstacles are around you and how you want to navigate around that. You saw a thing called PathNet in the demo. I saw a lot of lines that came from the front of the cars. Those lines are essentially options that we need to select which of these paths is the best, the most dangerous, and the most comfortable, with you right to the end of the road. You always make easy decisions. The second is ISAAC for robots. The wayfinding system is somewhat, if you’d rather, unstructured in the sense that you don’t have to look for lanes. The factories are poorly built. There are many people everywhere. Nothing’s often marked. You just need to go from one point to the next. The two times you will have to avoid obstacles, find the most effective route and not block yourself. You can get stuck in the dark, and you aren’t ready for that. There are various kinds of algorithms that allow you to do path planning.
In the ISAAC path planning system, you can see that in a game. There you can say, army, go from point A to point B, these points are much different. Between point A and point B, the character has to navigate across a mountain, boulders and bushes, to run around an river, etc. And so we could articulate the way it is done. You saw ISAAC do that, and there is another piece of AI technology that you might have seen in a demo called ASE. Basically, its Adversarial Skill Embedding. AI learned to speak in a person by watching humans at this point. You can say, hike the trees a lot, but take the trees. Swing the sword. Kick the ball. You can describe an animated individual in its phrases. I just gave you essentially the AI models that allow us to take multiplayer games and have AI characters that are realistic and easy to control. To that point, there will be some real people, some agents of AI, and some avatars that you saw using VR or other means. This technology is already up.
Q: How will you see the future of the autonomous driving business, since you introduced your new chip for a autonomous vehicle? Do you think that is still in the early stages of a business such as that, or see a wave coming up and cutting the industry? Can you tell us a bit about your strategic thinking?
Huang: First, the autonomous car has two computers. There’s the computer in the data center that has the ability to collect data, a process used by cars, can transform that data into trained models, can build the application, simulate the application, and regress the entire time, building the map, generating the map, and then reconstructing that map, if you want, then to do CIC and then OTM. That first computer is basically a self-driving car, but a data center. The car isn’t in the van anymore, because it collects data from the entire fleet. That data center is the first part of the self-driving car system. It’s an AI lab, training and mapping.
And then the second part is that you take all that and place it into the car, and a little of it. The small version is what we call a chip. The next version is called Thor. That chip has to do data processing. It is called perception or inference. It’s necessary to build a world model. They must do a map. It must plan the path and maintain it. And these two systems are continuously running, with two computers. Nvidias is done on both sides. In fact, you might say our data center has got an increased business capacity, so it’s certainly larger, and frankly, long term, the greater of both parts. The reason is that the software development for autonomous vehicles no matter how much work it will cost, will never be finished. Almost all companies are going to run their own stack. That’s a very important part of the business.
GeForce Now is accessible via Nvidia Drive.
The first client for OmniVerse is DRIVE Sim, a digital twin of the fleet. DRIVE Sim will be a very significant part of our autonomous driving business. We use this internally. Please allow other people to use them. And then the car is about a thousand things that we believe in the philosophy. As long as you study the way people built ADAS systems in the past, that is what Nvidia created for us, we invented the Xavier chip, the first of the world’s first software programmable robots, and that is a real computer programmable robotics. It was designed for fast sensors. It uses more advanced tools. It has Cuda in the area for localization and path planning and control. Many people, who said why anyone would need such a large SOC when I first introduced Xavier. It turns out that Xavier is not sufficient. We needed more.
Orin is a man. If you look at our robotics business now, which includes auto-driving vehicles, car-transmissions and trucks, and auto-synctic systems, our entire robotics business runs already a total of $1bn per year. Orin is on its way to the westend. The pipeline is now worth 11 billion dollars. My sense is that our robotics business is about to double in one year, so its becoming an integral part of our business. Our philosophy, which was extremely different from what happened in the past, is that many different technologies have come together to make robotics possible. Of course, one of them is intensive learning. We were the first ones who brought deep learning to a drive-off. I really thought it was from here. It was designed for computer-based vision. It was created by engineers. We used deep learning because we felt that was the easiest way to do that.
Secondly, everything we did was software-defined. You could easily update the software because there are two computers. There’s a computer in the data center that developed the software, then we put the software in the car. If you plan to do that on a larger fleet, do that fast and improve software on the basis of software engineering, then you’re definitely in need of a programmable chip. Our philosophy, to use deep learning and to use a platform fully software-defined platform, was really a good decision. It took a little longer since it cost more. People must learn how to develop their application. And yet I think it’s time for us to conclude, that we all will use this approach. It’s the right way for us. Our robotics business is already going to be a very large business. It’s already a great business, and it’ll become much bigger.
Q: On the AI generation mentioned for Ada, which is not just generating new pixels, but also increasing the whole of new frames, with the different sources that we have for AI-generated images, we see DALLE and all these different algorithms blowing up on the internet. It may not be the best use case for video games. But how can any other side of creation be done? Technology like broadcast and a bunch of other things get to the creators. How can other users, and then gambiters use it to generate new images, export new frames, and stream new frames at new frames? Are you studying that approach in a bid to use this AI technology more?
Huang: First, having the ability to synthesize high-priced graphics, using path tracing, is fundamental. That enables user-generated content. Remember, I mentioned in the keynote that nine of the world’s top 10 games today were game modes in one minute. That was because someone took the original game, which made it a more fun game, such as a MOBA, a 5-on-five, a PUBG. That required fans and fans to alter a certain game. That took time.
I think that in future, would have a lot more user-generated content. If you have user-generated content, it’s obvious that they don’t want big or large corporations to put up another wall or tear down this other wall or destroy the castle or forest or act any way they want. If you change those things, these structures, world, then the lighting system isn’t correct anymore. Using Nvidia’s tracing system and doing all in real time, we made sure all the lighting remained clean because they were simulating light. We are not required to clean up before taking a cold. This is really a big deal. In fact, as per standard, RTX and DLSS 3 and OmniVerseweve have created a version of OmniVerse called Remix of RTX. If you combine these ideas, then user-generated content will flourish.
A car is used to tell a story.
When you say user-generated worlds, what is that? People will say that’s the metaverse and it is. This metaverse concerns user-generated, user-created worlds. I think that everyone will be a creator someday. You’ll take OmniVerse and RTX and give it this neural rendering technology and build new worlds. Once you did that, once you was able to simulate the real world, the question is, how do you use your hands to create the entire world? The answer is no, because we have the benefit of mother nature in our world. We can’t do that in virtual worlds. But we’re living in AI. If you only say that, give me an ocean. Give me a river. Give me the pond. Give me a forest. Tell me a palm tree. You describe everything you want to describe and AI will assemble a 3-D world that will be presented to you. That you’ll be able to change.
The world that I described requires new techniques for computers. We call it neural rendering. That platform we call RTX is the same as the computing platform behind it. It’s important, number one, making video games, today’s video games, a lot better. Make the framerate higher. Many of the games today, because the worlds are so big, they’re becoming very limited on CPU. By using the time frame is produced in DLSS 3 we can maintain the frame rate, which is quite remarkable. However, the whole world of user-generated content is second. And then, the third is the environment in which they were today.
This video conference that was held today is somewhat archaic. Videoconferencing was created in the 1960’s. Videoconferencing isn’t being encoded or decoded in future. It will be perception and generation in the future. Perception and generation. Your camera will look you and then go with me. You can control that generation. As a result, a higher framerate will be achieved for everyone. Everyone has better visual quality. The amount of bandwidth used will be small; it’s just a small amount of bandwidth, maybe in kilobits per second, and not megabits. We’ll learn to use neural rendering for videoconferencing. It’s a different way of saying telepresence. There are quite a few different applications for it.
Q: I noticed that there wasn’t any NVlink connector on the card in the presentation. Has it gone for Ada?
Huang: There isn’t a NVlink in Ada. We’re getting it out, because we needed help with other things. We used my hordes and the area to make a huge difference in AI processing. And, because Ada is based on PCIe Gen 5, we can now do peer-to-peer across Gen 5 — the speed of that decision is sufficiently fast that it was a better balance. That’s the reason.
Q: The trade question is of some kind – does you feel that trade restrictions are a problem?
Huang: First of all, there needs to be fair trade. That’s questionable. There’s a need for national security. It always needs some attention. There’s a lot of things I know, who I know but have no idea. Even though nothing could be absolute, nothing can’t be the truth. You only need to get degrees. You can’t have trade open, open, and open trade. Without the attention of the national security, there is no unlimited access to technology. But not all is trade. And you can’t have any business. This is just a matter of degrees. The limit and the licensing restrictions affected by the fact that we still have plenty of space to trade in China with our partners. This gives us a lot of space to innovate and continue to serve our customers here. If we need the most extreme examples and the most advanced tools, we should obtain a license.
From my perspective, the restriction is not different from any other technology restriction that was placed on export control. There are many other technology restrictions in the field of memory. Several processors have been in a long time, yet a majority of them are now being used everywhere, yet often worldwide. The reason for the fact that we had to disclose this is because it came in the middle of the quarter, and it came suddenly. Since we were in the middle of the quarter, we thought that was material to the investors. That’s the key to our business. To those who were affected, it wasn’t significant part of their business, because accelerated computing isn’t much of a big part of Nvidia. Unfortunately, it was an important part of our business, so we must reveal. As far as I know, for the purpose of the Ampere and Hopper architecture, we have a large scale envelope to innovate and serve our customers. From that point of view, I’m no more concerned.
Microsoft Flight Simulator doubles its frame rate using DLSS3 on a new Nvidia GPU.
Q: 4000 will be here again, which in any case feels like a big launch. As far as the response is universally conceived, its cost is great, my God. Does there any question you want to ask about pricing on the new generation of parts for the community? Can they expect better prices at some point? Basically, can you address the loud screams that im seeing everywhere?
Huang: First of all, a 12-wafer is much more expensive today than yesterday. It isn’t a bit more expensive. It’s a ton plus expensive. Moores Law was taken down. The ability of Moores Law to perform the same twice the price or the same performance half the cost in every year and half is over. It’s nearly over. The idea that the chips will reduce in cost over time, unfortunately, is a story of the past. We have an immediate future ahead of the clock. It’s not easy to make a new architecture, to design as good as it can and then, of course, the computing isn’t a problem. Computing is a software and software problem. We call it a total challenge. We innovate in a single layer.
For all of our gamers there, here’s what I want you to remember and hopefully notice. At the same price, based on what I said earlier, although our costs, materials and costs are more as we used to be, Nvidias performance of $899 or $1599 dollars, previous two years ago our performance with Ada Lovelace is significantly better. The better off the charts. That’s really the foundation for him. Of course, numbering is just a numbering system. In the past, as far as 3080 compared to 1080 compared to 680 p.m., all the way back to 280a 280, obviously, was considerably less expensive.
With time, we must create to pursue a progressive evolution in computer graphics one hand, deliver price equal to the other, expand deeper into the market, and with lower and lower prices for all these things if you look at our track record, all three had done all three, and keep trying the other hand. We have brought the new frontiers of computer graphics further into new applications. Watch all the great things happening with the advancement of GeForce. At the same price, our value delivered generationally is off the chart, and then again is off the chart. If they could remember the price point, compare price point to price point, they will find that theyll love Ada.
Q: You talked about everything you are planning for your company and the big expectations you have from the robotics industry. Are there any things you keep on sitting at night, which could damage your business and how it is doing at the moment? Are there things you consider as challenging you’ve got to do?
Huang: This year I will say that the amount of external environmental challenges that are causing great challenges in the worlds industries is very large. It started with COVID. There were challenges to supply chain. There aren’t much supply chain stores in China. Cities that are locked up are up week-to-week. More supply chain issues. There is suddenly a war in Europe. Energy costs have risen. The inflation is up. I don’t know. Any things else that can’t go wrong? That’s not my fault that I am not at night, so we lose control of them. We try to be the best, make good choices.
Three or four months ago, we made some very good decisions when the PC market began to slow down. After seeing the sales, due to inflation, the consumer market starting to slow, we realized that we will have too much inventory coming to us. We have a stock and supply chain started at the beginning of last year. Those wafers and those products are coming at us. When I realized that the sell-through had to be limited, instead of continuing to ship, we shut ourselves down. We took two quarters of hard medicine. We sold our customers, and they sold everywhere a lot lower than those that were selling from the channel. That channel, just the mobile gaming channel, is now worth 2,5 billion dollars a quarter. We sold much less than two or three in Q2. We made the trailer ready and the partners prepared for Ada launch.
I think to do what we can do, we try to make wise decisions. The rest of it is continuing to innovate. We built a hopper during that awesome time. We invented the DLSS 3. We were able to draw neural signals. We built OmniVerse. Grace is being reconstructed. Orin is being put in the ramp. In the middle of all this worked to save the worlds financial and technological costs. If you can accelerate Hopper, Hopper is a factor five times bigger for large language models. If you add Hopper to the system, the TCO’s still improved by three. How can you improve TCO by a factor of three after the completion of Moores Law? It’s really amazing results, helping customers save money, and also to create new ideas and new opportunities for their customers to reinvent themselves. We were focused on the right things. If all of these challenges, environmental challenges, are to pass and then then the good old ones would not do any amazing things. None of this stops me going the night.
Q: You have started shipping H100. That’s all good news for you! The big rail station is from spring. But since Lovelace is out now, Im curious. We’ll be flying an L100? Can you give me any advice on how you are going to share these two architectural elements for the time being?
Huang: If you look at our graphic shop, let’s go back to Turing. In the meantime, it was only two generations ago, or about four or five years ago our core graphics business was basically two segments. One of these is desktop and gaming. One is desktop, and the other is computer and software. The two were really one. Desktop computers and desktop gaming systems. The Ampere generation opened up a whole lot of notebooks, because of its energy efficient ability. Thin and light gaming systems, thin and light workstations became a real major driving force. The laptop business is very large, as it is comparable to our desktop business or near this business. We were well-suited to the application of Ampere generation and the integration of it into the cloud, in the data center and in the cloud. It’s used in a data center because it’s ideal for inference. The Ampere generation proved successful on inference GPUs.
This generation are going to see more things. There are new dynamics, long-term trends that are very clear. One of these is cloud graphics. Cloud gaming is, of course, a great thing now around the world. In China, cloud gaming will become very important. There are a billion of phones that game developers can’t serve anymore. They make perfectly good connections, but the graphics are so sloppy they don’t know how to use the game that is built for the modern iPhone 14 and run on a 5-year-old phone. This technology has been developing fast. In China only there are a billion phones. I think, the rest of the world, there is a number of devices. Modern games don’t let developers serve those anymore. The best way to solve this problem is to use cloud computing. You can make an integrated graph. You can reach mobile devices and so on.
If you could do that for cloud computing, then you can do that for streaming applications that are not fully optimized for the game. For example, what used to be workstation applications that would run on PCs would be simpler. The future would be SaaS, which streams from the cloud. The GPU will be one of the current four-year-olds and four-year-olds. The Ampere GPUs will be based on graph intensive applications. And then there’s a new, very important, the augmented reality streaming device. Developing short-form videos, enhancing these images, or reposing, you’re using your eyes to contact everybody. Maybe its just an excellent picture, and youre animating the face. Those kinds of immersive reality applications will use GPUs in the cloud. In the Ada generation, would see a lot of new projects using GPUs for processing and capturing data from the cloud, particularly as the technology for augmented reality, televised devices and telephotos. It’s going to be the universal accelerator. It’s definitely coming. I actually did not call it L100, and I called it L40. Our high end GPU will be L40. It’s going to be used for OmniVerse, for augmented reality, for cloud graphics, for inference, for training, etc. L40 will be a phenomenal cloud graphics GPU.
Q: It seems like a lot of stuff youre releasing, the car side, the medical sideit feels like very few people are in AI safety. It seems like its more hardware accelerated. Are you talking about the importance of artificial intelligence?
Huang: It’s a huge question. Let me break down things into a few parts as a point of origin. Questions governing AI can be answered in general. Even though you develop an AI model that you trust that you, that you study with curated data, that you don’t believe in a bias and unnecessary bias as it is in the context of safety, so you want to have several things. The first thing you want to do is to get diversity and redundancy. One example is in the context of a self-driving car. You want to see where obstacles exist, but you want to look at where there’s no one else’s obstacles, which naive space that if there are no entrances, can’t you? Is there any obstacle to avoid, free space which you may drive through? These two models, even when overlaid on top of each other, give you diversity and redundancy.
TSMC is made for Nvidia.
We do it in a company. We do that in the medical field. It is called multimodality and so forth. We have an array of algorithmic attribution. Our calculations are diverse, meaning we are able to do processing in two different ways. We make our diversity a technology. Many of these are from cameras. Some of them are from radar. Some of them comes from space to motion. Some of it comes from lidar. You have different sensors, a different algorithm, and then a different algorithm. These are security layers. And then the next piece is – lets suppose you build an active safety system. You believe you’re resilient in that way. How do you know this isn’t tampered with? You designed the correct method. However, someone came in and tried to get it working. We must ensure we have a technology called confidential computing. To encrypt and make sure the system isn’t tampered, to to check the system and change it. You’re affected, right? And then back to the methodology of software development.
Once you sign and validate a full stack to be safe, you need to make sure that all the engineers in the company and everybody who contribute to it contribute to the software and improve the software in a way that retains its ability to remain certified and remain safe. It’s culture. There are the tools they’re using. It’s all method. There are standard standards for documentation and coding. I forgot about it. The data center isn’t tamper-proof. If no, somebody may tamper with the model at the data center just before we rolled the model on to the car. A real thing is, there is nothing but active safety, safety design, and security design, AI. We commit ourselves to doing that right.
Q: We had a limited number of production capacity at TSMC. Do AIBs also need to pre-order GPUs so far in advance? How many percent more these days are Is it better?
Huang: The supply chain was very challenged last year. Two things have happened. One thing is the time is extended. Time to ship the products would be about four months from the time you placed a PO on the wafer. Maybe slightly longer. In 16 weeks? It has been extended to a year and a half. Not just the wafer starts. You have structural conductive material, a voltage regulator, etc. in order to ship a product. It has whole system components. The number one time we had been out walking out was significant. Second in the main point, because everything was so scarce, it was necessary to secure your allocation by the advance, so you could continue to receive the allocation by approximately the year. Between normal working conditions of 4 months to the moment about 2 years or so, when you have to be able to arrange for this. And we got fast. Our data center business was growing nearly 100 percent every year. That’s a multimillion dollar business. You can easily imagine how much commitment we had to put for the growth rate and the time spent for the next cycle. That’s the reason why we’re having to take the difficult decision as demand slowed down, especially among the consumers, to really dramatically slow down shipments and let the channel inventory manage itself.
Those organizations don’t need to lead time orders, especially in regards to AIBs. We ordered the components no matter what. Our AIBs are agile. We carried the whole inventory. Since the market was very hot, the channel was exactly the same selling price. It never moved a dollar. Our component costs were high, as they known last year, but I absorb the all the costs in the project. We passed no money back to the market. Our product cost exactly exceeded the price we launched. Our AIBs created different scales that allowed them to see more value. The channel, of course, the distributors and retailers, did profit during the time when the product was hot.
When the demand slowed, we took action to develop marketing, what we called marketing programs. But basically discount programs, rebate programs, allowed the price on the market to return to a price point that we felt or the market felt would ultimately sell through. The combination of our commitments and youyou guys got to know that we made a minimum of about a billion dollars of inventory. Second, we put a hundred million dollars into marketing programs to help the channel get its price back on track. Between these two initiatives we took a few months ago, we should be in a good position in Q4 as Ada is hard to drive. I’ll do that with interest. The decision was painful, but they were necessary. It’s six months of hard work. That means hopefully that we can continue.
Q: I wondered whether he could explain why there isn’t an RTX 4070 and what it would be like to buy a 4070. Do you tell consumers to buy a three000-series card instead?
Huang: We don’t have everything ready to roll it out at a moment. One has laid back 4090 and 4080. Over time it’s time to sell other products in the lower end of the stack. But its not as complicated as we usually always start at the beginning, because that’s where the enthusiast wants to refresh first. We’ve found that 4080 and 4090 are a good place to start. As soon as we move further down the stack, we can move on. That is a great place to begin, right?
Nvidia GeForce RTX 4090.
Q: What do you think about EVGA blocking its production of graphics cards from the RTX 40 series onwards? Was Nvidia in close discussion with EVGA?
Huang: Andrew wanted to end the business. He wanted to do that for a couple of years. Andrew and EVGA are great partners and Im sad to see them leave the market. But he has several plans, and is thinking about them for years. I guess this is about it. The market has many great players. You will be good after EVGA. But they will always go missing me. They’re a key part of our history. Andrew is a great friend. He decided not to go to other countries.
Q: What do you say to the Jensen of 30 years ago?
Huang: I’d say so, to follow your dreams, your vision, your heart, as well as to come, you should. It was very terrifying at the beginning, because we invented a GPU, you probably know about it. When we invented the GPU, there wasn’t any GPUs in the market. Nobody liked GPUs. At the time when we came into the world to build a game console, the market for video games was small. She hardly existed. We sat all about video games in 3D, and there weren’t any 3-D design tools. You had to use 3D games to make it right. We talked about a new computing model, accelerated computing, that was the foundation of our company in 1993. That new technology was very effective, but nobody thought it was because of it. I had no choice but to believe in it now. We wanted to make that the company successful. We did that together with all the strength we can.
It was, steadily but surely, that between one customer and another, the partner and the developer became an extremely important platform. Nvidia has invented programmable shading, which now defines modern computer graphics. We invented RTX, and created Cuda, and developed accelerated computing. She led us to AI. We all talked about today. But, it is all, so far, only for every step, so no one knew it. GPU, programmable shading, Cuda, even deep learning. When I introduced an area of learning to the automotive industry, everyone thought it was ridiculous. In fact, one of the CEOs said, You can’t even detect a German dog. How can you find pedestrians? They wrote us off. Exactly now, the deep learning was very difficult, but it was finally completed.
The advice I would give to a young Jensen is to keep him alive. You’re doing the right thing. You must pursue what you believe. You’ll have a lot of people that don’t believe it early, but not because they don’t believe you. It is simple because sometimes it’s difficult to believe. How would anybody believe the same processor used for playing Quake would become the one who modernized computer science and brought AI to the world? The same processor that he used for Portal turned out to be the same one that led to self-driving cars. No one would believe it. Firstly, you must believe it, then, you’ll have to help others believe it. It might take a long road, but this is okay.
Today the previously announced version 3.1 was released by Genshin Impact developer HoYoverse. The new version comes with a new…
Liam loves to keep track of the latest Nintendo news and respect his games library. His favorite Pokémon character was…
It would be crucial that developers make a decision about a possible Ninja Gaiden 4 and outsourcing that development. In…
The noise of an air raid siren made many Ukrainians sad. In most cases, this notification can be heard very…
Last week, London police announced the arrest of a 17-year-old in Oxfordshire, England, as part of a hacking investigation. It…
According to director Fumihiko Yasuda, a release of Nioh on Xbox 360 is still unlikely. With The upcoming film “Ollen…
Asked why Nvidias latest 40-Serie graphics cards cost just one thousand dollars, CEO Jensen Huang said that Moores Law is dead. He explained that the days of constant rising costs are over, with the technology is continuing to slowed and the pandemic shortage pushed things further.