Great. On another subject completely, how does memory certification work? Is it Samsung that does the differentiation by testing it better or?

The way we view it is that we take a board that we manufactured and run it through burn-in. […] The memory is the same, but we use faster RAMs than our target. So if we got a 1GHz RAM, we'll run it at 800MHz.

Another difference with GeForce actually would be the TDP. The numbers I've seen are actually lower.

The computing TDP is lower because of triangle setup, texture and ROP. We don't use raster operations.

You've got a GeForce TDP, a Tesla TDP, and a CUDA-only TDP?

So, hmm, we actually don't. The TDP on the chip is the same on all parts. We use exactly the same chips. There's no extra binning right now but actually we will be binning differently for the systems products. […] [What's lower is the CUDA-specific TDP].

Finally, CUDA is becoming more and more central to the company, so I don't know how you look at that in terms of company strategy and how that affects Jen-Hsun's plans etc.?

So becoming programmable, that started a long time ago. What really happened is the workload even on the graphics side became more shader-centric.

But going forward… Does it have more of a direct impact on the roadmap than one year ago maybe? One year ago, you always talked of CUDA as a HPC thing exclusively, nothing about Consumer CUDA. That's the big difference.

Hmm, we're a very forward looking company. The major difference is that if we had told you one year ago that this technology was going to have an incredible impact across the board on everything we do, you probably wouldn't have believed that.

Hah, well if you had told me one year ago that we would have got H.264 encoding with low CPU utilization, everything offloaded [to the GPU], I would never have taken it seriously.

Yeah, so the way we plan products and the way we market products, they're different. Including a computing element in a GPU, we had to plan that a long time ago.

But isn't the roadmap a little bit different? I mean the ALU:TEX ratio for example?

No... no. You have to have… even a brand new GPU, it does take a while to design. We have to decide [everything very long in advance].

Hmm, and in the long-term then, in two or three years… are going to make Tesla-only chips?

Not at the moment, not at the moment… And the basic reason is the amount of R&D dollars that it takes to make a special chip, the manufacturing side (being able to amortize the manufacturing cost over many many chips)… Today, there are [nearly] no special chips [in the HPC industry][…], and even I would say Intel… their major chip is Xeon, which also shares a lot [with Core].

The market just isn't big enough in terms of volume right now?

Exactly, that's exactly the point.

But the point also is that the current GPU business isn't as big as the CPU business though. So if Tesla grows a lot…

You know, that's a business decision. […] It could be an incredible device, but [even then] it wouldn't have the same economics. It wouldn't have the same costs.

[For the next & final part of the article, click here]