Graphics Outlook 2003
First quarter: Matrox Parhelia 8X
As the name suggests, this is the AGP8x derivative of Matrox' Parhelia, supposed to directly replace it after remaining chip inventories have been cleared out. We don't know exactly when that will happen, but we don't think it can be too far into the new year. Just as uncertain is whether or not this chip will carry any further enhancements: The unconfirmed rumor of several hardware fixes and increased clock speeds is floating about. We'll have to wait and see.
First quarter: ATi RV280
Similar to nVidia (GeForce4 MX 8X/Ti 8X aka NV18/NV28), Trident (XP4e) and Matrox (Parhelia 8X), ATi too will deliver an otherwise identical AGP8x capable version of their Radeon 9000 /Pro (RV250), which goes by the code name of RV280. ATi will likely differentiate this chip via the numbering scheme, probably it will go on sale as Radeon 9050 /Pro or Radeon 9200 /Pro. Priced around 100€ and below later in the year, ATi will still be able to offer a cheap DirectX8 solution. It is likely that ATi has already stopped Radeon 9000 chip production, working on inventory cleanout and preparing a production shift to the RV280.
Second quarter: nVidia NV31
nVidia is planning an "MX version" of GeForceFX for the budget segment, and in direct competition to Radeon 9500 /Pro, to be unveiled during CeBIT 2003 towards the end of march. If everything goes smoothly - tape out has already happened - the first NV31 based graphics cards should hit the shelves quickly after CeBIT. It's almost a given that nVidia will use a trimmed down NV30 design for the NV31 to attain mainstream prices at around 200€.
The first step will probably be a reduction from eight to four rendering pipelines, along with a reduction in core clock to help yields. Very recently we also heard rumours about nVidia cutting back the flexibility of the chip's programmable pipeline to "normal" DirectX9 levels. That wouldn't help too much for transistor count but it would at least serve to further increase the gap between NV31 and the "real" GeForceFX. This idea just might appeal to nVidia.
Because we don't know the technical details - and most prominent, the clock speeds - we cannot really say how the NV31 will fare against the Radeon 9500 and Radeon 9500 Pro. But we can safely assume that nVidia will put some effort into comparative testing before settling for an official clock speed, and that they will then most likely make it so that NV31 will edge out ATi's Radeon 9500 Pro.
So we assume a victory over the Radeon 9500 Pro - after all that's not asking too much from a chip in the same segment, but coming out three months later. The real enemy will be RV350, scheduled around the same time and ATi's first chip manufactured on a 0.13 micron process. Both will probably perform similarly and given their coincident release dates, a fierce price war could result between the two.
While we're at it, just like the GeForceFX we can expect two variants of NV31, differing only in clock speeds. Since the introduction of the GeForce3 Ti, nVidia consistently put out several versions of the same chip, the sole difference being clock speed. This allowed them to serve multiple segments with little cost overhead while at the same time helping yields. The GeForce4 MX and GeForce4 Ti generation saw three versions each, probably one too many after all. With GeForceFX we're back to two versions, and this rule is very likely to also apply to NV31.
Second quarter: Matrox Parhelia 2
Also known under the code name "Pitou", this project is supposed to breath some new life into Matrox' Parhelia architecture with some enormous projected performance improvements. What exactly that means, we cannot tell. An obvious first step would be a transition to 0.13 micron manufacturing (down from the original Parhelia's 0.15 micron) to allow significantly higher clock speeds.
Another likely change would be higher memory clocks and improved DirectX9 capabilities. Matrox adding a few techniques for more efficient use of bandwidth would be ideal, as the lack of these is really holding back the original Parhelia. But that won't come for cheap, be prepared for seeing it in the 450 to 500€ price range again. This time around the assault on ATi and nVidia really must succeed if Matrox wants to secure its own niche in the graphics board market.
Sadly not only the technical details of Parhelia 2 are unknown, even the question whether it will ever be released is still all too evident. The tape out has occured, that's for certain, theoretically allowing for a 2nd quarter release. But there's also the wild rumour of the first chip samples largely disappointing, which led to Matrox scrapping the whole project, even contemplating leaving the 3D graphics market altogether. As Matrox, as usual, veil themselves in complete silence, we can neither confirm, nor plausibly dismiss this rumour.
Second quarter: S3 DeltaChrome
After looking at the DeltaChrome material posted on S3's website, we wonder: Was that an actual introduction of this chip? Not quite - so we're still waiting for an official presentation and launch. According to the early information we fathom the S3 chip as a DirectX9 solution aimed at the mainstream market - mainstream being our best educated guess.
The S3 Deltachrome implements an architecture similar to the GeForceFX and Radeon 9700 Pro, i.e. eight rendering pipelines and likely a 128 bit DDR memory interface. Clock speeds however will "only" be around 300MHz core and between 300 and 350MHz for the memory. We haven't seen any information on bandwidth saving techniques or at least a subchanneled memory controller yet - without them, DeltaChrome would almost be a lost effort. We need to wait for some more information from S3.
S3 will still not be able to compete with the GeForceFX and Radeon 9700 Pro - it's in fact unlikely that VIA, S3's current owner, would want to try to. Graphics chips designed by S3 primarily have had to be cheap since VIA took over, another good reason for S3 to go for 0.13 micron production. This will also make it easier to move the DeltaChrome core into the mobile market (dubbed as "Columbia-SMA") and later into VIA's integrated motherboard chipsets.
Accordingly, we definitely don't expect a high-end product. DeltaChrome sharing the DirectX9.0+ label with the GeForceFX (Radeon 9700Pro "only" does DirectX9.0) also bears little significance: Exceeding only one DirectX9 sub-feature would already justify the plus sign. This kind of classification cannot tell us whether the DeltaChrome exceeds DirectX9.0 up to the point of GeForceFX capabilities, or only a little. A might-be interesting feature could be what S3 calls "Advanced Deferred Rendering", but this could be either reality or pure marketing. We still can't tell from our information.
Mass production has been slated for the first half of the year, quite a time window. Currently we'd rather think the DeltaChrome graphics cards will hit the market in Q2, with an announcement in Q1 - good for marketing. DeltaChrome will have to battle it out with the other DirectX9 mainstream chips that will be available by then: ATi RV350 and nVidia NV31, by no means easy opponents.
But maybe S3 does surprise us with the DeltaChrome being either faster than we think or cheaper. In both cases, a healthy market could be found for the chip, otherwise it will be an uphill battle against mighty ATi and nVidia, at least without unique must-have features to differentiate the chip.
Update: On Jan-07 S3 released further information regarding the upcoming chip. Unimpressed with our skepticism, the chip will rightfully bear the "2.0+" designation for vertex and pixel shaders. DeltaChrome's vertex shader is equivalent to GeForceFX's and its pixel shader is equivalent to Radeon 9700 Pro's, but a little bit more than Pixel Shader 2.0. On another positive note, the first hard facts about "Advanced Deferred Rendering" have surfaced, S3's latest looks a lot more appealing now.
Update II: According to new information, the memory clock will amount to 300MHz. S3 seems to go for synchronous operation of chip and memory across their lineup. The memory interface width has now been confirmed to be 128 bit DDR.