AMD and Xilinx Reveal industry-shaking record for AI Inference

At the Xilinx Developer Forum, Xilinx CEO, Victor Peng, together with AMD CTO, Mark Papermaster, unveiled an industry shaking real-time AI inference processing record.

The AI Inference record was realized by using Xilinx new Alveo acceleration cards together with AMD’s EPYC CPUs. The hardware combination, with their respective technologies, was able to reach a whooping 30,000 images per-second inference throughput in real-time. The record might be an industry-first with a possibility to hold a world record, but it is yet to be verified by Guinness.

The system that produced the record was equipped with 8 of the freshly-announced Xilinx Alveo U250 acceleration cards and leveraged the industry-leading PCIe connectivity of two AMD EPYC 7551 server CPUs. The inference performance was handled by Xilinx ML Suite, which allows developers to optimize and deploy accelerated inference and support numerous machine learning frameworks such as Tensorflow. The benchmark ran a batch size of 1 and Int8 precision on GoogLeNet.

AMD and Xilinx have shared a common vision in revolutionizing computing to heterogenous system architecture and have a long history of technical collaboration. Both companies see a bright future in their collaboration and will continue to work together, as AMD EPYC server CPUs and graphics processors are tailored fit for Xilinx present Alveo portfolio and its forthcoming Versal portfolio.