Intel's recent claims that CPUs are better than GPUs when it comes to deep learning on neural networks has sparked a rebuttal from Nvidia.
In case you don't know what this is all about, here's a short recap:
Machine learning is currently a really big deal. It's a huge market with untapped potential in many industry verticals, which is why a lot of different companies are trying to get in on the action. It is widely taken as a fact that GPUs are a better solution than CPUs when it comes to deep learning, because neural networks require low precision computation, and not high-precision, which is what CPUs are usually made for.
But Intel has recently said that using CPUs actually yields better results.
In this brochure, it says that four Knights Landing Xeon Phi chips were 2.3x faster than “four GPUs”, and that Xeon Phi chips scale 38 per cent better across multiple nodes, among other things.
Enter Nvidia. The GPU-building company said Intel actually used old and outdated information, which renders their results invalid. The company recently switched from a 28nm planar process to a 16nm FinFET one, resulting in 'drastic increases in performance'.
Four Nvidia previous-gen Maxwell GPUs are 30 per cent faster than four Intel Xeon Phi servers, the company claims.
" It’s great that Intel is now working on deep learning. This is the most important computing revolution with the era of AI upon us and deep learning is too big to ignore. But they should get their facts straight,” Nvidia concluded. Ouch.
Image source: Shutterstock/Katherine Welles