The Radeon 5870 Reference Board
ATI launched both the 5870 and the 5850 together, albeit from what we know the 5870 should be more ubiquitous initially. Whilst we'll focus on what we have, which is the 5870, you should know that the 5850 is reduced versus the 5870 not only in frequencies, as was the case for the last 2 generations, but also in terms of ALUs (1440 versus 1600) and by extension TUs (72 versus 80). This gives us a bit of a hint about the kind of redundancy ATI has available to itself, since it seems that they can disable SIMD+TU clusters with ease. The 5850 is also shorter than its big brother.
The 5870 employs multiple clock profiles for differing usage scenarios: 850E/1200M for 3D load, 600E/900M as a down-throttle step in case of overheating (more on this in a few paragraphs), 400E/900M for video decoding and finally 157E/300M for idle mode. You also needn't worry about running 3D applications in a window, since the chip seems to properly detect them and use full 3D clocks, something which back in the "golden" days of the R600 wasn't happening.
In terms of raw numbers, a lot of muscle seems to have been crammed under the fashionable black shroud: 27.2 Gpixels/s, 108.8 Gsamples/s for Z and 68 Gtexels/s. Yes please! The rasterisation rate is also doubled, with the single unit (dual rasterisers was a bit disingenious on ATI's part in terms of education) able to convert twice the pixels in a single cycle, compared to the last generation.
If something looks slightly underwhelming in this otherwise idyllic picture, that would be memory bandwidth. The memory bus is still 256 bits wide, so the only way to extract extra bandwidth was to use faster GDDR5. ATI did this, going with a 1200 MHz (4800 MHz effective) GDDR5 for 153.6 GB/s, but it's a mere 33% improvement versus the 4870, and only 23% versus the 4890. For a chip that's generally twice as wide, this seems like too small of a bump. ATI argues that the 4870/4890 had more bandwidth than they could effectively use, and that the 5870 is properly balanced rather than bandwidth limited, but this is a contention we'll have to verify in practice. After all, the "bandwidth is king" common wisdom is still alive and well in graphics, even in this day and age.
A striking thing about the 5870 is its length: at ~28 centimetres, it's pretty long indeed, and it could theoretically interfere with SATA ports, or have clearance issues in very cramped cases. However, considering that the length is one of the reasons for the rather excellent thermal and acoustic performance that we'll discuss shortly, we can't fault it all that much.
The cooling apparatus itself is the now customary dual-slot affair, composed of a combination of an aluminium heatsink, heat-pipes and a blower that's likely to be manufactured by Delta Electronics. Asides from the PCI-E slot, the 5870 gets its electron supply via dual 6-pin power connectors, and powering up the card without these properly fed results in an brain-scorching whine – so don't do it because it won't work!
Moving to the backplane, we're greeted by an interesting assortment of connectors, namely two dual-link DVI ones, an HDMI one and a DisplayPort one, and it's probably the first time we're seeing DisplayPort included on the reference design of a high-end board. ATI claims their HDMI implementation is version 1.3a compliant, which means full support for Dolby TrueHD, DTS-HD Master Audio, AC-3 and DTS, which should make video/audiophiles quite happy. We've not verified these claims ourselves, mind you, so you may want to hunt for someone who did in order to ensure that these aren't merely claims.
Physicals
As interesting as testing the board's power draw may seem, we'll have to defer that to a later time when we'll have the tools to do it properly. What we can discuss though is thermal and acoustic performance...and both are very nice indeed. In terms of thermals, we're looking at very low idle temps in the realm of 35-40o C in our no-case scenario, whilst the highest load temps we've seen under normal operating conditions was 85o C, achieved when using the mighty Furmark, and the heat was being kept in check by the fan running at a rather low, but audible 37% duty cycle. We've mentioned normal operating conditions because our board initially suffered from a dubious condition in which the fan didn't up-spin correctly due to an improperly inserted fan connector (at least that's how we apparently fixed it, by messing with the fan connector).
This situation made us acquainted with ATI's new overheat protection: when a certain temperature barrier is hit by the GPU or its VRM, it down-clocks to 600E/900M, which is an improvement versus prior efforts. The VRM is also protected against over-current, dynamically reporting the the amperage going through it to the GPU, which down-clocks if necessary – in theory, this should've been present in the RV770 too, but that wasn't the case and the boards could end up overdrawing amps through the VRM and shutting down when faced with certain extreme loads, specifically tailored for the purpose.
The cooler itself is pleasingly quiet, and most of the time you'll not hear it, since in idle and video playback modes it's inaudible, and even when gaming it's very discrete. Overall, acoustics and thermals are a very significant step forward compared to the prior generation and we dare guess that they were a major focus during the design phases.
That being said, Beyond3D is hardly famous for its cooler and PCB reviews, so it's time for us to get to what we do best, which is analysing architectures. Respecting proper form, the first step is to describe the system we used, and the tools that aided us in our quest for knowledge.
