Thermal Performance

Putting realistic, repeatable load on a GPU to get a decent idea of it's real world thermal output has long been something we've experimented here at bit-tech. We've found that Synthetic benchmarks such as FurMark thrash the GPU constantly, which simply isn't reflective of how GPU will be used when gaming. It's such a hardcore test that any GPU under test is almost guaranteed to hit it's thermal limit, the mark at which the card's firmware will kick, speeding up the fan to keep the GPU withib safe temperature limits.

Conversely, simply leaving a game like Crysis running at a certain point also isn't reflective of real world use. There's no guarantee that the GPU is being pushed as hard as other titles might do, and the load will vary from play through to play through.

Eventually then, we've decided to use 3DMark06's Canyon Flight test as a real world representative, repeatable graphics test. It's a ferociously demanding test, pushing graphics cards to their limit, but also containing peaks and troughs in performance that match real world game play.

As the test is so demanding and GPU limited, we've set 3DMark to run the test at 1280x1024 with 0xAA and 16xAF (enabled in the driver), constantly looping the test for thirty minutes and recording the maximum GPU Delta T (the difference between the temperature of the GPU and the ambient temperature in our labs).



For analysis of these and all other results, please read the Results Analysis page.

Overclocking

With the Radeon HD 5850 still using very immature drivers, there isn't a lot of publically available overclocking support for the card just yet. As it stands, we're only equipped with the ATI Overdrive utility buried within Catalyst Control Center, which is usually fairly limited in its scope.



This time we were only allowed to push the card up to 775MHz on the core and 1,125MHz (4,500MHz effective) on the memory - clocks we greedily set about achieving. Needless to say then that the card easily coped with these improved clock speeds, representing a seven percent improvement on the core and 12.5 percent improvement on the memory. While the core overclock is fairly mild (we'll have to wait for better support from overclocking tools before tinkering further) the memory overclock isn't too bad at all, and represents an improvement in memory bandwidth from 128GB/sec to 144GB/sec.