Graphics chip specialistNVIDIA has essentially run away with the high-performance-computing accelerator market. The company says its Tesla GPUs accounted for 85% of the market as of the middle of 2013, with Intel's competing Xeon Phi product line barely registering.
Continue Reading Below
Last year I wrote that the next version of Intel's Xeon Phi, called Knights Landing, posed a significant threat to NVIDIA's enterprise business. With the release of Knights Landing anticipated in the second half of this year, and with the first details of its successor, Knights Hill, being disclosed, it's time to look at how Intel's efforts in the HPC space could affect NVIDIA's Tesla business, which generates a few hundred million dollars per year for the company.
Intel's plan for HPC dominance
Knights Corner, the latest version of Intel's Xeon Phi, has not made much impact since shipping in 2012. Knights Landing will be different than its predecessor in three important ways.
First, it will be far more powerful. Knights Corner could do about one teraflop, or 1 trillion floating point operations per second, at double precision, while Knights Landing triples this performance to three teraflops. This puts Knights Landing in line with the performance of NVIDIA's most powerful Tesla GPU, the K80.
Second, Knights Landing will be available as both an add-on card and a stand-alone processor. The stand-alone version allows Knights Landing to act as the main processor, although since it will be composed of dozens of low-power Atom cores, single-thread performance will be poor. A likely configuration, then, will be a powerful Intel Xeon server CPU coupled with a Knights Landing processor.
This configuration comes with one important advantage compared to add-on boards. NVIDIA's GPUs need to communicate with the CPU, sending data back and forth between main memory and video memory, and it must do this over the PCI-E bus. This creates a potential bottleneck in the system, something the stand-alone Knights Landing avoids.
Continue Reading Below
Third, Intel manufactures its own chips, and Knights Landing will be built on its 14-nanometer process. NVIDIA relies on third-party foundries, and currently all of its products are stuck at 28nm. Apple and other high-volume foundry customers took all the available 20nm capacity last year, and it's unclear exactly when NVIDIA will issue any products below 28nm.
This gives Intel an important advantage that NVIDIA can't match. Knights Hill, the successor to Knights Landing expected in 2017 at the earliest, will be built on Intel's upcoming 10nm process. What node NVIDIA will be on at that point is an open question.
Intel has significant advantages, and it might just be a matter of perseverance before it can claim leadership in the HPC accelerator market. However, NVIDIA has some advantages of its own.
NVIDIA will be tough to beat
The architecture of NVIDIA's Tesla GPUs and Intel's Xeon Phi accelerators are very different. Knights Landing will be comprised of dozens of Intel's low-power Atom cores, the same architecture found in many Intel-powered tablets. NVIDIA's GPUs, on the other hand, contain thousands of very specialized graphics cores.
This gives Intel another advantage: Since Knights Landing is based on the same x86 architecture as the rest of Intel's CPUs, programming for it should be simpler than programming for an NVIDIA GPU, which uses a proprietary GPU compute language called CUDA. However, so much software has already been optimized for NVIDIA's GPUs that it likely won't matter.
The number of applications that already run on NVIDIA's GPUs, ranging from computational finance software to scientific simulations, is staggering. NVIDIA maintains a catalog of this software, which includes commonly used applications such as AutoCAD and industry-specific applications like those used in the oil and gas industry. Ultimately, software support sells the hardware, and NVIDIA has a big head start in that area.
NVIDIA's Tesla GPUs should be fairly sticky, given the time and effort already put in porting software to the CUDA language. NVIDIA's GPUs have become the standard in high-performance computing, and displacing that standard will take more than competitive hardware from Intel.
NVIDIA also has a solution to the issue of sending data back and forth between the CPU and the GPU. The company has developed a technology called NVLink, which allows the processors to exchange data five to 12 times faster than PCI-E, and it is already set to be used in two new supercomputers being built by the U.S. Department of Energy. This reduces the advantage the stand-alone Knights Landing has over NVIDIA's GPUs.
Intel has far more resources than NVIDIA, and if it throws enough money into the effort it's hard to imagine the company not eventually winning a significant chunk of the market, especially considering its manufacturing advantage. The main question is whether the opportunity is large enough, relative to Intel's size, to warrant this effort.
In the end, the market might grow quickly enough to enable NVIDIA and Intel to both succeed. But if Intel's Knights Landing doesn't put up a much better showing than its previous efforts, the company may need to rethink its strategy.
The article Intel's Challenge to NVIDIA Is Coming This Year originally appeared on Fool.com.
Timothy Green owns shares of Nvidia. The Motley Fool recommends Apple, Intel, and Nvidia. The Motley Fool owns shares of Apple. Try any of our Foolish newsletter services free for 30 days. We Fools may not all hold the same opinions, but we all believe that considering a diverse range of insights makes us better investors. The Motley Fool has a disclosure policy.
Copyright 1995 - 2015 The Motley Fool, LLC. All rights reserved. The Motley Fool has a disclosure policy.