Accelerating personalized recommendations

Associate Professor Mark Hempstead and colleagues proposed a new near-memory hardware/software solution that could provide significant memory energy savings.
Computer memory chip

Personalized recommendation systems are found across the internet; they recommend content such as videos and information about projects. However, the systems that compile and deliver these recommendations can be limited by the system’s speed in accessing memory and its memory resources. In collaboration with colleagues at Facebook FAIR SysML Research Boston, Associate Professor Mark Hempstead developed a hardware/software solution that accelerates machine learning when it comes to these personalized recommendations for users.

In a paper presented at the International Symposium on Computer Architecture (ISCA 2020), Hempstead and colleagues proposed a solution to address the memory limitations: a new, near-memory hardware architecture that can accelerate personalized recommendation inference. The researchers found that that system, called RecNMP, provided 45.8% memory energy savings and a throughput improvement of 4.2 times existing systems.

Read more from ISCA 2020: “RecNMP: Accelerating Personalized Recommendation with Near-Memory Processing.”