Graphics Outlook 2003

January 5, 2003 by Leonidas / translated January 10, 2003 by zeckensack & nggalai / page 1 of 4

  Deutsche Version anzeigen

   Introduction

This article is meant to provide an outlook towards new graphics chips expected during the course of 2003. Largely lacking verified data, we'll have to resort to inofficial information quite a lot. Some parts are our own speculation, but they'll be clearly marked as such. The data we'll present here is almost certainly subject to change, but it's still the best we can offer.

Note that all release dates mentioned in the following headlines are our own guestimates (oder estimates). Some vendors' official schedules may be different from our data, but the comparison would only be skewed by using these official launch dates as they usually don't coincide too well with actual 'on the shelf' dates.

Also note that mobile chips (like the ATi M10, nVidia NV31M & NV34M, and the S3 Columbia-SMA) will not be covered here, this will be purely about desktop graphics chips. We'll also skip workstation graphics which rules out 3Dlabs' future products, among others. Not to mention the complete lack of information about 3Dlabs plans to begin with.

   January:   ATi Radeon 9100

Beginning in January, the first graphics cards based on ATi's Radeon 9100 will hit the streets at prices of about 100€. ATi has only briefed their board manufacturers about this chip so far, an official presentation or note still hasn't shown up on their website. In fact, this chip isn't really new. It's purely a cleanout of R200 (Radeon 8500) chip inventories, which ATi still seem to have plenty left of. With clock frequencies of 250/250MHz the specifications are equal to the Radeon 8500LE.

This also indicates that ATi won't produce any new R200 chips for these Radeon 9100 cards, they simply rebrand the remaining items instead of writing them off. After these are sold out, Radeon 9000 cards will retake their place, offering better profits for both ATi and their partners. This doesn't mean that Radeon 9100s are bad graphics cards. DirectX8 support as offered by the R200 is still up to date. And at the intended price of about 100€ it is certainly a competitive offering. But still, the chip is nothing "new".

   February:   nVidia GeForceFX

As it happens, the GeForceFX was already announced in mid November, but the cards that nVidia promised to release to the American market in December 2002 are still not available, yet. Mind you, we expected the cards to hit Europe in February anyway, January at the very earliest. Meanwhile some board manufacturers made their announcements, we'll have to wait and see whether they'll keep their promises. Not a single GeForceFX board made it into the hands of an independant reviewer, so we won't hold our breath for a January release.

As we know the GeForceFX is based on an eight pipeline architecture running at a clock speed of 500MHz in its "GeForceFX 5800 Ultra" incarnation, while the vanilla "GeForceFX 5800" will run at 400MHz. Both versions will be relesead simultaneously. Either way, the raw rendering power will be significantly higher than that of ATi's Radeon 9700Pro (eight pipelines as well, but "only" running at 325MHz), so nVidia's new generation looks pretty appealing.

nVidia's disadvantage lies in theretical memory bandwidth: They do run the memory at 500MHz (400MHz for the non-Ultra, respectively), way faster than the Radeon 9700Pro (310MHz). But the Radeon sports an 256 bits wide DDR memory interface, while nVidia limit themselves to 128 bits DDR/II. Ignoring clock speeds, "DDR/II" is largely irrelevant for performance, leaving nVidia with a 24 per cent disadvantage in the theoretical memory bandwidth department.

But that is really only a theoretical issue with current graphics chips from both ATi and nVidia: Bandwidth saving techniques are what really matters most. nVidia rules the roost with a 4x32 bit DDR/II subdivision, which has better granularity and likewise operates more efficiently than the Radeon 9700 Pro's 4x64 bit DDR. It remains to be seen whether this will be able to compensate for the 24 per cent nominal disadvantage. The comparative quality of the other bandwidth saving features is also still pure speculation, only benchmarking production cards will truly be able to clear this up.

However, we do believe that nVidia will win the rendering power discipline and that effective memory bandwidth will work out about even. In most benchmarks, ATi's chip has proven to be limited by core clock rather than memory clock, so nVidia's core speed advantage will pay off nicely. While carrying a much higher price of about 500€, we expect a clear victory of the GeForceFX 5800 Ultra against the Radeon 9700Pro. The slower clocked GeForceFX 5800 non-Ultra will only manage to pull even.

This lesser GeForceFX's price is still undetermined but will have to mimic that of the Radeon 9700Pro. This would mean somewhere around 350€. Rumour also has it nVidia will release another GeForceFX variant running at 600MHz in Q2 to counter ATi's R350. If faced with a successful R350 this would only be rational. But it's difficult to reliably predict whether they'll be getting that kind of clock speed out of a new architecture.

   First quarter:   Trident XP4

Already announced in April, Trident´s XP4 has been with us for most of 2002. Shipments were initially planned to begin by the end of 2002 but Trident didn't manage to stick with their own schedule. Only in November did we have enough information to offer an overview of the chip's specifications, even though an affordable DirectX8 solution as the XP4 would have been a very welcome introduction to the market.

However, this year's first quarter should really see the chip's introduction. In case Trident's delays it again, the whole project could just as well be scrapped altogether. The XP4 was never meant to compete in the high-end or mainstream markets, and there simply is no market below the initally targeted low-cost segment (if we disregard integrated chipsets). Also, the XP4's architecture is the most appropriate for the low-cost segment: 4 rendering pipelines with dual texture units each, manufactured in 0.13 micron technology and sporting DirectX8 functionality.

Two things about the XP4 design are truly interesting: First, the image generation works with tiles instead of lines (though the basic rendering primitive is still the triangle), saving some memory bandwidth. Besides that, there don't seem to be any plans for furhter bandwidth saving techniques, nor a division of the memory controller into narrower subchannels. Second, the rendering pipelines "share" resources. This explains the remarkably low transistor count, only 30 million. This allows the chip to be very low-cost and to work with passive cooling without any trouble.

Three different versions were - and probably still are - planned. The T1 flavour with a mere 64 bit DDR memory interface is probably not too relevant. The T2 with 128 bit DDR and - according to latest information - 230/230MHz clocks, and the T3 version with 250/250MHz are much more appealing. According to previews carried out with beta boards, it didn't quite reach GeForce4 Ti levels, more like 20 per cent less. We're not sure how much Trident will be able to improve on this with final chips and board designs.

So, what remains as XP4's striking advantage is its price. T2 can be expected to cost around 100€, T3 will go for around 120€ - at least according to Trident's last year's plans. But market introduction must happen asap. GeForce4 Ti4200 cards still cost somewhat more, but Radeon 8500/9000 models have now almost reached these prices. The longer it takes Trident to get the XP4 to market, the less fruitful will be its low price strategy.

This also means that Trident should have already lowered the initial price target to be able to compete against the marketing power of ATi and nVidia, and to find its niche. Unfortunately, Trident has fallen silent, not too good a sign. The possibility of the XP4 now never hitting the regular market can't be denied. It surely would be a pity for the chip's interesting design.