Abstract
Machine learning (ML) recommendation workloads have demanding performance and memory requirements and, to date, have largely been run on servers with x86 processors. To accelerate these workloads (and others), Esperanto Technologies has implemented over 1,000 low-power RISC-V processors on a single chip along with a distributed on-die memory system. The ET-SoC-1 chip is designed to compute at peak rates between 100 and 200 TOPS and to be able to run ML recommendation workloads while consuming less than 20 W of power. Preliminary data presented at the Hot Chips 33 Conference projected over a hundred times better performance per watt for an Esperanto-based accelerator card versus a standard server platform for the MLPerf Deep Learning Recommendation Model benchmark.
Highlights
Author Bio: DAVID DITZEL is the founder and Executive Chairman of Esperanto Technologies
The light green bars compare the relative performance of each accelerator card to the performance contributed by one Intel Xeon server processor
The dark green bars compare the relative performance per watt of each accelerator card to the TDP power of one Xeon
Summary
ML Recommendation workloads in hyperscale datacenters have some of the most demanding performance and memory requirements, and to date have largely been run on servers with x86 processors. Often these servers have an available slot for a PCIe accelerator card, but an accelerator card needs to meet some key requirements:. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Esperanto picked the RISC-V instruction set as the foundation for its general-purpose programmable solution
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
