Intel Xeon Phi Benchmarks And What Nvidia Has To Say About Them
Intel published some Xeon Phi benchmarks recently and the company claimed that their architecture that consisted of small Atom CPUs was more efficient than GPUs when it came to deep learning. It seems that Nvidia does not agree with these claims and the company has made a post about how these claims are false.
Whether the Xeon Phi benchmarks claims hold true or not. One thing is for sure then it comes to deep learning GPUs are the way forwards. CPUs are built for high-precision computation the deep learning requires the exact opposite. This is why GPUs are used for this purpose. There might come a day when GPUs are replaced with something better but today is not that day.
Over the course of several years, Nvidia has made GPUs more efficient for deep learning and the company has also invested in software in order to make things easier for developers who are looking to work in this field. This is one of the reasons why people go with Nvidia GPUs instead of AMD.
According to the Xeon Phi Benchmarks, Intel claims that four Knights Landing Xeon Phi chips are 2.3 times faster than four GPUs. The company also claims that these chips can scale 38 percent across multiple nodes. Intel claims that a system consisting of 128 Xeon Phi servers is 50 times faster than single-Xeon Phi servers.
Nvidia, on the other hand, argues that Intel has used rather old data to make these claims and that this is not the fact. Nvidia claims that GPUs have become more powerful after they starting making them on the 16nm FinFET process. Besides that Nvidia has also optimized GPUs using a different software as well.
Nvidia argues that if Intel used up to date data for their tests then they would see that four previous generation GPUs are actually 30% faster than four Xeon Phi servers. Nvidia shares it’s beliefs that having fewer stronger nodes is far better than having multiple weak ones.
It is unclear why Intel chose to use older generation GPUs in their test and software that is 18 months old. The Maxwell GPUs have been in the market for some time now and are readily available.
These are the Xeon Phi Benchmarks. Who do you think is right there? Intel or Nvidia?