Gaming, Graphics, Software Programs

UPDATED: Folding@home: ATI Radeon HD 5870 is as slow as 4870!

Besides PeakStream [now owned by Google], Folding@home was the first app supported by ATI under its Stream initiative, and FireStream GPGPU card [back in 2006].

However, as the time went by, the performance of ATI boards in Folding@home remained on the same level – a mediocre one, when compared to nVidia hardware. So, what happened? nVidia didn’t support Folding@home until the GeForce GTX 280 didn’t came out and then it was a CUDA-fest of yielding high awarding packets, while ATI Radeon 3000 and 4000 series stayed in doldrums, achieving around 3000 PPD [points per day] for the Radeon 4850, 4870 and 4870X2. There were tricks and hacks how to get the second GPU working on the 4870X2, but overall – it was far from plug’n’play experience.

With the launch of Evergreen series and the Cypress GPU, we had high expectations – 1600 shaders packed in a single ASIC looked promising, and after all the talk we heard about GPGPU optimizations inside the Evergreen architecture – our ATI Folding balloon grew to sizeable dimensions.

Since we did not receive ATI Radeon HD 5870 at launch, we had to wait until first board partner managed to send the board in. Personally, I’ve lost a lot of family members to Alzheimer, Parkinson and cancer – thus I donate my spare cycles for both GPU and the CPU to research that will hopefully, benefit us all.

PowerColor was the first company to send the board in [extensive review coming in the next two weeks], and we immediately put the card in our computer and load DVD with the drivers. Here comes the interesting bit – Catalyst 9.9 Beta drivers offered us an option to install Folding@home, a shortcut that we gladly clicked on it. Upon the installation of the GPU client, we were welcomed with a message that we don’t have supported GPU.

Upon contacting good friends at AMD, we were told that AMD is also waiting on Vijay and his Pande Group to release an updated version of Folding@home, one that would properly detect boards such as nVidia Quadro CX, ATI Radeon HD 5850 and HD 5870, for instance. But there is a way to get Folding@home running on an ATI 5000 series card.

All you need to do is get the F@h shortcut on your desktop, right click on properties and then force the R700 mode with a "C:Program filesFolding@homeFolding@home-gpuFah.exe" with two switches:

  • -gpu 0
  • -forcegpu ati_r700

However, the performance wasn’t what we expected – Radeon HD 4870 and HD 4870X2 will both yield around 3000 points per day and we definitely didn’t expected that Radeon HD 5870 1GB will join the club. In our measurements, different units showed that the number of points will vary between 1650 and 3150 points per day, dramatically lagging behind nVidia GeForce cards.

In comparison, Gainward’s GeForce GTX 275 returns between 5400 and 7100 PPD – depending on how complex the units are. Direct competitors to Radeon HD 5870, GeForce GTX 285 and GTX 295 are simply leaving HD 5870 in the dust.

We asked Dave Baumann for explanation on this unpleasant surprise at2:17AM Croatian time, and in a brief discussion we learned that theunderlying problem is the fact that ATI client for Folding@home isbased on Brook+. As we wrote ages ago, Brook+ is pretty much "put outto pasture", i.e. the support for it has been killed long time ago.

AMDkept banging on the OpenCL drum and stated that DirectCompute is fullysupported, including Double Precision mode. At present, this means little for Folding@home community as Pande Group is unfortunately late with the second [or third?] generation of GPGPU clients.

FahMon - ATI's HD 5870 Folding@home performance isn't exactly stellar...
FahMon – ATI’s HD 5870 Folding@home performance isn’t exactly stellar…

Bottom line is, if you want to fold – you have to either wait until Pande Group switches to OpenCL in some foreseeable time in future. OR – keep on using GeForce boards for Folding, and use Radeon 5870 for supported applications such as MilkyWay@home.

Overall, quite a disappointing start of our Radeon HD 5870 evaluation. Luckily, games are a different thing.

Update October 8th, 2009 12:05 GMT – We received word that the new version of Folding@home is in the works. According to a blog by project leader, Folding@home in its third generation of GPU code will adopt OpenMM, Standford’s open molecular model. Next in the works is bringing OpenCL support into the OpenMM library, resulting in OpenCL-acceleration for all GPU cards, including previously non-supported parts [but OpenCL-supporting], such as Quadro CX, FirePro and the like.

Update October 8th, 2009 15:16 GMT – After reading the comment from Mr. Andreas Przystawik [Gipsel], we could not but to quote his comments. You can read the whole comment in our complementary comments section, just scroll down.

"Look at two other distributed computing projects using ATI cards: Milkyway@home and Collatz Conjecture. Both scale virtually perfectly with the higher shader count. That a HD5870 achieves this with the Collatz Conjecture project is also showing that ATI has done a tremendeous job with the memory controller efficiency as it is quite bandwidth hungry."

"Both projects use the Brook+ layer as a base for the ATI applications. So that has nothing to do with the possible scaling."