SK Hynix has introduced one of the industry's first computational memory solutions (CMS) — a 512GB memory expansion module with compute capabilities and a PCIe 4.0 x4 interface with the CXL protocol on top.
Aimed at big data analytics applications, the device is said to increase performance by several times when compared to CPUs in specific computations.
SK Hynix's CMS carries four 128GB conventional DDR memory modules and is equipped with a system-on-chip (or maybe a set of chips) that can handle machine learning KNN (k-nearest neighbors) classification algorithm acceleration, data filtering, and load balancing that features a PCIe 4.0 x4 interface which supports both CXL.mem and CXL.io protocols. The CMS card also has two network connectors to connect to other nodes.
We're not sure about the card's power consumption, but it has an eight-pin auxiliary PCIe power connector that can deliver up to 150W of power.
SK Hynix says its prototype CMS can perform data filtering and real-time data processing operations faster than various many-core CPUs in its big data analysis platform designed together with SK Telecom.
SK Hynix did not quantify advantages of its computational memory solution over traditional processors in its press release. However, one of the photos demonstrated by the company appears to show graphs and numbers depicting benefits provided by the device.
"Through the internalization of computational functions, CMS enabled performance several times faster than that of dozens of CPU cores in specific computations," said Park Kyoung, Head of Memory System Research at SK Hynix. "Considering that this is just a prototype, we think we can improve the performance even further and are considering applying the technology to other applications such as big data."
The CXL CMS prototype designed by SK Hynix was demonstrated first at the Open Compute Project (OCP) Global Summit. The company also demonstrated its big data analysis platform optimized together with SK Telecom.
Maybe RAM could use some "local" compute.
Seems a bit "Anti-cloud" - but maybe I'm just seeing what I want to see.