Xilinx uses Micron’s clever NOR flash to speed up its AI accelerators

Xilinx Versal

High-speed Xcella memory for the adaptive compute acceleration platform

by Max Smolaks 1 October 2019

American chip designer Xilinx, which specializes in programmable logic devices, has announced that its Versal accelerator lineup, intended primarily for AI inference, will be equipped with high-speed Xccela flash memory from Micron.

Xilinx is the granddaddy of custom silicon, credited with inventing the first field-programmable gate arrays (FPGAs) back in 1985. FPGAs have since become a market in their own right, being employed for tasks where using conventional CPUs is impractical.

Xilinx’s long-time competitor Altera was acquired by Intel in 2015 in an all-cash transaction valued at approximately $16.7 billion, serving as a testament to the growing importance of this chip category.

According to research by Tractica, the overall market for deep learning chipsets – which include CPUs, GPUs, FPGAs, application-specific integrated circuits (ASICs) and others – will reach $72.6 billion by 2025.

Not resting on its laurels, Xilinx recently developed a new chip type called the adaptive compute acceleration platform (ACAP), especially suitable for AI workloads and commercialized as the Versal family. And it’s Versal that is going to receive a shot in the arm from Micron’s new memory.

Xcella is a brand of NOR flash – which is different from NAND flash widely used in both enterprise and consumer storage products. Whereas NAND memory may only be written and read in blocks – like a hard drive – NOR memory allows the device to read and write individual bytes. From this pointy of view, NOR is like RAM, and programs stored in NOR flash can be executed directly without needing to be copied into main memory first.

According to Micron, Xcella will boost the boot, dynamic configuration performance and overall system responsiveness of the Versa platform by up to eight times, when compared to prior-generation FPGA platforms using older NOR flash.

The company added that Xcella delivers up to 400MB per second in double data rate mode while consuming 30 percent less effective energy per bit over traditional quad SPI NOR flash.

“Xilinx’s choice to support Xccela flash in its Versal ACAP is a testament to the growing importance of bandwidth for memory and storage used in artificial intelligence applications,” said Richard De Caro, director of NOR flash for Micron’s Embedded Business Unit.

“As autonomous driving vehicles incorporate higher levels of artificial intelligence inference capabilities into advanced driver-assistance systems (ADAS), Xccela flash enables Versal ACAP-based systems to power up and configure rapidly to meet ADAS application requirements.”