[Digital Today reporter Chi-gyu Hwang] Meta has developed a technology called Vistara that installs DDR4 memory removed from older servers into new servers and shares memory via its own CXL (Compute Express Link) chip without application-to-application latency problems, The Register reported on June 29 (local time).

A paper released by Meta said about 40% of its servers, or millions of machines, are handling some tasks because they cannot add more memory. Server lifetimes are 3 to 5 years, but memory can be used for 7 to 10 years.

Meta therefore applied an approach of removing DDR4 memory from older servers, installing it in new servers that use DDR5, and building a memory pool.

The problem is that mixing different types of memory leads to lower bandwidth, increased latency and greater management burden. Existing CXL products blocked reuse of memory modules, did not support DDR4, and were not suitable because of high power consumption and costs.

Against this backdrop, Meta designed the Vistara ASIC in-house and connected DDR4 memory to the host processor via a CXL 2.0 interface.

Meta put the technology into production use for ML inference, big data processing, databases and distributed caches, and said incidents of work stopping due to a lack of memory fell by 33%. In distributed inference, it cut the number of servers by up to 25% and is also achieving cost savings amid soaring memory prices, The Register said.