RV770: ATI Radeon HD 4850 & 4870 analysis

Final Thoughts...

There is a hell of a lot to like about the Radeon HD 4800 series and RV770 is arguably one of the best ATI graphics processors we’ve ever seen. AMD has made some really big changes in its latest RV770 graphics architecture and the focus has clearly been to make what was an inefficient design an efficient one.

Aside from some transistor-level optimisations, the shader core has remained largely the same – but almost everything around it has been tweaked in one way or another.

With both R600 and RV670, AMD had a serious amount of compute power on tap and that’s even truer with RV770 – the difference is that the company’s architects have made some clever decisions to better harness the power that’s available inside the shader core.

The new cache layout, for example, is something that really needed to be done – shared cache works well in many scenarios, but when you’ve got a chip as wide as today’s modern graphics chips are, you need more than one shared cache, otherwise there are literally hundreds of execution units running thousands of threads all trying to get a bit of the same piece of silicon. With RV770, the cache hierarchy is much better suited to the task at hand – there are global data stores available for particularly large tasks that cannot reside in just a single SIMD core, and there’s enough bandwidth available inside the chip to move data between the various caches.

One of the biggest improvements has to be anti-aliasing performance – it’s nothing short of incredible and it takes AMD (or ATI of old) back to where it used to be. The Radeons always used to make more efficient use of AA than Nvidia’s equivalent products and with the Radeon HD 4800 series, that has made a comeback in style.

Obviously there was a fairly low barrier for entry when it came to improving the situation, but AMD has gone well beyond improving it – the architects have gone from zero to hero almost as quickly as you can say just that. Improving the render backend throughput in certain scenarios was a clever way to improve anti-aliasing performance and I’m sure that GDDR5 has had an influence as well.

That’s not to say that the Radeon HD 4850 is starved of bandwidth though – its anti-aliasing performance is better than anything Nvidia has to offer at the moment. The Radeon HD 4870 and 4870 X2, on the other hand, are just in a completely different league. Playing games at high resolution with 8xAA enabled is verging on plausible on a card that costs quite a bit less than £200 (including VAT). And with the Radeon HD 4870 X2, it’s almost the default – at anything less than 1,920 x 1,200 4xAA, there’s simply no benefit to owning the card; it’s all about cranking the image quality right up in today’s games.

What’s disappointing for me is that AMD sees a future of multiple ‘small’ GPUs to fill the high end market, when in actual fact we don’t want to be greeted with the headaches associated with today’s multi-GPU technology – it just doesn’t scale well enough across the board. I’d love to see a bigger chip derived from RV770 with well over 1,000 stream processors because at that point we’d be seeing some rather crazy and consistent performance gains.

With that said though, it’s very clear that RV770 was all about architectural efficiency and on that front, AMD has really achieved. RV770 caught Nvidia by surprise and it wasn’t until the company cut its prices that it became competitive again – the GTX 260 is an attractive proposition, but it’s still a little too expensive for my liking. It needs to be on price parity with the 4870 if it’s going to steal back some market share from the big red rose in AMD's lineup.