So you must be happy to split things up like that, where DP is only available on Tesla or certain Quadro and GeForce doesn't get that. You're happy with that?

Yeah, that maps to our product stack and business and markets that we're in. Professional customers don't buy mid-range boards.

Well that's similar to one of the worries we had when CUDA showed up, that you'd just limit the software to running on Quadro only and a whole bunch of people with G80 boards couldn't do anything with it, so it was nice to see NVIDIA let CUDA run on anything.

Right, CUDA is knit in to the fabric of the GPU, so it's not just one piece off to the side on the die that we can just add in or take away. But those things that make sense to limit to certain markets, like double precision, those are pieces that we can modularise and take on and off of a GPU. There will be some things there in terms of roadmap features that we're not talking about yet, but we've announced DP and it's coming so that's one thing that'll appear in that fashion where it's market specific.

Is there a chance we'll see Tesla be the first GPU product line where a new GPU will debut, because of that? Could the GeForce brand not be the first product to grab a new chip?

It's possible, but in general the roll out process in the professional market is much longer than in the consumer space, so there's integration and qualification and other processes that don't necessarily apply to GeForce. The cost of not having something working in the IT space when you come to deployment is too high, so introduction takes longer, and these are still graphics processors so GeForce is always likely to go first.

Thinking about that then, where a current GPU has area dedicated just to 3D rendering, with fixed-function blocks like texfilter or the ROP. But you can see a future where the ROP is just another bunch of programmable ALUs or not even a separate bunch, and the same with almost all the other fixed units that have a need for high performance math. So almost the entire chip is just a big FP machine and there's no big area for fixed-function units any more.

I don't think I can speculate things will go in that direction!

Right, but if you imagine a chip like that where it's just massively suited to HPC with almost all the area dedicated to generating programmable flops, you could see a chip like that hitting Tesla first?

That's something I couldn't speculate on, because the heart and soul of NVIDIA is still one that serves the consumer graphics market and it's maybe unlikely that you could build a chip so dedicated to FP and still have it do as well in that space as a chip with more area dedicated to graphics-only tasks. So there's a centre to the company and we add applicable markets around those efforts, and the centre is still 3D graphics acceleration and we have to serve our core markets.

Some physical stuff then about the products you're about to ship. You're using external PCI Express cabling, and we notice you support PCI Express Gen2 with that, but has the external cabling spec been finalised for Gen2? We're thinking that it's just Gen1 that has a finalised spec for external there.

I think things are good for Gen2 there and we've run tests for that.