Meta is reusing old DDR4 RAM in its servers instead of buying new hardware

Meta developed Vistara, a custom chip that allows it to use old DDR4 RAM from obsolete servers in new servers that rely on DDR5.

The global hardware shortage isn’t exactly news, as the entire world has been struggling with rising component prices for quite some time now. And while big companies certainly aren’t as affected as the average consumer, even they aren’t opposed to the idea of saving a few (million) bucks.

Meta appears to have found a way to spend less on new hardware while also putting its outdated infrastructure to use, essentially killing two birds with one stone. The company has built a custom chip that lets it reuse memory from retired servers rather than buying new hardware. The chip is called Vistara and allows for connecting old DDR4 RAM from obsolete servers into new servers that rely on DDR5.

The problem Vistara solves goes back to a basic mismatch in how long hardware lasts. Meta replaces its servers every three to five years, but the memory modules inside them are good for seven to ten. When a server gets decommissioned, perfectly usable DDR4 RAM goes with it.

Meta is presenting the new method at today’s ISCA symposium, but The Register has got hold of a paper that explains how Vistara works.

It's a custom ASIC that bridges DDR4 memory to newer processors via aCXL 2.0/1.1 interface over PCIe Gen5 x16. Meta pulls DDR4 sticks from old machines and installs them in dedicated units it calls MemServers, each of which pairs 768GB of DDR5 with 256GB of recovered DDR4. The operating system sees the DDR4 as an additional memory node and draws from it when the primary DDR5 is running low.

Off-the-shelf CXL hardware couldn't do this, so Meta built its own. Existing interfaces bundle their own memory with the controller, which makes reusing old RAM sticks impossible. But Vistara separates the controller from the memory entirely, so Meta can plug in whatever DDR4 sticks it has on hand.

Meta plans to deploy the new architecture in hyperscale infrastructure with millions of servers, which should mean that Meta’s AI datacenters will now be more efficient. The company is investing heavily in AI infrastructure, especially with its new AI model, Muse Spark, now widely available.

All of this doesn't mean that Meta will exclusively rely on "recycled" RAM, but the company is still looking at considerable savings at scale.