Future Graphics Products
X-bit labs: It is obvious that it is crucial to be the first with introduction of next-generation products. ATI was first-to-market with DirectX 9 generation and won the round, whereas Nvidia was first to introduce DirectX 10 GPU and won this time? Are you confident that you will be the first to launch DirectX 11 GPU?
David Cummings: Currently, we are the only vendor supporting DirectX 10.1, another first for us. As all features of DirectX 10.1 will be supported in DirectX 11, many developers are starting to develop for DirectX 10.1 as a preparation for the eventual move to DirectX 11. We feel very confident that we will make that transition to DirectX 11 right along with the developers.
X-bit labs: ATI has explained the reasons why it decided to switch to multi-GPU strategy for high-end market a number of times. But modern multi-GPU configurations are homogeneous multi-chip setups. Perhaps, it makes sense to look at so-called heterogeneous multi-chip configurations (e.g., one chip does load-balancing, other chip computes, etc)?
David Cummings: Certainly we are considering our options with respect to heterogeneous computing, however we have not made any decisions at this point.
X-bit labs: ATI has been heavily criticized for its 3:1 and then 4:1 texture to arithmetic performance ratio for DX9 and DX10 generations of products. But can we expect 5:1 or even 7:1 ratio for DirectX 11 generation of products, considering the fact that DX11 will have more computational power-based features?
Dave Nalasco: There is always some guesswork involved in determining the optimal configurations for GPUs, especially when programming models change and evolve. That said, we do an extensive amount of up-front simulation and characterization, and that is something we are improving with each generation. I think our process yielded some great results in our most recent designs, and I expect that trend will continue in our upcoming products.
X-bit labs: What levels of performance (in TFLOPs) should we expect from GPUs in the next 2-3 years? We do remember that Phil Hester (ex-CTO of AMD) promised petaFLOP performance in the next decade :).
David Cummings: More :).
X-bit labs: What do you think about external graphics cards? ATI and Fujitsu Siemens have launched one called Amilo Booster, but it has numerous limitations and is only supported by one laptop by one vendor…
David Cummings: I think the Fujitsu implementation of our XGP technology was a very good beginning for a technology that has tremendous potential.
X-bit labs: Do you think that sometime in the future there will be external add-in graphics boards for desktop?
David Cummings: The notebook is the logical opportunity for external graphics as it has the dual need of being highly power efficient when on the move and yet having the same demands as a desktop placed on it when in a home or office environment. I could possibly see a desktop utilizing external graphics as a way of incrementally increasing graphics performance.
X-bit labs: ATI has been seriously investing into new memory technologies, which helped the company a lot back in the days. Will you continue to invest further into the evolution of GDDR? When can we expect the GDDR6 standard?
David Cummings: We are still actively involved in the development of GDDR5 memory technology. The payoff of that investment is clear in that we were the first (and still the only) provider of GPUs that take advantage of this standard. GDDR5 is still early in its lifecycle, and we believe there is plenty of headroom available to continue improving its performance and power consumption before a new standard is required.
X-bit labs: What are the reasons why you decided to drop ring-bus memory controller? Can we expect its return in the future?
David Cummings: The 512-bit ring bus interface used in the HD 2900 required a lot of room. Returning to a smaller memory interface allowed us to keep the overall size of the HD 4800 series down, an important requirement with more and more HTPC and SFF-based systems on the market.
X-bit labs: ATI advertised the “Sideport” communication bus on the ATI Radeon HD 4870 X2, but the technology is still disabled. Can you re-emphasize potential benefits of Sideport when enabled? Can you reveal the reasons why it is still turned off and when, if ever, it will be activated?
Dave Nalasco: The ATI CrossFireX sideport provides an additional communication path between GPUs, with the potential to increase the available bandwidth. Improvements in our CrossFireX drivers have since reduced the amount of bandwidth required to achieve expected levels of performance scaling, which obviated the need to enable the sideport on ATI Radeon HD 4870 X2 boards. We still have the option to enable it at a later date if we encounter cases where the additional bandwidth can provide a performance benefit.
X-bit labs: Apart from Sideport and unified memory [for two GPUs], are there ways to improve efficiency of homogeneous multi-GPU technologies?
Dave Nalasco: We are continuing to work on our CrossFireX technology, and we believe there are still opportunities to improve efficiency with the current hardware architecture.