NVidia unleashes Cuda-enabled GeForce 8800

[ndat]

Um... lol, I read all of that and I still have really no idea of what it does or what it's used for.

[Szos]

^^ That's cuz it is all "marketing slang" - a quick translation is simply:
"is be good and fasterer"

I am actually surprised that they didn't use marketing terms like "thinking outside the box" and "paradigm shift" and "synergy".

[DotPainter]

Market speak it is indeed. Cuda is basically Nvidias architecture for implementing the requirements for DirectX 10. Some of these can be seen here (http://www.news.com.au/couriermail/story/0,23739,20654430-8362,00.html). DirectX 10 is designed to bring about a new level of interaction between GPU /CPU and OS. Everyone right now is talking about Vista and the Aeroglass experience and this is just the beginning. Aeroglass is designed around the concept of the GPU being able to take geometry and shader information perform some sort of alorithm on it and RETURN the result to the CPU. This is what makes many call the next generation of GPUs more general purpose like CPUs. But this does not mean that the GPU is a general purpose processor. What it means is that the new architecture of the GPU, as specified by the next generation of graphics apis in the DirectX 10+ generation, will require a more OPEN architecture for the GPU. This more OPEN architecture allows for easier access to the low level internal "core" functions of the GPU, the vertex, geometry and pixel shader cores, allowing output from each to be stored in buffers and chained together to create complex routines called pipelines. Being able to store the result of intermediate stages of processing and create chains of shaders that operate on buffered data and store the result in buffers is what makes this new architecture more general purpose, along with the fact that the shaders themselves can be written in a higher level language like C. The benefit of this is that a next generation DCC programs could have shaders that are all GPU based which create effects in near real time and are able to be saved to disk at a basic level. On a more advanced level, you could render a scene in passes, where each pass is mapped to a GPU routine allowing complex multi-pass scenes and effects to be generated in near real time, with each pass stored separately and combined together using the GPU to produce a result. But dont expect that level of functionality any time soon.

But this is just the first generation of the new GPGPU architecture. In this early stage, things like the geometry engine are somewhat limited relative to the pixel and vertex engine. However, over time the capabilities of the geometry engine will improve, allowing large data sets to be passed to the GPU for complex shader routines to be executed in near real time.

More links:
MSDN (http://www.microsoft.com/indonesia/msdn/wvddirectx.aspx)
BeHardware (http://www.behardware.com/articles/631-1/directx-10-and-gpus.html)

[JoshBowman]

All I hear is "blah blah blah, it's prettier and faster and will cost a lot" :)

[Tripdragon]

All I hear is "blah blah blah, it's prettier and faster and will cost a lot" :)

Same here. And zero developers will use it to make a realtime rendering engine just more games.

[DotPainter]

Actually, it is not so. The next generation versions of some of your favorite renderers, like mental ray are being designed so that they can be run on these new GPGPUs as shader programs. Compositing apps already use some form of GPU accelleration and that will increase. So suffice to say, much of the work in DirectX 10, Cuda and the ATI implementation will make it easier for existing apps to take more advantage of the GPU "under the hood", without exposing a whole bunch of technobabble to the user directly. In one sense, it is good that microsoft pushed GPUs in this direction. It is the basis for CPU makers starting to talk about CPU/GPU cores all integrated together. And if you think that DCC software wont have to change to take best advantage of all this power, then you are sadly mistaken. Or to put it even better, if these apps dont change to take advantage of these changes in the future, then we are suckers for paying for outdated technology. These changes are MOST beneficial to DCC apps, since most OTHER consumer software, including games, has yet to even take advantage of DUAL core processors, let alone multi-core CPUs/GPUs. Bottom line, the new APIs for DirectX 10, Open GL and unified shader cores are laying a foundation that will become very important to computing in the next 5 to 10 years.

[ndat]

Thanks for that Dot, I understand it a lot better now.

[Signal2Noise]

If I see the word "unleashes" in thread topic titles or ads again I think I'll be throwing up. Or is that "unleash some vomit"? ;)

[NOTORIOUS-ro]

Ok.. so I'm here using my P4 1.7 processor 9200 (before had a gf2) and doesn't bother me... that much... This is the first time I'm thinking about upgrading my pc in 5 years. I mean look at the power jumps in the last 1-2 years. I can buy a new, twice as good pc just by cutting of the extra smoke and coffe. Even if this is marketing we can't say that dual cores and dx10 aren't a BIG evolution. As for the prices.. common.. remember the prices on the X2..well I thought those were ufos at that time. DotPainter, thanks for making things easyer to understand... please help out whenever thing like this get us confused. It might sound dumb but I'm not that good with this gpgpu thing. Isn't the gelato frome nvidia a gpgpu thing. From what I understand this helps with viewport feedback by giving high quality image using gelato shaders. I mostly use Max so, for example, I'll have two types of shaders (gpu and normal)? Do these new technologies help in the final render or just with the working enviorment? I've heard of using gpu for physic simulation, sound and pure processing power (like a cpu).

[dmaas]

The C language support is a very big deal to me. I can't wait to learn more about it. This seems aimed at the big problem with using GPUs for non-graphics computations, which is that right now it takes approximately a zillion OpenGL calls to set up the card just to do something simple like add two numbers. It should also finally get rid of the absurd abstractions you need to make to shoehorn general-purpose computation onto GPUs, like pretending that your vector data are actually texture maps or that your matrix multiplier is writing RGB pixels into a framebuffer.

Oh yeah, and it will run games faster :).

Can I please get a passively-cooled version, or at least one that doesn't sound like a B-17 taking off inside my room?

[HellBoy]

wow I see lot of coding talk, especially from you dot and all the text sort of looks like this "%%%%%%%%%%%%%% %%%%%% %%%%%%%% %%%%% %% %%% %%% %
%% %%%%%%% %%%%%% %%%%%5 %%%" and my eyes is like :eek: and in the end I'm like I :love: :deal: :D

[pgp_protector]

...snip...

Can I please get a passively-cooled version, or at least one that doesn't sound like a B-17 taking off inside my room?

That's what water cooling is for :)

[trickypixel]

dmaas is right in that nvidia has a new, simpler interface for using using the GPU as a general purpose parallel processor. One benefit of the new chip architecture that GPGPU applications do not have to use the combersome graphics interfaces (OpenGL, D3D). However, what may not be clear from the description is the new architecture is much more general-purpose and flexible than previous architectures. The 8800 was designed with features that are not required by *any* graphics interface (DX10 included), but are necessary for efficient general parallel processing.

Don't be surprised to see the new nvidia chips used in a lot of non-graphics applications. On the other hand, the new architecture and interface is a boon to those working on ray-tracing and physical simulation so it should enhance graphics applications too.

[Tripdragon]

ALL I see is, give me a tutorial and a card and lets see if i and others can program some crazy stuff out of this new C like language in it...
I want real realtime raytracing

[DotPainter]

The real time raytracing will come, no doubt. However, what is going to be the main sticking point is the fact that both Amd(ATI) and Nvidia have competing APIs that are not interoperable AFAIK. Therefore, if you want to take advantage of the power of these new processors directly, you have to code to the API of either AMD or Nvidia. On top of all that, with Intel, Amd and Nvidia all claiming to be working on integrated GP graphics cores in upcoming CPU designs, it is clear that there will have to be a new instruction set to cover the new features of the GPU/CPU combo. Until someone comes up with a UNIFIED instruction set to cover this new generation of GPUs or CPU/GPU hybrids, it will indeed only be used in certain niche applications.

I just wanted to point out the benefits of such a new architecture, because indeed, if someone wanted to, they could rewrite their render engine to take advantage of these GPGPUs within existing 3D apps. But, alas, things are not that simple, especially when designing programs that run on different OSs and hardware. So, we are just at the beginning and truly most of what is being touted now is pure market speak, for bragging rights to say who was first. Practical use of such architectures in consumer applications may be a bit delayed as the future direction in CPU/GPU processing is made clearer.

Here is a link to AMDs (ATIs) new General Purpose processing API:

http://ati.amd.com/developer/siggraph06/dpvm_e.pdf

[Szos]

^^ and that right there is why you need someone with a lot of clout to step up and 'force' ATi and Nvidia to develop ONE standard. I really see only one company being able to do that - and that is MS (and honestly, even if it is a Windows-only standard, so be it - as long as something happens). The only other company even close to having that much sway, is AutoDesk, but doing this takes initiative and innovation, something that AutoDesk knowing nothing about.

In all honesty, I am surprised and disappointed that this new version of D3D doesn't allow GPU rendering in DCC programs directly. Apps like MAX are already D3D biased, so even if it was only possible when working in D3D mode, that would be fine by me.

[dmaas]

Real-time raytracing is here already. At SIGGRAPH this summer I saw a CPU-only software ray-tracer doing 1-2 fps at 1024x1024 on a fairly complex scene. The developers were getting up around 20fps by running it on a cluster.

[martinw]

There are existing cross platforms APIs to program GPUs. There are some Open Source ones and some commercial ones too, eg for the latter, RapidMind:

http://rapidmind.net/products.php

[CGTalk Moderation]

This thread has been automatically closed as it remained inactive for 12 months. If you wish to continue the discussion, please create a new thread in the appropriate forum.