» Current | 2022 | 2021 | 2020

Linley Newsletter

Xilinx Delivers Server Acceleration

February 19, 2019

Author: Bob Wheeler

Embracing vertical integration, Xilinx is shipping complete solutions for AI inference acceleration. They comprise Alveo FPGA-based PCI Express cards, programmable-logic configurations, and software that works with common neural-network frameworks. This move places the company in direct competition with Nvidia’s Tesla T4 line for inference. The Alveo cards address other markets—including financial technology, database acceleration, video processing, and scientific computing—by supporting third-party intellectual property (IP) and software.

The initial Alveo line includes the U200 and U250, which feature 16nm UltraScale+ Virtex FPGAs as well as DDR4 SDRAM and are already in production. In 1Q19, Xilinx sampled the Alveo U280, which has an FPGA with in-package HBM2 memory; this high-speed memory should increase performance on large neural networks. All three cards have a 225W TDP and provide a PCIe Gen3 x16 interface. The U280 additionally offers PCIe Gen4 x8 (for compatibility with next-generation servers) and a CCIX cache-coherent host interface.

Previously, customers had to program the FPGAs using standard hardware-design tools, but Xilinx now provides its ML Suite, which includes an “overlay” that preprograms the FPGA as an AI inference accelerator. The ML Suite also has drivers for Caffe, MXNet, and TensorFlow that allow neural net­works from these frameworks to run on the FPGA accelerator. Employing the FPGAs’ DSP blocks, the cards deliver up to 33.3 TOPS for 8-bit-integer (INT8) data.

The Alveo cards compete directly against AI accelerators from Nvidia, the market leader in data-center deep-learning acceleration. Xilinx will also compete with several new entrants. Intel is shipping what it calls a Programmable Acceleration Card (PAC) based on its midrange Arria 10 GX FPGA. Like Xilinx, it provides precompiled FPGA code for neural-network acceleration that works in concert with its OpenVino tool kit. Startups targeting data-center inference include Habana, which is now shipping its Goya HL-1000 card.

Subscribers can view the full article in the Microprocessor Report.

Subscribe to the Microprocessor Report and always get the full story!

Free Newsletter

Linley Newsletter
Analysis of new developments in microprocessors and other semiconductor products


Linley Spring Processor Conference 2022
Registration now open: April 20-21, 2022
Hyatt Regency Hotel, Santa Clara, CA
Linley Fall Processor Conference 2021
Held October 20-21, 2021
Proceedings available
Linley Spring Processor Conference 2021
April 19 - 23, 2021
Proceedings Available
More Events »