Massively Parallel GPU Computing. Pretty cool stuff, right? Who would have imagined 10 years ago that today, we’d be using GPUs (otherwise known as Graphics Processing Units) for computational research & calculation rather than for eye-popping 3D graphics! Ok, MAYBE it’s not too far-fetched to think that Jen-Hsun Huang had already figured that out… Today, NVIDIA® enjoys accolades from both consumers (GeForce™ / Tegra™) and professionals (Tesla™ / Quadro™) alike. They’ve effectively & successfully created GPUs that are capable of handling a range of tasks, from creating compelling, real-time visual effects in games, to modeling a Boeing 787 Dreamliner’s engines, to simulating the aerodynamics of a Ferrari LaFerrari. Incredible, really.
AMAX ClusterMax SuperG GPU Cluster configurable with the latest Nvidia Tesla Kepler K10 / K20X GPU Accelerators
At AMAX, we may not do games, but we do build High Performance Compute clusters & servers that take full advantage of NVIDIA’s Tesla line of GPUs. Our ClusterMax™ SuperG, for example, can be configured with Tesla K10 and K20 / K20X GPUs and is a real computational powerhouse of a machine. The question is… which GPU is for you, and why? That’s a common question that we come across and it’s not an answer that is ubiquitous among all users or all applications. By solely looking at the specs, the K10 has greater SP (single precision) floating point performance than the K20 / K20X, but clearly lesser DP (double precision) performance. So the choice is simple, right? In the case where your application needs DP, then the choice is easily made – K20 / K20X takes the cake. But with SP-focused applications, the decision is a bit fuzzier. Let’s give a live example..
One of our customers, who requested to remain anonymous, recently acquired a startup company that specializes in neural networks. The acquisition was made to enhance search algorithms and go beyond the traditional ability to identify pieces of content, images, voice, text and more. The startup’s research is known to have profound implications on speech recognition, computer vision and machine language understanding. But I digress… our friends have found that their proprietary application runs much better on a Tesla K20X compared to a Tesla K10 – even though their application is SP-focused. When investigating why, they deduced the following:
“A K10 consists of two rather slow GPUs. Together they’re quite powerful, and so a K10 is great if you intend to run two independent jobs on it.. or jobs that don’t need to communicate often. A K20X consists of one VERY fast GPU. This is what you should get if you need to run code that requires frequent communication. We happen to run very big models, and because the various parts of the models have to communicate with each other frequently, they have to live on one GPU. Therefore even though the K10 consists of two GPUs, the larger & faster single GPU ends up being more efficient for our purposes.”
By now you’re probably saying, “So I won’t really know which GPU is best for me without running my code on both!” And in many cases, that statement would be correct. But we can help! At AMAX, we’ve set up a Tesla Test Drive program where you can run your code on a small-scale ClusterMax SuperG cluster… for FREE! K10 vs K20 / K20X ? We’ll help you find out first hand which GPU is right for you. Pretty cool, right?