The original printed, but somewhat shortened version was first published in issue 05/2007 of the german PC Games Hardware magazine.

Here you'll find the unabridged and uncensored interview transscript of PCGH with Vijay Sharma at Cebit 2007. It was a freely held conversation - no scripted questions and answers. So please bear with us the casual manner of speech and grammar. The Interview was conducted by Carsten Spille on behalf of PC Games Hardware, all rights reserved. (2007)

PCGH: Hi Vijay. AMD decided not to launch R600 officially at Cebit. Can you give us a clue, what was your reason to do or not do so?

Vijay Sharma [Director of Desktop Discrete Product Marketing at AMD]: Sure. Our initial plan as you're probably aware was to do an NDA press media event this week just before Cebit at Amsterdam. And we realized that we had two derivative products of R600 coming. So we had a whole family of R600-products coming within 45, 60 days of each other. So we said rather than have the press come to Amsterdam before Cebit and then another month later do another event lets talk about the whole family at once and include desktop and mobile products as well. So we have a very strong story and we choose to hold on the R600 and have a combined event later in april.

PCGH: So that means the reason for the delay is not R600 not being ready...

Vijay Sharma: No, no. The R600 schedule hasn't changed. So we're building boards, we're on the same hardware-schedule that we've always been. We're just not launching the product until we launch the family.

PCGH: Given the problems your competitor has with Vista drivers what is your feeling about your Vista drivers for R6x-Generation.

Vijay Sharma: We are very confident in our Vista drivers. We had our first Vista driver that was WHQL-certified I believe in November...

PCGH: So even before Vista...

Vijay Sharma: Even before Vista we had WQHL-certification. So our driver is in great shape. And the other point is we are actually on our second generation of DX10.

PCGH: You're talking about Xbox360-GPU...

Vijay Sharma: Yes, Xbox360 was our first generation and we're on our second generation and you can see our lead in the drivers. This isn't the first time we're doing this. We have a lead, we have established technology, so we're very confident.

PCGH: A couple of days ago there was a blog entry by one of the FSX developers stating that they obviously had a R600 to experiment with and that it's geometry shader performance was much higher than G80s. Can you comment on that?

Vijay Sharma: I haven't heard of that specific instance or example...

PCGH: He didn't give any numbers...

Vijay Sharma: No, but you know even in our [...] I can tell you in our X1000 series architecture our predicted branch performance in shader model 3 is seven times more efficient than our competitors. So we've done things in our architecture (and we continue to) that allows for maximum efficiency in executing shader code. And so I'm not surprised that the geometry shader is working very well in R600.

PCGH: So are you hinting at that branching performance or granularity is somehow another key to the speed of geometry shader or was it just an anology.

Vijay Sharma: Oh, it was just an anology.

PCGH: Okay. What could possibly slow down a geometry shader. Were could save transistors to support geometry shaders but to make it slower. Is it a matter of cache size or... Because since you've got a unified shader architecture every shader has to have the same capability - so it's not possible that you've just got fewer geometry shader units.

Vijay Sharma: Honestly I don't know how it's going to play out. Because Geometry Shader's - i believe - is fairly new construct - it didn't exist in DX9, does in DX10. And I think we'll have to see and wait for ISVs to use geometry shaders extensively and then we'll see were the limitations and where the advantages are. So i don't know what it is today.

PCGH: Early drafts of DX10 proposed something like a programmable primitve processor. Did you have any plans to include it or how advanced were your hardware design when it was cancelled out of the specification.

Vijay Sharma: Oh, I don't know in terms of that details. And even if I did know I could not tell you at this stage - because it's before launch.

PCGH: On another matter: Your R5x-GPUs were quite superior to Nvidias GPUS in terms of image quality. G80 has much improved on this. Do you plan to improve image quality in terms of anti-aliasing and texture filtering - I'm not talking about DX10-effects - but in terms of FSAA and AF- over R580.

Vijay Sharma: Oh yes, there'll be those improvement put in. But mind you the trend we're seeing thaht with higher resolution the benefits of AA are diminished. So if you're running 1.920x1.080 the effect of 4x or 8x AA is less than if you're running 1.024x768.

PCGH: Of course. But it's still visible. But nonetheless: Many people still think "hey, i got a 16x12-display and i don't need AA anymore."

Vijay Sharma: Sure, but there's a tradeoff. If you take the performance penalty...

PCGH: You're talking about the law of diminshing returns.

Vijay Sharma: Yeah. So i think from a trend perspective we're not gonna see, we probably won't in the coming generations 32xAA. You know even today with crossfire we support 14xAA and that's continue to do that.

PCGH: You told me about a different between 4x and 8x AA - you didn't mention 6xAA...

Vijay Sharma: Or 6x - i was just using multiples of 4...

PCGH: OK. Rumors are that you are using, ok actually we've seen already some slates here lying on the floor stating that R600's using GDDR4 - not being used obviously. But rumors are you are using very high clocked GDDR4 memory on R600 with a 512 Bit bus width. You would get massive amounts of bandwidth, about 150 Gigabyte/s with that. What are you doing with that bandwidth. It is higher requirements of DX10-style applications or are you just trying to be on the safe side never to be bandwidth limited with your expensive chip.

Vijay Sharma: I don't want to get to far into the architecture at this stage. But certainly as you increase the compute capabilities of the graphics engine you'll need more bandwidth. We think that our competitors who have a 384 Bit interface they are probably going to be bandwidth limited given that they're building with GDDR3 - I don't think they have a GDDR4-design - so i think they don't..

PCGH: They state that G80 supports GDDR4, they just don't have a design ready.

Vijay Sharma: Yeah, they don't have the design - which is fine. But i think the 512 Bit inteface gives you a tremendous bandwidth and so we think on the high end the last thing you want is to be bandwidth limited.

PCGH: But a broad bus interface and high speed memory - GDDR4 - is a cost factor and it would make R600 more expensive than if you were using GDDR3 and a 256 to 384 Bit interface.

Vijay Sharma: If you had a 512 bit interface doesn't mean that you can't use just 256 bits. And you can use different memories. So from a bandwidth perspective we can certainly match or exceed anyones bandwidth with the architecture.

PCGH: Are you sticking to the ring bus memory controller concept? After all you've put much effort into it in R5x-architecture.

Vijay Sharma: In the 5x-architecture we have the ringbus memory controller - there have been some changes in the 600-family.

PCGH: ...obviously for the better, I guess.

Vijay Sharma: You'll have to wait for the details...

PCGH: On another matter: Nvidia stated that the G80-Chip in it's current form is not compatible with their AGP-Bridge - maybe future chips are again, but G80 will not be. Can you tell us about R6x?

Vijay Sharma: We have a brigde chip and it works with 600 family.

PCGH: So there's no reason it should not work.

Vijay Sharma: No.

PCGH: So it's up to the board partners again to build AGP cards just as they do now.

Vijay Sharma: Yes.

PCGH: Power consumption. We've seen the photos on the internet of the so called OEM-designs of the R600, we've seen the photo in the press conference of your teraflop-box. The coolers look very very massive. Is this to be on the quiet side of thing or do you just have to get rid of massive amounts of thermal energy.

Vijay Sharma: Both. You know, there's no surprise that the next generation - whether it's G80 or R600 - these are big chips. And you saw the two 600s put out a teraflop. And they don't get that at low wattage. You need power. They do consume power and we have air-cooled dual-slot cooling solutions for the R600. And we've designed it such that we meet acoustic requirements and also provide the cooling that's required. We will have third parties making liquid cooled solutions. I think in the high-end and in the enthusiast space we're going to level off in terms of power consumption.

PCGH: So you don't expect power to go much higher than R600/G80-level.

Vijay Sharma: Well certainly not from AMDs perspective - I don't know what Nvidia's going to do - but certainly not AMDs perspective I think you're going to the a levelling off of power certainly in the enthusiast space. Some people have asked will power consumption come down in the enthusiast space. It can because we can go to 65nm where the per gate leakage is much lower. Or we can go to 55 and employ different techniques in the process. But in the enthusiast space the game is to deliver maximum performance.

PCGH: No matter what cost...

Vijay Sharma: No matter what cost - well reasonable. But what we do is we say this is the power budget for the enthusiast space. Use every last watt and deliver performance. That's why I think it will level off.

PCGH: Where you considering in some point in R600 development that air cooling would not be sufficient anymore. Because there were rumors there were early that where by design... [broken]... They had some problems with power leakage and they where rumored to be above 300 watts.

Vijay Sharma: Wow - 300 Watts I don't recall. But what I can tell you that we have a whole thermal team in Markham in Toronto and what they do is develop cooling solutions for the graphics products.

PCGH: So you have dedicated specialists. You don't go to Taiwan to the "cooler market and go like "hey, give me 500k of those"...

Vijay Sharma: We do everything. I mean, we have a dedicated team...

PCGH: I mean for high-end.

Vijay Sharma: The purpose of this cooling team is to look at all possible ways to cool GPUs. We've looked at liquid cooling designs, we've looked at air-cooling, we've looked at vapor-chambers. We've done Heatpipes. We've gone to third parties around the world who work specifically on cooling solutions to see what they can come up with. We actually want to take the best in innovation and put it on the product.

PCGH: It was the first time since - I don't remember when - when we saw G8800 GTX and GTS having such a large discrepancy between the top-model and the second in line. Do you think it was a mistake. Do you plan to take advantage of that big gap that's opening up there to introduce a intermediate model with your fastest model being fast than GTX and your second fastest model being almost as fast as GTX.

Vijay Sharma: I don't want to talk about what we're going to do on those products. Soon we will announce that too. But I think you've discovered that the graphics market in the enthusiast space is very wide in terms of pricespan. So it really goes from 349 or 399 Euro-dollar all the way up to 599 and the volume at 599 is very small - the market is very small. And I think our competitor realized that. So the key in enthusiast graphics is price-performance. So I think on a price-performance-basis the R600-series is going to be very strong.

PCGH: Given the fact - as I take it - that the R600 is produced at 80 nm. How much of an advantage does that give you because after all it's just one-ninth from 90 nm. So how much of an improvement do you think that gave you.

Vijay Sharma: In terms of...

PCGH: Overall - transistors fitting on a given die size, power...

Vijay Sharma: Yeah, I think the big advantage that 80 gives you over 90 from a process perspective is transistor density. You just get more and the chip will cost less for a given transistor count than it would at 90.

PCGH: But did you go for less cost or maximum transistor count?

Vijay Sharma: In the high-end, methodology is when we develop the chips, we do go for cost target. So we specify a cost we need, and then the engineers need to put as many transistors and as much performance into that cost budget. And although a lot of people say that it's "Deliver the highest performance no matter what there's always a "matter what - whether it's cost or power or something else. So we do do that. So the choice of 80 was aligned with what we need to deliver and it did give us a cost advantage over 90. So we choose to go that way.

PCGH: You felt it was worth the hassle compared to the proven process of 90nm?

Vijay Sharma: Oh the R600 wasn't the first 80nm part we've done.

PCGH: But I'd bet definitely by far the most complex...

Vijay Sharma: O certainly it was the most complex. But from a process perspective we have not had 80nm process issues on the R600.

PCGH: One question about performance: I know you can't answer to any details, so I'll be very vague about it. Did you hit you initial clock target for R600 or did you even exceed it?

Vijay Sharma: Oh, I think we actually exceeded the clock target that we had. So we're quite pleased with the results.

PCGH: Another word on the family of R6x. You plan to introduce all products at once you said. Does that include mobile products?

Vijay Sharma: Right now we're looking at a whole family of products. So there's a whole set of R600-class mobile products and desktop products. And so we do one big tech event for the press under NDA. And then following that will be the announcements, formal public announcements for the products.

PCGH: So the press event will cover all products and the mobile devices are going to launch a little later?

Vijay Sharma: Yes. Well they're very platform dependent.

PCGH: Ok, one of the last questions... The RV530 was in terms of performance was in terms of performance and functional units I think exactly a fourth of R580 and you put a later chip out - RV570 - to cover that large gap that opened up there. Do you think that R6x-family is a little bit better balanced. Or didn't you feel that it was a mistake at all to make RV530 that weak compared to RV570?

Vijay Sharma: Your comparison is accurate. But we don't look in our product line that way - to say that a 530 is a quarter of a 580 - which is true. What we look at is that the 530 right now is selling in the market for a 256 meg part for 89-99 Dollars or Euros. Which is a wonderful price-point and it's selling very very well. The 570 is substantially more architecture and it's selling at 199 and it's selling very very well. The key really is to make sure that you deliver the right amount of architecture, the right amount of 3D at the right price point. And then it comes down to timing. Because not all things necessarily will be available at the same time. And so the 570 came after the 580 and the 530 so yes: There was a gap at 199 and the 570 filled that gap. In that respect, looking at the product plan, now everything is correctly sized.

PCGH: What price range are we looking at at R6x-launch? From 99 to 599?

Vijay Sharma: I think I'll say that it'll cover the full spectrum of price points so you'll have the best DX10 available in the value-space, mainstream, performance and enthusiast.

PCGH: But not integrated graphics on motherboard chipsets?

Vijay Sharma: That's not planned at launch. Not at this launch.

PCGH: Rick told us on the press conference that you had HD-Audio and HDMI integrated into the RV6x. Is that true for all R6x-products from top to bottom?

Vijay Sharma: Yes.

PCGH: So I guess that's all I wanted to know - well, all I could reasonably expect you to give me an answer to..

Vijay Sharma: [laughs] We have to have an excuse to see you again in a month, so...