Bitcoin Currency and GPU Mining Performance Comparison
Testing Configuration and Software Setup
True ease of use is something that the Bitcoin ecosystem doesn't really have yet though they are steadily improving on it. You'll need a couple different items up and running on one or more machines to really start with your mining experience.
The first thing you'll need is the Bitcoin client application that acts like your wallet and actually accesses your wallet.dat file. While this doesn't necessarily need to be running on the same hardware that is doing the mining, you'll need to run this to get your key information to share with the mining apps.
For your mining application, there are several options including some command-line based apps and graphical ones. For the quickest setup and configuration time we liked GUIMiner, seen above. The interface you use does not necessarily determine the kernel you use for computing the Bitcoins and which kernel you use can alter performance pretty dramatically. In its infancy the Bitcoin community ran CPU-based kernels until the performance difficulty got to a point where they were incredibly inefficient leading to the creation of several GPU-based designs.
For our testing we went with the poclbm kernel that is built around OpenCL and works with AMD Radeon HD 4000 series and above and NVIDIA GeForce 8000 series and above graphics cards. There definitely are other options out there for Bitcoin mining and many enthusiasts argue that some perform better than others across different ranges of CPUs and GPUs but in terms of popularity today, poclbm seems to be the winner.
The above image shows us actually running a pair of the kernels, one for each GPU on a multi-GPU graphics card. If you have more than one GPU in your system, whether on a single card or multiple, you need only assign a kernel to each available processor to max out your processing performance.
(Side note: because it is built on OpenCL, you can actually run this on CPUs that have compliant OpenCL stacks. However, it is not the most efficient on that class of processor by any means.)
When running a Bitcoin mining application be prepared for a lot of GPU utilization but not much on the CPU side of things.
Here you can see our Core i7 Sandy Bridge based processor is not getting a heavy workout while running the GUIMiner application with our OpenCL-based client focused on the GPU. Looking at the graphics card workload however...
The ASUS ARES card (dual Radeon HD 5870 GPUs) is working hard with a 99% GPU load reached and temperature slowly rising. If you have a GPU with a loud fan or are sensitive to the heat created by your graphics card then this mining process might not be for you!
For our testing we ran the Bitcoin clients on our standard GPU testing bed built out of the following:
- Testing Configuration
- ASUS P6X58D Premium Motherboard
- Intel Core i7-965 @ 3.33 GHz Processor
- 3 x 2GB Corsair DDR3-1333 MHz Memory
- Western Digital VelociRaptor 600GB HDD
- Corsair Professional Series 1200w PSU
- NVIDIA Driver: 275.33
- AMD Driver: 11.6
Our graphics card selection was based on trying to compare some current options to some previous generation cards that are likely to already be in the hands of potential GPU Bitcoin Miners. Here is the lineup with a few curveballs tossed in:
- GeForce GTX 285 - $300
- GeForce GTX 295 - $289
- GeForce GTX 460 - $160
- GeForce GTX 560 Ti - $319
- GeForce GTX 580 - $469
- GeForce GTX 590 - $749
- Radeon HD 4890 - $240
- Radeon HD 5750 - $115
- Radeon HD 5830 - $129
- Radeon HD 5970 - $620
- Radeon HD 6850 - $159
- Radeon HD 6990 - $750
- Radeon HD 5870 x2 (ASUS ARES) - $1100
- Radeon HD 5870 x2 (Overclocked) - $1100
- AMD A8-3850 APU - $139
- "The Beast" - $1710
We have covered the bases of the last several years by starting with the HD 4890 and GTX 285 cards of yester-year. We included a range of modern cards including the very popular GeForce GTX 460 and the lower end Radeon HD 5750. Dual-GPU cards make a frequent showing with the GT X 295, GTX 590, HDF 5970 and HD 6990 as well as the ASUS ARES in a standard and overclocked setting. Standard clock rate on the ASUS ARES is 850 MHz and our overclocked setting pushed that to 1005 MHz - an 18% increase. Basically, we just wanted to see how high we could push that $1100 graphics card.
The big outlier is the new AMD A8-3850 APU released this month that combines a quad-core CPU and "discrete class" GPU on a processor. The Radeon HD 6550D GPU on that die (as it is branded) has 400 stream processors and uses a DDR3 memory interface that is shared with the x86 cores. Because the computing process at work in Bitcoin mining is not memory dependent, we kind of expected the APU to do well for its price and position.
The pricing listed here is used throughout our performance review to judge value and profitability. Keep in mind that some of these numbers were hard to really nail down especially for cards like the GTX 285, GTX 295, HD 4890, HD 5970 and ASUS ARES that are hard to find anywhere but eBay and very small online stores. The prices here are my best estimates at what you would have to pay (on average) to acquire a card like this today.
You might also be wondering what "The Beast" is in our list above. That is a mega-crunching machine we put together after doing all of our other card testing to see just how much we could push out of a single system. Using the same base test bed, we installed the Radeon HD 6990 4GB, Radeon HD 5970 2GB (both dual-GPU cards) and the Radeon HD 6970 2GB single-GPU cards. While we wanted to include the ASUS ARES in this configuration we weren't given that option since it required three PCIe power connections and our Corsair AX1200 power supply only supplied us with six of them. You will have to wait until later in the article to see the results of that setup as I decided to leave it off the single card result graphs as it tended to skew the scale quite a bit.
What to look for
The first thing you are going to notice is that the AMD graphics cards solidly outperform the NVIDIA GPUs for reasons we are still diving into. The VLIW architecture at work on the 4000/5000/6000 series of cards is seeing some very high utilization by the poclbm kernel and it is definitely one of those few applications nearly reaching the theoretical limits of TFLOPs claimed by AMD over the years.
What else is there to evaluate?
- Pure Mhash/s rates - how fast is each GPU in computing the math required for Bitcoin mining? The higher the Mhash/s rate the faster the card and quicker you will get to finding the next coin in the currency.
- Performance per Dollar - Mhash/s/$ - This is probably the most important factor for users that might consider Bitcoin mining as a way to make money and pay for things they want to buy. Which card is going to bring the most "value" the mining experience?
- Performance per Watt - Mhash/s/watt - If you value your air conditioning bill more than most or maybe want to cram as many cards into an enclosure as possible for a mining power house you might want to know which cards and GPUs are the most power efficient.
- Dollars per day - Step 1: Mine. Step 2: ?? Step 3: Profit. How much money can you make on a given card on a daily basis? This metric will fluctuate from day to day based on the actual exchange rate of a Bitcoin with USD (or your own currency) but we will evaluate it based on the numbers as of this writing.
- Time to Graphics Card Payoff - Based on the amount of money you can earn per day, how long will it take you to pay off the card you purchased for this purpose and start making the aforementioned profits? Obviously for cards that are either end-of-lifed or just plain hard to find this is going to be a rough estimate (go ahead and find me an average price for a GTX 285 today) but it provides another useful data point for professional miners.
- One Year Profit - If you took that daily earned amount and could apply it perfect for one year (which we know you can't really because of the changing algorithm) and subtracted the cost of that graphics card, how much could you possibly MAKE in a year? The numbers might surprise you!
So there you have it - let's jump into the results and see what our testing brought forth with more details and explanations along the way!