GPGPU Performance

I wanted to briefly cover GPGPU performance – the act of using your GPU to perform tasks other than gaming. There are plenty of benchmarks out there, such as the always useful SiSoft Sandra’s OpenCL benchmarks. Using it, you will get a result like this:

What exactly does this tell us though? Is the Radeon 7870 “twice as fast” as the GeForce 660? Is the Radeon 7870 XT then twice as fast as the 7870? In this test, they are! But unless you are into running SiSoft Sandra’s OpenCL benchmark all day, it won’t be of much use. I have found that OpenCL applications can vary a lot in performance.

So instead, we are going to focus on real world GPGPU computing, using software that someone might actually use at home.

Folding@home GPU Performance

First we’ll look at what is I assume the most popular non-profit Distributed Computing projects around,Folding@home. It can also be considered one of the pioneers in GPU computing.

We will be using what I think will become the standard in Folding@home benchmarking – FAHBench. It supports both OpenCL and CUDA, and can run on CPUs as well. We will be including the results from our CPU just for reference.

Previously, the GeForce 660 was able to overtake the Radeon 7870 by using CUDA, which is only available to Nvidia cards. The 7870 XT is able to overtake it however, due to hardware design.

GPU BitCoin Mining

Another interesting use of GPU computing is BitCoin mining. Mining is a way of maintaining the currency itself, and a way to reward everyone for participating. We are only going to consider mining as a casual user – if you are seriously considering mining as a venture, you are going to want to spend a lot more time researching this.

In fact, the days of GPU mining at all are numbered – as more people mine, and specialized hardware hits the market, mining difficulty increases. So where a 400 M/hash card could make about ฿0.07 BTC per day last January when the difficulty was around 3,000,000, you are now looking at about ฿0.03 generated instead. Using today’s exchange rate, that’s going from $221 USD a month in income, to $97.70. It will only go down from there, and depending on how much you pay for energy, will eventually not be profitable at all.

When the flood of mainstream ASIC units hit the market – such as Butterfly Labs’ $129 units that can calculate at 4.5 G/hash per second while consuming just 6 watts of power – difficulty will likely skyrocket to the point where it will not be feasible at all to use even the best GPUs.

But if you are just interested in checking out mining for yourself, and perhaps making a few BitCoins on the side (they are more popular than ever now, with yet another European bank collapsing and furnishing money from its citizens), you will probably want to keep it simple. The easiest way to get started is to use GUIMiner, a Windows client that supports several miners, including some GPU miners.

We used GUIMiner with the default OpenCL miner, as well as a CUDA miner (RPC Miner, CUDA version). This required quite a bit of research in itself, since the latest stable version of RPC Miner doesn’t support the GeForce 660. I was able to find a binary that works over at bitcointalk, but it turned out not to be worth the effort:

If you were expecting the 7870 XT to mine significantly higher than a 7870 for the same price, these results have to be disappointing. They are especially disappointing when you consider power consumption (which you must, if you’re doing this for profit)

Despite having a TDP that isn’t much higher than a standard 7870, the power consumption of the 7870 XT goes through the roof during mining. This means that the standard 7870 is a far more efficient miner. Punching these numbers into a Bitcoin profitability calculator, we find that with the current difficulty and exchange rate, using an energy cost of $0.12 USD:

Radeon 7870 XT 3 month income: $254.38, power cost: $83.63, net income: $180.76

Radeon 7870 3 month income: $266.37, power cost: $56.41, net income: $209.96

GeForce 660 3 month income: $40.19, power cost: $27.88, net income: $12.31

The 7870 XT isn’t useless, but it is disappointing that you don’t benefit from the superior architecture, and end up paying a lot for power cost.

By the way, if you are into Bitcoin, we accept donations in that form only. If you found this information useful, please consider it with my gratitude!

GPU Video Encoding

Another common use for GPGPU is video encoding. There are a few apps out there that let you use your GPU to encode video. The problem I have with these is that they don’t offer much control over the resulting file. Once a fully capable GPU powered x264 client is released, this will truly be something worth considering.

For now, we have CyberLink MediaEspresso, which can use GPU hardware to encode. We converted a 1080i MPEG-2 recording from a terrestrial feed to each program’s “iPad 2” profile. We normally use ArcSoft MediaConverter as well, but there was a compatibility issue with the 7870 XT.

There seems to be a compatibility issue with both programs – while MediaConverter didn’t work at all, MediaEspresso doesn’t seem to use the 7870 XT fully. It is faster than CPU encoding, and uses less power, but these results seem off.

Personally I’d stick with Handbrake and wait a bit longer for the CPU to do its work, for now.