Diane Bryant, head of Intel's (NASDAQ:INTC) datacenter group, said at Computex in Taipei that the company's Knights Landing ("KNL") processors could be the perfect fit for AI (artificial intelligence) in terms of running ML (machine learning) or DL (deep learning) workloads. KNL represents Intel's second generation (x200) Xeon Phi family.
According to Bryant, since the latest Xeon Phi coprocessors feature up to 72 cores, and each core has 2 Intel AVX-512 SIMD processing units for delivering improved per-core floating-point performance, these processors are ideal for running ML/DL workloads. However, it remains to be seen if these processors can surpass Nvidia's (NASDAQ:NVDA) new Tesla P100 GPU accelerators, which are currently considered the best fit for ML/DL.
Image Credit: Intel
Nvidia Needs To Retain Its Competitive Edge
Nvidia jumped on the ML/DL bandwagon with its powerful GPUs coupled with better algorithms. Although it's not clear yet if Intel can displace Nvidia's Tesla P100 GPU accelerators which are based on the latter's new Pascal-based GP100 GPU, Intel is certainly well positioned than before.
Pascal is Nvidia's 5th generation CUDA architecture. The company took three years and spent $3 billion to build Tesla P100, which is currently in volume production. Players like IBM (NYSE:IBM), Cray (NASDAQ:CRAY), Hewlett-Packard Enterprise (NYSE:HPE) and privately-held Dell are already using P100 for their upcoming HPC (high performance computing) servers and supercomputers.
Delivering the required computing horsepower for HPC and supercomputers via accelerated computing is nothing new. Regarding the performance of its earlier version of Tesla K80 GPU accelerators compared to Xeon Phi, Nvidia said that simply recompiling and running applications on Intel's Xeon Phi without changing CPU code usually resulted in deceleration instead of acceleration. Further, programming a GPU and Xeon Phi used to need similar effort, but the GPU delivered meaningfully better performance. According to Nvidia:
Once you see the facts, a better understanding of accelerated computing emerges. Today, a GPU provides double the performance for essentially the same developer effort. GPUs are the logical choice for accelerating parallel code. In part, this could be why scientific researchers have published with GPUs more than 10:1 over Intel Xeon Phi this year. And why NVIDIA GPU is favored more than 20:1 over Xeon Phi in HPC systems today.
Image Credit: Nvidia
However, these facts were presented three years ago, and today's scenario could be entirely different with Intel significantly improving its latest Xeon Phi and supporting tools. In the following section, I will try to assess if Intel can displace Nvidia's GPUs this time around for running ML/DL workloads.
Intel Better Prepared This Time Around
Intel's KNL x200 Xeon Phi is capable of delivering 3 teraflops of FP64 compute horsepower, i.e., 64-bit double precision arithmetic. Although consumer grade GPUs don't support double precision arithmetic, Nvidia has made Tesla P100 to deliver 5.3 teraflops of FP64 compute power for beating the KNL Xeon Phi.
However, despite Nvidia's Tesla P100 offers higher standalone FP64 performance, in terms of performance per dollar, this is insignificant at the system level. Beating Intel in running ML/DL workloads could be difficult for Nvidia due to three reasons, which are given below:
- Intel's AVX-512 (advanced vector extensions 512) SIMD processing units will support ML/DL algorithms like floating-point multiply and fused multiply-add.
- Intel has empowered the latest Xeon Phi so that the OS can boot itself via a self-boot socket.
- Finally, with Intel's unique omni-path architecture (OPA) fabric technology supporting the Xeon Phi, the processor will be a supercomputing beast.
For programming the second generation Xeon Phi, similar effort by developers as for Tesla P100 might not result in significantly better performance for the latter, which was the case three years ago. In fact, Intel has made it a lot easier than before to parallelize ML/DL code with Intel MKL (math kernel library). The latest version of Intel MKL, i.e., the Intel MKL 2017 beta, includes new tools for optimizing ML/DL neural networks.
Better Choice For ML/DL Workloads: The Verdict
For raw computing horsepower, Tesla P100 should be the obvious choice. But for running ML/DL workloads, raw computing horsepower isn't always the priority. With cloud computing increasingly impacting ML/DL research, a comprehensive HPC platform with massive storage capacity is what researchers need. In addition, better optimization tools are also required.
I strongly believe that Intel's scalable system framework (SSF) HPC platform based on its omni-path architecture will make all the difference in favor of Intel. I shared my view regarding Intel's SSF approach in a previous article. When the latest Xeon Phi coprocessors will be used alongside Intel's new highest-end Xeon processors and 3D XPoint memory-based Optane SSDs, ML/DL workloads will certainly get a performance boost.
In addition, Intel is providing more tools in the hands of developers via software libraries. It is optimizing its core architecture also for open-source ML frameworks like Caffe and Theano. If raw computing horsepower isn't the priority, latest Xeon Phi should be the best choice.
AI will go mainstream very slowly with the advent of autonomous cars, smart homes and the like. Nvidia has developed sophisticated AI-based autonomous vehicle technology which could take a decade before going mainstream due to lack of adequate infrastructure.
Oftentimes, building a technology takes lesser time than building adequate infrastructure for its proper utilization. Nvidia will continue to generate most of its revenues from its graphics business despite developing advanced technology for autonomous vehicles. Intel, on the other hand, is continuously creating new revenue streams for the future. That's why I am more bullish on Intel.
Disclosure: I am/we are long INTC, IBM.
I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it (other than from Seeking Alpha). I have no business relationship with any company whose stock is mentioned in this article.