NVIDIA DirectX 11 Questionnaire – A Response to AMD’s Views

[Final8y]

AMD and Intel could for example disable support for Nvidia graphics cards for their processors. After all, the CPU and GPU is interdependent and it’s hard and expensive for them to test their processors with all possible configurations involving Nvidia cards. I’m sure they want to be able to ensure a great experience for their own customers, and the only way I can see that happening is to disable Nvidia hardware. It’s good to know that Nvidia will support such a decision completely.

 

[NickK]

nVidia buying Mental Ray to me says that they're doing R&D on rendering on the GPU using compute shaders.

They are.

Shader 'fragment' programs are basically massively parallel computations. The main instruction in a GPU is result:=A*B+C (multiply-and-add). Except you provide 2D textures to it. The real way this is done is that you load your X,Y,Z into the R,G,B of the pixel as non-IEEE single point floating 'clamped' numbers 0.0 to 1.0. Although double point or extended point (IEEE) is was really required for the main applications such as computation rather than games. Also when double did appear it was 1/2 the speed..

For example - an old X1950XTX 512MB GDDR5 card can process 120GB/sec of data with multiply and add. Why is this important/useful? Well it's a dot product, it's a matrix multiplication... The slow part in a GPU is the offloading via PCI-E to the memory at 1-2GB/sec because the upload game textures was historically more important than the offload into main memory hence offload being much slower.. Memory is about 12GB/sec.. Add to that the GPU could only upload/download whilst the card was not actually processing/rendering meant there were some hard changes required in the GPU memory controller to get the system actually working efficiently.

The new 'compute' cards will be quicker as the hardware is evolving to support the requirements of compute rather than games. Hence Fermi's support of double fp and proper IEEE error handling.

Apple have Renderman, so it will be interesting to see nVidia's choice in creating nV only software to create rendering cross swords with their GPU client.. I would suspect that nV are producing a portfolio of GPU-enabled products, starting with this because it'll be of use to game studios... if they decide to go with nV for their product.

Don't make the mistake - there IS a compute market out there. Especially as the basic x86 desktop machine has really reached it's limit and HD/3D is creating a multimedia processing void. I think the mistake is the over estimation of gamers wanting to be tied to nVidia to play their games because all the technology in the game requires nVidia hardware. That's too much of a lock in - in fact it smacks of nVidia producing it's own games console for the mass market and then using the same technology for high power computing applications.

So what does everyone thing the nV console will be called? nV, nX720/1080, ..

 

[NickK]

Additional.. I think you may find that Fermi appears in the PS4 because sony need something to outshine MS and Sony's native market will pay premium for something this bleeding edge - especially if it provided 1080p 3D gaming and possibly mainstream media in 3D etc.

I think nVidia are going to go through a lot of pain. Innovation as a new market is risky (I know I've done that ), especially when you've stretched your neck our as far ahead of the current gamer market as they have. They have vision, it's just they need to make sure that mine field they're walking through isn't going to kill them first.

 

[carrot1401]

To be honest, while I was there, there weren't many people talking to him.

[ZiiP] Carrot was there while I was talking to the nVidia guy, but he was moving about a bit.

Most of my talk was one-to-one with him.

I kept out of the way, I'm a student so just smiled at all the pretty free things

 

[BoozerDawg]

Sounds like the Nvidia guy had some fun .