How to apply Ethereum mining hardware in deep learning
by using a small platform for GPU based computing
When we think of computer processing power, our first instinct is to consider the speed and architecture of a computer’s central processing unit (CPU). As it turns out though, the consumer graphics cards produced for the PC gaming market have far more computing power for many types of tasks.
It’s common knowledge that the Proof of Work algorithm that the Ethereum blockchain depends on is largely computed by consumer graphics cards produced for the gaming market. These same cards also outperform all other options for machine learning and artificial intelligence both in terms of raw power and cost-per-performance.
That’s always surprised us given that NVIDIA also produces dedicated cards for this type of processing in the USD 3000–5000 range, yet they lag behind much cheaper consumer graphics cards in terms of machine learning performance.
Since we were looking for new ways to apply Ethereum mining hardware, we spend some time developing a test platform using 6 NVIDIA GPUs. As anyone who has done cryptocurrency mining knows though, the mining rigs crash pretty often when they’re optimized for maximum productivity. This is OK for mining, but not acceptable for high performance computing so we needed to make a few changes.
The first step was to improve ventilation and cooling. The open-air cases that most mining rigs use are terrible in this regard — the purpose of a well built case is to promote sensible airflow and keep parts cool. Open-air mining rigs more or less just blow the hot air from some parts onto other parts. The solution didn’t need to be fancy — just sectioning off the graphics cards into their own small section so it can have dedicated ventilation.
The second step was choosing fans. We went with 6 small metal 230VAC ball-bearing industrial fans (3 ingress, 3 egress). These are designed to run for long periods of time, and receive power independently of the computer.
Finally, we installed Ubuntu Linux 16.04 LTS. Linux has good support for GPU computing through the CUDA toolkit, as well as fantastic remote management out of the box.
To stress test the new machine, we let it mine Ethereum for a week at maximum stable overclock in a small room with no air conditioning and only passive ventilation (an open window). Previously, the graphics cards could hit temperatures of nearly 80 degrees centigrade under these conditions, and now they never reached 60 degrees.
Overall the performance was acceptable and we now have a small platform for GPU based computing. Our plans for it mainly include blockchain research, but we certainly can’t rule out AI yet!
If you’d like to read more about using consumer graphics cards for machine learning, we highly recommend this excellent article detailing the costs and benefits of different cards by Slav Ivanov : https://blog.slavv.com/picking-a-gpu-for-deep-learning-3d4795c273b9
