Accelerator cards adapt for data centres and AI
At the Xilinx Developer Forum, Xilinx launched what it claims to be the world’s fastest data centre and AI accelerator cards. The Aleveo U200 and Alveo U250 are based on UltraScale+ FPGA and are claimed to deliver four times the performance, compared with graphics processing units (GPUs) and 90 times the performance compared with computing processing units (CPUs).
The accelerator cards are designed for use in industry-standard servers across cloud and on-premise data centres. Xilinx says customers can expect low latency performance for real-time machine learning inference as well as video processing, genomics, and data analytics.
The FPGA-based accelerator cards’ hardware can be reconfigured to accommodate workload alterations and new standards and updated algorithms.
The Aleveo U250 is optimised for machine learning applications. It increases real-time inference throughput by a factor of 20 when compared with high-end CPUs, and more than four times for sub-two-millisecond low-latency applications compared with fixed-function accelerators like high-end GPUs, says Xilinx. In addition to reducing latency by a factor of three, compared with GPUs, the accelerator cards also accelerate applications such as database searches up x90, compared with CPUs, claims Xilinx.
Although only launched last week, there is an ecosystem of partners and OEMs; companies have developed and qualified applications in AI/ML, video transcoding, data analytics, financial risk modelling, security, and genomics. OEMs, including Dell EMC, Fujitsu and IBM, are also collaborating with Xilinx to qualify multiple server units with Alveo accelerator cards.
“FPGA-based acceleration solutions in modern data centres are gaining popularity as accelerators that can be programmed and reprogrammed easily as users see fit,” said Ravi Pendekanti, senior vice president, product management and marketing, Dell EMC Servers & Infrastructure Systems.
“With the IBM Power Systems AC922 server, IBM has already demonstrated that we have the best platform for enterprise AI training,” said Steve Sibley, vice president of IBM Cognitive Systems. “IBM sees inference as a key component of a complete, end-to-end AI platform, and POWER9’s leadership I/O bandwidth for data movement makes it an ideal pairing with Xilinx’s new Alveo U200 accelerator card to bring inference to the enterprise.”