Reflection on Silicon Process

Looking at the market at present we have Matrox with a recently released Parhelia (83 Million transistors) product on .15µ, Creative/3Dlabs with the P10 on .15µ (76 Million transistors), ATI with R300 (110+ Million transistors) on .15µ and NVIDIA's NV30 (120 Million transistors) in design on .13µ.

The fact that three out of these four 3D vendors opted to take their next generation products on to .15µ rather than .13µ signified the uncertainties surrounding the process at the moment. So-called issues with the process have long since been discussed and these would appear to be apparent given that all the products that have actually made it to silicon so far are based on .15µ. 3Dlabs even stated to me they didn't think .13µ would be commercially viable until next year.

Conversely we have NVIDIA stating NV30 would not be possible on any process other than .13µ. On the introduction of P10 3Dlabs also stated that floating point pixel pipelines, as DirectX9 demands, wouldn't be feasible on a .15µ process and would have to wait until .13µ.

When you add these things together it's almost as though ATI caught everybody else off guard. R300 is a part that's not terribly less complex than NV30, and yet ATI found that it is possible to do this via .15µ. Likewise, ATi also managed to squeeze floating point pixel pipelines in to their .15µ part, something which 3Dlabs thought wasn't feasible (or least, not for their architecture). On top of that, ATI hit speeds of 325Mhz for R300 in the Radeon 9700 PRO product, which is higher than their previous part which was a little over half the complexity and based on the same process!

All in all it's hard not to think this part as quite an engineering feat. But also where did this type of performance increase come from? R300 is still basically on the same process as R200; sure enough there may have been refinements to the process but what is there that a part this complex could not have been achieved a year earlier? Apart from the silicon engineering probably the single most important element that facilitates these types of speeds is the introduction of BGA memory packages, which has facilitated the adoption of 256bit busses at feasible (if still somewhat high) consumer pricings; without these wide busses, chips this powerful wouldn't be of any use as there just would not be significant bandwidth to make use of it.

One other element to consider with the R300 chip is that of die size and hence cost. A chip this large is likely not to be achieving the yields of chips like the Radeon 8500 which will drive the price up. In the face of .13µ based competition it's a wonder if this will have much room to go down in price and how soon yields will stabilise on the .13µ process for that to become the more attractive proposition. However, presently there is no .13µ based competition and ATI's decision to stick with .15µ, although initially sounding risky, has played right into their hands and could pay dividends in the short term.

Conclusion

This was my final comment in my previous Radeon 8500 review some 5 months ago:

"In my eyes, the original Radeon marked somewhat of a turning point for ATI, something which has been emphasised by the technology of the Radeon 8500 and the acceptance of ATI's new business model. It's certainly going to be interesting to see what ATI have in store for the future..."

I think we can say that is certainly has been interesting seeing what ATI has in store for us, for what started with Radeon and started to show a little more with Radeon 8500 really has pushed to the fore with Radeon 9700. The performance of the board is reasonably staggering in some cases in comparison to their previous generation and they have done this whilst introducing the most complex and feature rich chip on the market place today. The chip itself isn't quite a departure from the norm, nor as programmable at 3Dlabs P10 part, and feels more like a sensible, evolution from where they have been going over the past few years. The chip does, however, offer much more programmability than their previous parts and full DirectX9 compliancy.

The Radeon 9700 PRO 'feels' much more of a complete, well rounded solution than I felt Radeon 8500 to be, certainly in its initial stages. One area that ATI has listened to is the comments concerning their driver development; the 'CATALYST' push has shown some improvements with their current products and this is really beginning to show for Radeon 9700 PRO. None of the tests used in this review showed any issues, and there were no odd stutters or serious performance drops as characterised so often by 8500. Given the how new the board is it's quite surprising to see it as driver problem free as it is; ATI did supply a list of known errors, but most of these seem relatively minor.

Something to note, though, is that this is one hot board. It's not surprising that this gets hot considering the chip size and running speeds, and the heat-sink size is a fairly big giveaway, but the entire board and many of the components do get quite hot to the touch. It should also be noted that the heat of the board did not manifest any stability issues or rendering issues even after extended periods of running, though I did encounter a minor issue with the particular board sent to Beyond3D.

There should be a small heat sink / spreader attached to four small chips on the underside of the board, yet as I went to remove the board after running some tests the heat sink dislodged. Having contacted ATI it seems they are aware of this and have since changed the adhesive used for the production boards. ATI explained that the spreader is there it appease some OEM manufacturers and the board sent to Beyond3D has remained perfectly operable without it.

As I've gone through reviewing this card I've been incredibly impressed with the performance numbers being posted, though in another sense it's also a little depressing. As the 3D industry has progressed, speed has always been an important issue and it's always been something to strive for; however, as the pure rendering performance is picked off, so we've strived for higher quality FSAA and filtering. Now Radeon 9700 PRO, and likely any high end offering after it, offers us these on a plate - so where next? We need content that will make use of these boards. The delivery of the hardware has far outstripped the speed of delivery of software that is able to make use of the hardware. It's when you realise this you really begin to understand the importance of projects such NVIDIA's Cg and ATI's RenderMonkey (a public beta, of which, has become available today) - something has to occur to deliver content that can make effective use of this hardware. You can bet that ATi and NVIDIA have both realised this, which is why they have gone into such ventures - if they don't do something to accelerate content development to make use of the incredible flexibility and programmability of these architectures there will not be a drive for such high end boards or the need for replacement as often--and thus goes a large chunk of potential sales. So while many of us are wanting DoomIII to appear, you can rest assured that the hardware vendors also want it and many more games like it.

Still, Radeon 9700 PRO is here, today, and if high resolution, high quality filtering, high level FSAA rendering at high speeds is what interests you then it has all of this to offer you right now. On top of this you also get DirectX9 compliance when Microsoft releases the new API and a healthy dose of features and programmability to see it into the future. As consumers, having a market with at least two healthy, innovative, companies is a very good thing.

• Please feel free to comment on this article here (after you gone back and actually read the article because you just skipped straight to the benchmarks the first time and missed some interesting stuff!).