ATi surprised us with 4870 and 4850 graphics card and after that ATi did it again with 4870X2. This product was available to all users from the day it was officially launched and it reached the goal as number one by performances and by price range, as it was promised by ATi. This sudden success urged us on some thinking about new accepted ATi’s philosophy about multiGPU cards. If we look back in time, we could remember Rage Furry Maxx card as antic example and nVIDIA 9800GTX2 and ATi 3870X2 graphics cards from newer history and all of them were examples of fastest products in their own time.
Even though these examples were really successful in their attains to beat the competition, we still got the feeling that those were more forced acts and that they were more acts of parrying competition than they were result of long-term plans of manufacturers. With introduction of 4870X2 and now with 4850X2, new business philosophy rise to the surface. So we’ve decided to get some more detailed information about this multiGPU philosophy for graphics cards. Sasa Marinkovic, Senior Product Manager, Desktop Discrete Group AMD, was a right man for this task, because Mr. Marinkovic was included in few very serious products (X1950 PRO, HD2600 series and 4800 series) as SR. Product Matketing Manager, Graphics Production Group.
As always there will be users that want to have only the fastest graphics card, regardless of price and for those is intended this “king of the hill” (4870X2). All the rest of us, mortal souls, that need serious performances for reasonable amount of money or simply want better price/performances ratio, can wait for 4850X2. As Mr. Marinkovic stated it is clear that ATi have changed business model that was based on principle of making one top-class model (2900XT for example) and its derivates that were crippled versions of basic (top-range) product (2600 and 2400 models) intended for mid and low range products. New approach in building graphics cards that ATi uses has many similarities with process of building new multi-core CPUs. It’s clear what that new approach is if we reflect on the way that CPU manufacturer few years ago were racing who has CPU with higher MHz margin and now who has CPU with more cores.
As Mr. Marinkovic explains, very similar path is expected in future for graphics cards development. Of course this is not the only goal. Because it takes few years for a computer game to get developed, game developers couldn’t switch to new way of thinking over night and most of optimizations are actually done by ATi’s teams responsible for new driver development. They already had huge knowledge base thanks to experience gained with CrossFire configurations.
Despite nVIDIA with its SLI was first on market with such solution ATi’s solution performed better in most cases. Since GDD5 memory is expensive we asked Mr. Marinkovic why they didn’t use single memory buffer for both GPUs instead of implemented two memory buffers (each GPU has 1GB of video memory at its disposal). “Basic mistake is to expect higher performance levels with single frame buffer for both GPUs. That can be applicable to system memory (RAM) that has few applications that have access to it. When memory on graphics cards is concerned it’s not the case and gain in performances wouldn’t be great. Off course complexity of implementing that solution shouldn’t be disregarded along with expenses that would follow that kind of development and it is actually cheaper to implement double amount of video memory in this case”.
Good example of faster hardware then software development is Side Port technology. This technology supports direct link between two memory buffers on graphics cards that allows speeds in communication up to 21.8GB/s as opposed to PCI express link with maximal bandwidth of 8GB/s. At this moment even PCI express link is enough since maximal bandwidth used for communication between two frame buffers is less than 6GB/s because there is no software that can use more than that.
Our experience with Crysis benchmark confirms such opinion. With Radeon 3850 with 256MB and 512MB memory we saw big differences in performance levels although we didn’t notice that kind of difference in other games.
One idea crossed our mind in conversation with Mr. Marinkovic so we asked him, would less memory on graphics card resulted in lower power consumption. He stated: “Power consumption of memory on graphics card is very small so we could expect consumption levels just a few watts lower in that case”.
Low price for 4850x2 must be achieved thru some other cuts in production process besides usage of cheaper GDDR memory. We expected that GPUs for 4850x2 cards are taken from edges wafer while those from center of wafer are used for faster 4870x2 series. Mr. Marinkovic denied this and pointed us to some other ways of cutting production cost: “Most of GPUs used for both cards are the same so final work frequency depends on rest of elements used on circuit board. In production process of 4870x2 series we use top quality (pricier) components that allow higher GPU frequencies. Thanks to cheaper components used in manufacturing process of 4850x2 cards ATi achieved lower price and because of this GPU on these cards cannot achieve working frequencies of 4870x2. Beside GPU frequency, bandwidth and memory frequency are key factors in determination of final performances for graphics card. When you sum up all this you get cheaper but slower product”.
At this moment ATi already has new graphics cards from 46xx series. With extremely acceptable prices, low power consumption and minimal thermal dissipation they offer performances of faster graphics cards from 38xx series. Although 46xx series is important we are eager to see 4850x2 in action. Its biggest significance is that it will offer graphics card with two GPUs for acceptable price. Because of that we expect that ATi changed its philosophy in the long term and not because of marketing goals linked to new generation of product being launched. By making multi-GPU products accessible to wider user group, ATi has started a new chapter in graphics card business. Off course we shouldn’t expect that nVIDIA will sit back and watch as this train leaves the station without them. On contrary, even Mr. Marinkovic’s boss Ian McNaughton stated that GTX 280 is last “monolith” graphics card from “green” team as far as top-range products are concerned and he expects that nVIDIA will follow their philosophy of producing multiGPU graphics cards.