Nvidia Launches The G80 Graphics Chip; An Interview With Jen-Hsun Huang
Here is a quote apoppin and others are sure to find interesting:
Here is a quote apoppin and others are sure to find interesting:
Jen-Hsun: We don't need to talk about Fusion in order to get excited. We absolutely have more than we can do right now GPUs and will stay focused there. We are just not working on CPUs. There are a lot of things I am working on. Without any exception, I?m not working on CPUs.
Wednesday, November 08, 2006
Nvidia Launches The G80 Graphics Chip; An Interview With Jen-Hsun Huang
Dean Takahashi, 11:01 AM in Dean Takahashi, Gaming
I took some time to interview Nvidia CEO Jen-Hsun Huang about the company's latest graphics chip, a "general purpose GPU" that looks and smells suspiciously like a CPU. But it's really not an attempt to bump Intel aside. It's Nvidia's way to find new markets for the latent potential of the processing power of its PC graphics chips. Read on to hear more.
What is this non-graphics oriented processing that can be done on the G80 and the notion you can assume more responsibilities from the CPU. Can you explain that?
Jen-Hsun: Andy Keane who is the general manager of our computing business unit. His job is to work with architects create products and to take to market to serve the need for high-performance computing or general purpose applications. You?ve seen the part of the presentation that explains that the G80 is the worlds first DX10 unified architecture. We used to have geometry processors that were in the CPU and vertex shaders that were in the front end of the chip, etc. We?ve taken all that functionality and created one unified processor architecture. It?s a scaler architecture with a lot of floating point performance and energy performance, 128 processors running in a chip. Altogether, 3 Teraflops or so. All of these different processors could support a whole bunch of threads themselves and run concurrently. There are thousands of threads that run inside chip concurrently. We created a programming model of a chip supported by a complier and the model is a language extension of C. Architecture for interface is called ?CUDA.? It?s a general purpose programmable processor that is optimized for very heavy weight, large amount of data and data-parallel applications. So those are the high-level bits.
How do you answer whether or not you can also do non-graphics related processing?
Jen-Hsun: That?s right. We are really good at applications and programs where the instruction complexity is not extremely high, but the data is enormous. Gigabytes of data and moderate amount on construction complexity. Whereas the CPU deals with infinite amounts of construction complexity. Almost any lousy programmer who could write some C code to the extent that you could compile, could pretty much run that program. A GPU or G80 with compute, you?d still have to be moderately thoughtful about how you?ll structure/factor the program. If you structure/factor it right, the amount of throughput you get out of it is extraordinary. At that point you can utilize the 3 Teraflops of floating point performance inside the chip.
Andy: If you look at what applications naturally divide between GPU/CPU, Jen-Hsun talked about varying workloads, irregular data, structure that exist is operation systems and applications so that naturally migrates to CPU because that?s what it was built to do. There are another set of problems where data is large, relatively parallel. That?s where things will migrate to the GPU. An extreme example of that is in a game. You have advantages in game, AI, input, a lot of these inherently disconnected or random functions and then the graphics processing extending into geometry and physics now migrating to GPU because it fits architecture well. Take it to next step, and find elements of code or anything that?s decomposed into irregular data--this will find a home on GPU. Architecture now maps problem.
When you a running a game on the G80, what do you need the CPU for? Is it less necessary than it was in the last generation?
Jen-Hsun: Creation and destruction of geometry was never done before on a GPU. With Direct X10 and GeForce 8800 we can now create and destroy geometry on graphics processor. Geometry shader is a whole new class of applications that can move from CPU to GPU. That has a number of applications from hair rendering, shadow rendering, etc. A number of key graphics algorithms have major components that deal with geometry migration. And as a result, a number of games have bottlenecks because of this you still have a number for things like AI, database management that happen on CPU, it is still an important component. But now the GPU is no longer bottlenecked by the CPU in terms of physics effects, geometry, and shading.
Jen-Hsun: Now its boiled down to two things ? control and processing. Processing will go on the GPU. Control will remain on the CPU. More things you can Control. Huge things amount to control. Games are richer, faster, more alive.
What?s the state of tug-of-war for CPU and the GPU given AMD?s acquisition? Reports on the Inquirer suggest that Intel is doing stand-along graphics chip. There are reports of Nvidia using the former Stexar team to create something more CPU-oriented with GPUs. Given all of these things, what is the state of the tug-of-war?
Jen-Hsun: We don't need to talk about Fusion in order to get excited. We absolutely have more than we can do right now GPUs and will stay focused there. We are just not working on CPUs. There are a lot of things I am working on. Without any exception, I?m not working on CPUs.
But the definition of what is what is now changing. I guess that?s the reason for the fog.
Jen-Hsun: Only thing that changed is the AMD acquisition of ATI. The GPU is about to go through a revolutionary time. That more than ever before the GPU will be perceived as a general purpose processing unit to help solve vastly complex data problems. We happen to think there are all kinds of applications from image processing to computer vision to high-performance computing, to stock option price calculations to physics simulation to geometry processing. So many interesting problems that the industry hasn?t tackled. We?ve finally created a revolutionary architecture to tackle that in a thoughtful and elegant way. And a way that is consistent with the programming model that they understand, which is C.
There are some things moving around with the PC platform, with the Intel Geneseo project and AMD with Torrenza platform as well. How do you build something that will fit in with one of these schemes?
Jen-Hsun: The two architectures are very different at the lower-levels. But at the conceptual level they are very similar. Makes a GPU share for the first time coherently the memory system of the computer. We love the fact that finally the GPU memory system is coherent to the CPU system. Now that the GPU is a co-processor to the CPU its a fabulous advancement of where people see the important of CPU computation. Its a recognition that the GPU is becoming a more important part of general purpose computing. Whether it?s hyper-transport based or PCI-express based, we are neutral.
Is the GPU average selling price going to go up relative to the CPU?
Jen-Hsun: I don?t know. The market will have to decide that. One observation I?ve made ? the GPU has been the only one to increase consistently over 15 years. The GPU has an ASP of $17 when NVIDIA first started. The lowest price was $5. In the course of last 10 years, the ASP of the GPU has gone from low teens to low 30s. Don?t know of any component in last 10 years gone up by factor of 3. Or up at all.
I noticed when I did a check with a few folks that the CPU ASP is still higher than the GPU ASP.
Jen-Hsun: It sure is. The CPU ASP, when you weigh in the server price, is in the thousands. And there are 6 million servers built each year. That shifts ASP of CPUs. I?m not sure what the ASP in exclusion servers, I think it?s below $100. But nonetheless, certainly still higher.
Is this an important thing to watch? To see who is winning?
Jen-Hsun: Not really. The important thing is that we continue to add value to GPUs, and hopefully the value will go up. It?s not important to me who wins. NVIDIA has always been a GPU company. We really believe our purpose in life is to transform the computing experience and solve the most complex visual computing problems. That is our purpose. It does not exclude AMD or Intel. We support both microprocessors. CPUs are a good thing. It also helps us achieve our purpose. Intel and AMD have a war going and we are not part of that war. We are friends with them.
Windows Vista will drive more adoption of GPUs from your point of view?
Jen-Hsun: If you look at OS10 as a surrogate for the future of Vista, people will realize that OS10 laid down the user interface. Things like the windowing system, icons, went to 3D, had drop-shadows, etc. PPTs, spreadsheets, iTunes will go to 3D. Everything goes to 3D once you have the basic foundation of 3D. The important thing about Vista is that the foundation is in 3D. All application developers from Adobe and iTunes ? various apps will go to 3D without much concern.
DirectX10 making GPGPU possible. What does Microsoft think?
Jen-Hsun: Microsoft sees, and I?m only interpreting from the movements they make, the next wave of apps are heavily consumer oriented, heavily weighted on the experience of the consumer. They believe that the intersection of Image/graphics/database processing could enable a new class of applications. One of my favorites is called Photosynth ? do a search. It is an extraordinary application. It is taking photography and 3D graphics and merging them together.
Is this what they talked about last week?
Jen-Hsun: Possible, I wasn?t there. Search Photosynth, we are working with that group. Dragonfly is the code name for project. It?s really incredible. Definitely the next-level. You can do that without intense image processing to correlate control points and photos and navigate it in 3D.
I guess at 3 Teraflops, you are at least 50 percent faster than the PlayStation 3?
Jen-Hsun: It?s basically bigger. Cell has been around 2- 3 years. It?s later in time so not surprised it?s more.
So that means the PC has come back to the forefront in graphics?
Jen-Hsun: Yeah, I don?t think anyone denies that. It?s expected that PC evolves and advances on 6-months cycles. And consoles advance on a 5 year cycle. In very beginning, game console is faster and more powerful than the average install base of 150 million PCs, quite an achievement. Over time, the PC continues to evolve, and you sell over 250 million, over next few years, PC becomes more powerful than game console again. Programmers extract every last drop out of game console. It gives consoles a factor of 2 advantage, without software overhead, using high-level tools to write programs. PC has advantage over the game console.
So with the unified shaders ? ATI went with that architecture with Xbox 360. Do you have a different take on it?
Jen-Hsun: They went to a unified shader architecture that unified vertex and pixel shaders. We are a unified processor architecture, we unified 3 types of shaders: geometry, vertex, and pixel shaders. In addition, our processors are C-programmable processors. It?s a radically different thing. G71 had shaders and were programmable, but they were not programmable by C. G80 is unified and programmable through C.
What would that mean on a strategic level?
Jen-Hsun: The fundamental revolution of G80, is that we are making programmable element, the processor accessible for general purpose computing. GPGPU means you can still program triangles and texture maps in a clever way, numbers you read out of frame buffer are answers. To teach someone to do that is weird. Only a few researchers around world can do that. With G80, we?ve given it a programmable architecture similar to kind people can understand.
It sounds a lot like the GPU is contending with CPU. It?s taking on more of CPU?s functions. Is there a semantic problem here?
Jen-Hsun: We are not taking on the work of the CPU. We are doing the work that the CPU could do before. It?s not like people weren?t using X86. We are doing the work that wasn?t being done before in a satisfactory way. Right now I visualize that there are two axes. The vertical axes is the CPU ? designs for any instruction level complexity. It needs to be syntactically correct for the CPU to run it. Doesn?t matter if its? written well. The CPU will compile it. The infinite levels of instruction level complexity and frankly the amount of data complexity; as long as you can fit it in four vector it?s happy. Amount of floating point is 1- 2 issues per pop cycle. So not a whole lot of data complexity, but an infinite amount of instruction complexity. In the case of G80?s computer, we have a lesser instruction complexity capability. You could pile on gigabytes of data and we?ll stream right through it. So the two axes ? one is data level complexity, and one is instruction level complexity.
What?s your thought about the low-end?
Jen-Hsun: (Integrating for the) high-end doesn?t make sense. Too much innovation left in the GPUs. Integrating worst of either worlds makes high-end product completely unsuccessful. On the low-end the question has got to be asked ? integration of two different chips. The Southbridge chip is another chip. The CPU uses fewer transistors, but fast ones as you know. GPUs don?t do that. I?d rather have lower-leakage transistors. Frankly that tends to be quite similar to Southbridge. It?s either integrated GPU or CPU with MCP. In the low-end no one cares anyway.
AMD is making the argument that it can cut costs out of the system.
Jen-Hsun: Yep, but you still have two chips, unified memory. Maybe they?ll save a dollar, but they won?t change the world.
On the G80 itself, how long did it take?
Jen-Hsun: Four years. We built is ground-up brand new.
Do you consider it late?
Jen-Hsun: Yes, we tried to get it out last year. But it was just too big. So it?s late by our own standards. In the end, it cost us 4 years. It was important to get it right. There were 600 man years total. We started with 10 people working on it and grew to 300 eventually.
Competitively, the G80 can get you ahead for a while for what reasons?
Jen-Hsun: From an architecture perspective. It opened up scope of apps we add value to. It?s the first GPU where the application is larger than graphics processing, with physics, high-performance computing, and image processing. We are working with folks like Adobe and Apple. We are bringing image processing to the desktop. It is so important to us. Ultimately, the applications where we add value to will reflect what kind of company we are. Up until now, we were known for just video games.
And what about power consumption?
Jen-Hsun: It?s two times better than the last generation.
What?s the actual wattage?
Jen-Hsun: 177 watts. The power per watt of architecture is twice as good as last generation. It is all about efficiency. If I know that architecture had better performance per watt, reducing power is easy.
I saw the frog demo ? what will this enable? The quality of games?
Jason: The stream output will come through to the geometry shader. The output of data into memory without going through the process pipeline. Take data back in for much more advanced shader calculations and effects. Hair and shadow algorithms leverage technology for much higher quality of effects. Deformation aspect ? using objects to actually deform objects. Mix that up with geometry shaders, and realistic images. Rigid chunks of triangles, etc? will go away. It will create a world that is really soft and malleable and stretchable. When people animate faces, joints, and hair it will be soft and flowing.
Are you falling in love with the female characters?
Jen-Hsun: I?m still in love with the first three. I love them all.
Is there anything else? Will waiting for Vista impact you?
Jen-Hsun: Microsoft has programs that allow for upgrades. That should help keep the momentum.
What?s your favorite PS3 game so far?
Jen-Hsun: Metal Gear Solid is the most beautiful to watch. I like beauty, but also games that you can learn in 30 seconds. I won?t learn Metal Gear Solid in 30 seconds.
I want to play a game with you one of these days.
Jen-Hsun: Let?s do it!
Facebook
Twitter