Intel Gearing Up To Compete With Nvidia In The Coprocessor Segment Of The HPC Data Center Market

In our previous analysis, we discussed how Intel is competing with Nvidia in the data center coprocessor market. In this analysis, we specifically discuss details about the product offering of the two companies in this market that differentiates the two companies. It is worth noting that a coprocessor can offload processor-intensive tasks from the CPU, resulting in improved system performance. Further, it is economical to have a coprocessor for a specific task, as compared to having a costly integrated CPU that caters to a large number of functionalities even when you don’t need them.

Nvidia’s GPUs Highly Suitable As A Coprocessor

Nvidia is the leading manufacturer of GPU(Graphics Processing Unit), a processor with multiple cores that is optimized for the compute-intensive functions involved in processing graphics. These computational capabilities make GPUs ideally suited for use as coprocessors in High Performance Computing (HPC) environments.  Over the last decade, in particular, the use of GPUs in this capacity has gained traction. This is because the nature of computations involved in deep-learning algorithms used in HPC and computer graphics are similar. It is worth noting that GPUs have a parallel architecture with hundreds of cores, making it highly suited for matrix and vector operations in both deep learning and 3D computer graphics. For example, Nvidia’s Tegra X1 GPU has 256 cores running its Cuda operating system for partioning and load balancing workloads. The Tegra X1 CPU, in constrast, has for ARM 64b cores, while the Tegra K1 GPU (designed for automotive applications) has 192 cores.

Intel’s Xeon Phi Coprocessor And Its Recent Acquisitions Have Expanded Its Reach In The Coprocessor Market

In contrast to a GPU with thousands of cores that allows parallel computations, a single core processor can perform only serial computations, processing only one element at a time. Though Intel has multi-core processors that can allow for parallel computations, they cannot match up to the speed of GPUs. However, they are designed to be coprocessors and share a common code base and development tools with the main processor.  In 2012, Intel launched its Xeon Phi processors, a series of massively parallel multicore processors, which expanded its offerings in the HPC market. And it updated this coprocessor family earlier this year.  Additionally, Intel acquired Altera in 2015, gaining access to its FPGA (Field Programmable Gate Array) technology, further expanding Intel’s capabilities to address the coprocessor needs of computing in the future with the FPGA technology.

Furthermore, Intel recently announced the acquisition of the deep learning technology startup Nervana Systems, which according to Diane Bryant, has a fully-optimized software and hardware stack for deep learning and an advanced expertise in accelerating deep learning algorithms. This acquisition which can help Intel expand its capabilities in the field of AI (artificial intelligence) and compete directly with Nvidia. Sources report that Nervana has gained traction against Nvidia with its Cuda-compatible Neon software offering. The company is also developing a Deep Learning accelerator (i.e., coprocessor ) that is expected next year.

Who Leads The HPC Segment Currently?

Though both Nvidia and Intel are competing head on for coprocessor sockets in  the HPC segment of the data center market, it is the performance that will hold the key to success of either of the company. Currently, it is debatable as to which one – Intel’s Xeon Phi processor family (formerly code-named Knightsbridge) or Nvidia’s Tesla processors – is better in terms of performance. Both the companies claim to beat the other in terms of performance, though in truth they offer distinctive archetectures and compete on highly technical features.

However, it is quite clear that currently, Intel dominates the market for CPUs in high performance computing, while Nvidia captures a large majority of the coprocessor sockets.  According to the most recent Top 500 Supercomputer report, around 455 (or 91%) use Intel microprocessors, 55 (or 26%) use IBM Power microprocessors, and 13 (or 3%) use AMD microprocessors. However, of the 500, 93 (or 19%) are designed to use accelerator/coprocessor devices, of which 67 (or 72%) use Nvidia coprocessors. Going ahead, it remains to be seen if Nervana’s advanced HPC stack will successfully integrate into Intel’s resource-rich environment, and if it can help Intel to gain a larger share of this market.