After NVIDIA announced their earnings from last quarter, everyone has been bullish about the company. After huge growth in the gaming sector, NVIDIA also presented a substantial (205%) growth in the data center field. This means that as AI and machine learning start to find new uses for the enterprise world, CTOs are turning to NVIDIA for the chip sets necessary to keep the data centers running. Even AMD’s new offering of it’s super fast GPU system doesn’t stand a chance against the market dominance of NVIDIA. However, a lesser known technology is making a play for market share in the AI/big data world. The technology is known as FPGA, or field programmable gate array.
A group of engineers at Intel recently published a paper detailing the comparison between the FPGA systems with current GPU standards. The paper made it clear that as FPGA technology and algorithms increase, they will soon become far more adept at handling the big data stress load that GPUs, and all at a lower ops/watt ratio. This means that the technology is more geared toward the future than the GPU, which may be seeing it’s industry domination fading. The FPGA has far more extensive customizability than the GPU and therefore is capable of being integrated into the developing systems around AI needs. Long story short, the FPGA is better equipped to handle what’s coming than the GPU, and that includes the best stuff NVIDIA can offer the market today (Titan X Pascal).
The top makers of FPGA technology are currently Xilinx and Intel, with these two companies holding 90% of market share. Microsoft is using the Intel version of the FPGA in all it’s data centers now, and is promising to increase it’s usage in the near future. What’s more, Xilinx just announced a huge new contract with Amazon to provide them with the accelerated AWS (Amazon Web Services). Amazon has chosen the FPGA model with the intention that it provides far more customizable solutions over greater data sets with less power – all the advantages over the GPU. Such huge market players moving toward this technology should indicate that the market shift is coming away from single processing systems toward the array system. This major win for the company was nearly ignored by the market, but the truth is always better than fiction. While this is still a tiny portion of the overall market in the processor field, the writing is on the wall.
It seems that a good play for companies looking for AI/ML solutions is to analyze which system provides the best solutions for their infrastructure. If the solutions offered by simple processing units are not comprehensive, it may be worth looking into a FPGA solution. Further, for companies and investors looking for a smart money play, Xilinx may be a great option. If you had the chance to go back in time and buy NVIDIA in 2007, you’d be on a beach somewhere right now. That chance may have just come again.