ATI Radeon HD5870

THIS WEEK, AMD's long anticipated HD5870 GPU, codenamed "Cypress" and the new kid on the ATI Radeon graphics block, took the "fastest GPU chip" crown away from Nvidia's GT200b in the GeForce GTX285.

You've probably read plenty about the new 2.15 billion transistor GPU with 1,600 shaders and 4.8GHz GDDR5 memory throughput, which is also the first non-classified chip of any kind to deliver 2.7TFLOPS in floating point single precision and over 500GFLOPS in double precision. Oh yes, it's true, but you've got to code under AMD Stream or OpenCL to get anywhere close to these figures, though.

The HD5870's initial GPU clock is set at 850 MHz, although overclocked editions running at up to 1GHz should be expected to appear over the next month or so. Asus's HD4870 edition card, for instance, already supports GPU clock speeds of up to 1,035MHz right now with built-in voltage tweaks.

However the real excitement will be the upcoming HD5870X2 implementation, which is expected to support up to 2 x 2GB memory configurations per card, and, if you're into GPU computing, up to 10TFLOPS of combined single precision computational power using two such cards per system.

Here we have a quick look at the new ATI GPU platform, pre-release drivers bound, on a matching high performance computing foundation - an Intel Core i7 975XE processor on the Asus Rampage II Extreme mainboard, with 6GB of Kingston DDR3-2000 RAM and a CoolerMaster V10 heat sink fan combo with TEC cooling, on a 30-inch 2560x1600 LCD display.

We ran the CPU at 4GHz across all cores, and memory at the DDR3-1600 CL6 low latency setting. Right now, what we have is an early reference sample card, and we expect to see higher performance from the final card version with newer drivers.

So, at this stage, it's just an initial look at a reference sample card using the first software driver, with a couple of 3DMark Vantage runs.

Before we run the usual 3DMark Vantage test, keep in mind that this is a 1GB card without the EyeFinity six display feature. This standard model can run 'only' three monitors at once. The EyeFinity six display port model will come with 2GB of GDDR5 RAM to handle three more screens.

Here are the 3DMark Vantage "Performance" results:

And the 3DMark Vantage "Extreme" results:

As expected, these are record setting results for a single GPU. Compare it to the GTX285 OC 2GB on the same machine, where we get just 12,102 on the GPU score in Performance mode and 5,866 in Extreme mode. The various game benchmarks on other sites also paint a similar picture.

During these benchmark runs, this early sample card was noticeably less hot than the ATI HD4870 or HD4890 cards I've handled before. And this is factory reference cooling we're talking about, and not running in the idle mode, but in the full load 3Dmark test. The black backplate, while not providing any extra cooling, does help aesthetically, as well as preventing finger scratches from the PCB back side. The card is large, just a bit smaller than its ATI Radeon HD4870X2 dual GPU predecessor, however it should still fit in most ATX casings without problems.

In summary, the new king of the GPU hill has been enthroned. For how long, we'll see - it all depends on how well the next Nvidia graphics processor, the GT300, performs. Intel's Larrabee seems to be still far enough in the future for both of these big graphics vendors not to get too worried yet. µ