Semiconductor company Xilinx said Tuesday it has created a new classification of processor called ACAP — Adaptive Compute Acceleration Platform — to be used in an array of technology areas, including artificial intelligence.
Xilinx, based in San Jose, Calif., is known in the semiconductor market as the company that created the FPGA (field-programmable gate array), an integrated circuit that can be configured by a customer after manufacturing. FPGAs are used for quick deployment of specialized processing algorithms and for prototyping fixed-logic chips called ASICs. FPGAs provide flexibility in design and the ability to reprogram after deployment, but at the expense of die size, cost and power. Microsoft MSFT, -0.06% used Intel-built FPGAs in its Project Brainwave cloud-based AI compute solution.
Xilinx XLNX, +1.25% is the biggest player in the FPGA market, with $2.3 billion in revenue last fiscal year and $622 million in profit. It competes in areas of network processing and AI compute against INTC, +3.98% Nvidia NVDA, -0.33% and others, but this new product line shifts its direction dramatically.
The Xilinx Versal family will still include a programmable portion called the “Adaptable Hardware” block, but the addition of two high-performance Arm processor cores, two low-power Arm processor cores and dedicated AI compute engines custom-designed by Xilinx for a complementary set of processing.
Xilinx says its Versal chips will offer two to eight times the inference benchmark performance against Nvidia GPUs available today, while being up to four times more power-efficient. Those same chips could be 43 to 72 times faster in the same inference workloads when compared to Intel Xeon processors.
Potential hurdles
The new Xilinx Versal chips aren’t slated for release until the second half of 2019, so the market still has 10 to 12 months to react. It’s not only likely, but assured, that Nvidia will have different products on the market by that time. New competition could drastically shift the messaging that Xilinx is presenting today.
Another area of concern for Xilinx is software integration and development tools. FPGAs are notoriously difficult to program for, and with a brand-new AI engine architecture for engineers to work with, Xilinx will need to pour significant money and resources into infrastructure and marketing. Nvidia spent more than a decade building the software ecosystem for its GPUs with CUDA, including tool sets, books, college curricula and more. Xilinx will need to be just as aggressive.
A big change for Xilinx
Though the emphasis is on the versatility of the new product portfolio, this marks a stark shift in how Xilinx positions its products. Versal is targeted at specific customers and market segments as opposed to primarily being a development-tool provider to others entering those same markets.
Xilinx believes this new chip will allow it to address applications ranging from data-center workloads in the cloud, to networking infrastructure products both wired and wireless, to edge compute solution for industrial automation, autonomous vehicles and more. The key driver for the company’s confidence lies in the adoption of AI models across all of those markets, and the performance and efficiency of its custom designed AI engine.
Nvidia is the clear leader for AI training systems, but Xilinx believes if it can capitalize on the growth of the inference market (the application of deep-learning models in consumer or cloud environments), it can position the company in a new direction. Xilinx showed data from Barclays Research projecting the (total addressable market) TAM on inference processing to match that of AI training by 2021 and to more than triple it by 2023.
By building a product with customized AI computation, along with standard Arm processing cores and programmable FPGA capability, Xilinx said the Versal product family creates a platform offering “whole application acceleration.” GPUs can only solve part of the problem for AI workloads, and customers will migrate to options with more flexibility, the company said.
Ryan Shrout is the founder and lead analyst at Shrout Research, and the owner of PC Perspective. Follow him on Twitter @ryanshrout.
Get the top tech stories of the day delivered to your inbox. Subscribe to MarketWatch's free Tech Daily newsletter. Sign up here.
