ATI Radeon HD 5870 Architecture Analysis

Summary

It's fair to say that, from a high level, ATI's Cypress GPU has built upon the successes of RV770, and has added some extra goodness in for good measure. As a result, Cypress is the most powerful GPU ever to be released, and it's also the most complex, too.

It generally runs rings around Nvidia's fastest single-GPU card, the GeForce GTX 285 in the scenarios we've tested here, and makes it an effectively irrelevant graphics card at its current price of around £250. However, the tests we've run have focused on general graphics performance. The Radeon HD 5870 is an excellent performer in that respect, but the quirks that were there in the previous generation are still generally there.

What's not clear at the moment is how AMD fares in more general purpose computing applications. With the advent of programming standards and APIs such as DirectCompute and OpenCL, the good thing is that we could finally start to see GPU-accelerated applications that work on both AMD and Nvidia hardware.

We've waited a long time for this day to arrive. Well, I guess we have to wait for Nvidia to actually launch its DX11 GPU, but it's pleasing to finally be on the cusp of the GPU computing revolution we were promised with DX10. We're still not 100 per cent convinced with AMD's stance on GPU computing though. It's been something of a mish-mash over the years, but the firm says it’s settled on one stance now.

AMD appears to be placing all of its eggs into the common standards basket, as it expects developers to use DirectCompute and OpenCL to harness the GPU computing capabilities on all of its GPUs - especially with Cypress. It's unclear at this time whether every application developer will be 100 per cent comfortable with using an API to develop their applications though – they’re likely to stick to more familiar environments and languages, such as C.

As game developers are used to developing at least the graphics portions of their game engines with an API, ATI is likely to see quicker uptake here. However, most game engines are coded using C or C++ and 90 per cent (or more) of their code is portable across all of the platforms they're developing their game on. Being tied to DirectX or OpenCL might not suit many developers, especially the cross-platform guys.

Despite our concerns, we are pleased to hear that AMD is working to adopt physics engines that will run on all GPUs. This is what we've really been waiting for, because until now developing a game with physics gameplay, rather than physics effects, has meant severely reducing the target audience of the game. For example, despite Nvidia’s aggressive promotion of PhysX, we’re yet to see a compelling enough reason (i.e. one or more must-have games) to opt for an Nvidia graphics card over an AMD one based on PhysX support alone.

If physics engines can be developed to an API that is hardware agnostic, and therefore runs roughly as well on ATI and Nvidia cards, developers can get started on creating games where physics can affect gameplay, as the target audience will be much wider.

However, while we can't fault AMD's support for DirectX and OpenCL, we really want to see more than just that. We say this because we believe there are developers out there who don't want to be forced down the API route (and would rather code in C, or something similar, for example), but still want to accelerate their applications on a GPU that has as large an install base as possible.

Conclusion

The Radeon HD 5870 is undoubtedly the fastest, most complex and most feature-rich GPU on the planet today. It's also the most future proof graphics card available today, thanks to its support for DirectX 11, OpenGL 3.1, OpenCL 1.0 (and potentially OpenCL 1.1, pending driver certification). In that respect, it brings back fond memories of the GeForce 8800 GTX when that launched, because only today has that cards really been superseded. G80 was miles ahead of its time and it paved the way for a whole era of game development and its architecture has also helped to shape OpenCL and DirectCompute.

We can talk all day about feature differentiation too, as both AMD and Nvidia now have their own display technologies to improve immersion in games. I'm not convinced that either is the right solution at the moment, because they each have drawbacks. We’re unconvinced by 3D Vision as while the 3D-ness is quite fun, the performance hit and the potential bouts of nausea aren’t.

Eyefinity, on the other hand, looks to be slightly more accomplished because it fixes the 'tunnel vision' problem that plagues gaming in general. By surrounding you with displays, it fills your peripheral vision and makes the experience more immersive. However, ATI still needs to polish the driver and even so we see very few people actually using it. You’ll likely need to buy a whole new set of DisplayPort screens, and a massive desk to fit them on too. Let us know what you think in the forums though!

As our Radeon HD 5870 review shows, the card is roughly equal in performance to the top-end dual-GPUs of the DX10 era. As such, it’s not the min-blowing leap that Nvidia’s G80 was. If you’ve got a fast card now, there’s no need to rush out and buy a HD 5870 – you should wait to see what the HD 5850 and Nvidia’s GT300 is like, as you have the luxury of playable frame rates at the moment.

We are happy to see the excellent anisotropic filtering of the HD 5870 - anything that raises the image quality bar gets our juices flowing. Equally, pixel shader performance is epic, as is texture fillrate, so possibly a more mature driver will unlock more fps. What’s really impressive is that the HD 5870 is very power efficient with an idle board draw of just 28W (causing our powerful Core i7 PC to draw only 150W when it’s idle). The card also draws the same power as a HD 4890 when gaming, and yet is much faster, again a significant achievement. Those fancy power circuits really work well

Whether the HD 5870 becomes the next GeForce 8800 GTX isn’t yet clear - we're not convinced it will be, frankly - but what it will do is shake things up, and hopefully lead to further image quality improvements. The early delivery of the card (before DX11 is released and well ahead of Nvidia’s competing part) should swing the balance of graphics dominance back toward ATI however – if we had to upgrade our graphics cards now, a Radeon HD 5870 or HD 5850 would be our only choice. Furthermore, games developers will be optimising their games for these cards as you read this. All we have to do now is play the game of wait and see what Nvidia fights back with.

Oh, and discuss this endlessly in the forums, of course!