It's no secret that there appeared a lot of preliminary data (for example, on www.anandtech.com and www.vr-zone.com) on future ATI products based on the new R520 chip and its crippled low-end modifications (RV5XX). Let's try to sum up the available information, pick out the most likely (logically) scenarios and parameters, and analyze the prospects. On we go.
To all appearances, the following solutions will be announced:
There are no doubts that the key feature of the new products is their SM 3.0 support. Good stuff that! At last!
R520 possesses significantly (we emphasize this word) overhauled pixel processors. At least in comparison with the previous R4XX architecture, burdened with rudimentary phases - the heritage of the dark transitional period of the first shaders and DX8. Such architecture differed much from the more flexible NV3X-4X (designed initially with regard to more or less arbitrary, in structure and length, shaders and dynamic branching). As a result, pipelines became flexible, they learned to branch and other SM3 features. Thus we have more complex (versus previous generations) pixel processors both in terms of the number of transistors and the die surface as well as the algorithm. But having become more flexible and complex, have they become faster? Not necessarily. Such major changes have been applied to ATI's pixel processors for the first time since the R3XX times. The architecture of R3XX and R4XX was simpler than in NVIDIA. In fact, having a weakness (no SM3 support), it also has an advantage - high efficiency in simple tasks. Has this efficiency been lost now that the flexibility is attained? One trouble most likely makes up for another - we get more flexible pipelines, which thanks to various tricks and innovations are comparable to the old (equally clocked) in terms of performance. Perhaps they are a tad faster, but obviously not by many times, not even by 50% — just about 10-20% of the peak advantage per pipeline in case of equal frequencies.
Funny thing. On the one hand, the ATI architecture became closer to NVIDIA. But on the other hand, due to its initial differences, it made a step towards what we already had seen in the R3XX-4XX and NV3X-4X. It's a step to the future, represented by Xenos (really revolutional architecture of the MS Xbox 360 graphics accelerator) and future architectures of graphics processors of the new Microsoft OS (previously known as Longhorn). It's not that long as in case with Xenos, but it's essential. We shall dwell on this issue in our future article devoted to new architectures and their usage in game boxes. But so far we can notice the existence and unity of dialectical opposites of tendencies in the R520 (smart phrase, eh?) ;-)
So, from the performance point of view, we should expect pixel pipelines to be more (but not critically) efficient per unit of frequency than in the R4XX, thus comparable with G70 pipelines. But don't forget that G70 has 24 pipelines instead of 16. ATI will gain absolutely nothing with such 16 pipeline chips. On the contrary, it will most likely run into problems. See for yourselves: the core frequencies are not as successful as intended at the design phase. Who knows how many 600 MHz cards will be actually available, while G70-based cards have already swarmed the shelves. The market is on the verge of potential release of the advanced G70 (7800 Ultra) based on the same easily-overclockable chip (most likely with faster memory). No matter what, there is some risk to lose the prestigious race in the top sector. Our tests will reveal everything. But even now we can forecast rather close results without major advantage for ATI. The calculation is simple: 24*430=10320, while 16*600=9600. Even if we assume that ATI has the 20% frequency advantage (720 MHz*16=11520), it does not look like super excellence. This is the theoretical maximum for shaders. But in fact, everything will depend on a given application, thus testers will not have a unanimous opinion. It's impossible to write miraculous divers here, the similar shader optimization level is dictated by the SM3 architecture and DX9. Much depends on drivers of course, but it is mostly about not losing rather than defeating NVIDIA. Taking into account how long ATI had struggled to streamline the R520 production, we can assume that the drivers are not just ready, but also fine-tuned/optimized for many times. The first pilot chips appeared very long ago.
There may be parity or the 600 MHz cards may be even slightly outperformed. But will the 600 MHz (or higher) cards be really available? 550*16=8800. Even +20%=10560 - in case of the best theoretical fillrate it means parity, in reality - even a small lag in many applications. And now imagine that NVIDIA manufactures G70 at 500 MHz (or a tad lower) and with faster memory? That's that.
Let's draw a conclusion. Faster memory bus and pixel pipelines in the R520 will be compensated by fewer pixel processors. However, you shouldn't underestimate the role of the local memory — it may work wonders, if used correctly. Another matter is that NVIDIA is not asleep either — faster memory is available and can be installed, the overclocking potential of the G70 is well known. As a result, the most likely situation in this market segment will be performance parity (or close to it), everything will be up to the price, preferences, and additional features. Their capacities are comparable. The prices may be comparable (especially if these products start competing with each other). But what concerns availability, you will agree that NVIDIA has a noticeable head start.
But as is well known, much depends on a vast Middle-End sector in financial terms (Low-End has been taken up by integrated solutions of late). That's where belong two other chips — RV530 and RV515. ATI enjoys a more peachy position here, especially in case of the RV530 - 12 pipelines and a good core frequency — up to 700 MHz (that's what the more complex R520 lacks). There is a theory that this chip is the crippled R580 rather than R520 - hence the higher frequencies and some other features. Something tells me that we should also hope for significantly higher yield of effective chips than in case of the senior model. It's manufactured by the 90 nm process technology. That is it's potentially more profitable, especially if the production volume grows. On the other hand, it's no secret that as process technologies grow thinner, they become increasingly freakish. Their advantages are getting absolutely debatable, even in Middle-End (that is after a long phase of streamlining the manufacturing process and with large production volumes.) This very freak, the large gap between the desired and real parameters of the R520, resulted in the moderate core frequency (500-550 MHz on the average, 600 MHz in the top and most likely difficult-to-get model) and significant delays in launching these chips. Do you remember how much time and effort it took NVIDIA to streamline the new process for the NV3X?
The RV515 is the cheapest chip, it has only one quad, which significantly facilitates some aspects. If it's manufactured by the 90 nm process technology (we don't know ATI's plans in this respect so far), it may become a very inexpensive chip and still offer all typical features of this series, including SM3. A wide range of available frequencies, memory bus width, and memory types supported allow a flexible response to the market demands and competition. There is also the HyperMemory (HM) mode planned. Moreover, a vendor can order it have the narrowest memory bus possible — 32 bit. The fate of this chip will be decided only by its price.
Despite the rumors about 24- or even 32-pipeline R520 modifications, we are inclined to think that even if such (trial) chips had been really designed and manufactured instead of being just modeled (it cost ATI a great deal of additional expenses in case of this scenario), they didn't pay off. The yield of effective chips or the peak frequencies of these monsters forced the company to stay with the base 16-pipeline R520 modification. To all appearances, there are no additional pipelines on the production die. If there are any, they will be granted only to the optimized R580 modification, which is currently in the design phase. It will probably be used with the XT PE suffix (see the table with cards below) or with a new number, for example, RADEON X1900.
X1800 series: XT will most likely be a card for the press, at least for the first several months. Enthusiasts will get the XL card (if all goes well). The PRO modification will be the main card, as always. Partially defective chips with one locked quad will be used in the LE modification, more or less available to a normal user, not a gamer. By the way, top solutions are subject to overheating and take up two slots (due to their cooling system).
The X1600 series looks the most attractive in terms of competition with NVIDIA – 12 pipelines, Shaders 3.0, the advanced technology. Performance, which used to be available to the previous generation of top cards, but at a reasonable price and with moderate power consumption taking up only a single slot. 700 MHz in the senior model holds out a hope for a good overclocking potential in cheaper models. Both modifications planned possess 12 pipelines so far – it speaks of high projected yield of effective chips. The Middle-End series will offer much room for vendors' high-handedness – they may install various memory types and capacities, including cards with halved memory bus width (64 bit) and old inexpensive DDR memory at low frequencies. If the prime cost of the chip is reasonable, it will lead to well-balanced and competitive inexpensive cards, intended to replace the previous generation of cards and to give it hot to the 6600 and 6800 series from NVIDIA.
X1300 series – inexpensive RV515-based cards. Low power consumption, all features, and hopefully a moderate price accompanied by the rich choice of various modifications from independent manufacturers. Thus ATI closes down the entire line of discrete products and moves it up to a new architectural convolution, supporting SM3. All that remains is to design a proper chipset with an integrated solution. It's most likely in the design phase already.
It goes without saying that the current who-is-who for the entire line will be determined by the following three parameters as usual: performance, availability, and price. There is not much time left to wait.
On other features and interfaces
The new accelerators support HDCP format for both DVI interfaces. Senior R520-based models are capable of applying HDMI (High Definition Media Interface, the interface for outputting video and audio to digital theatres and other audio-video playback devices of the new generation) to DVI connectors. You can read about popular interfaces here; have a look at connectors here, and read about the HDMI standard here). All you have to do is to plug the DVI-> HDMI cable. We don't know yet whether this feature is available to all R520-based cards or only to the All-In-Wonder line. Of course, sooner or later (later is more likely) X1800 CrossFire will be launched (CF, I had nearly typed SLI). All regular X1800 cards are initially compatible with it and can work as slaves. Besides, all cards with an external TV-Out chip will support analog component HDTV-Out.
R580 is the most interesting. When and how many pipelines, that is the question. It will be obviously based on the same architecture as the R520, but probably with more pipelines (24 or even 32). It's no secret that the new architecture, similar to Xenus (R500) from xbox 360, with unified processors, and besides fully compliant with WGF 2.0, will be implemented only in the R600 and in the next NVIDIA architecture (previously known as NV50). We shall not see both chips soon, only in 2006. But as for now, we shall witness severe competition in the Middle-End sector and an attempt to become a formal High-End leader by launching optimized cards, probably with more pipelines, higher-clocked cards and revisions, etc.
That's the future. As for now, there are no prospects of the R520 leadership, there will be no revolution. The most vital question is the same: the real price and availability of top ATI cards. Shall we see the 600 MHz R520 on the shelves right after the announcement?
Alexander Medvedev (email@example.com)
September 19, 2005.
Write a comment below. No registration needed!