Over the past year, memory prices have skyrocketed. DDR5 memory prices have generally increased more than fivefold compared to this time last year, with some models seeing a tenfold increase. Even AI giants find the prices outrageously high.
If AI servers were fully equipped with high-spec DDR5/LPDDR5X memory, supply would fall short and costs would spiral out of control. Hence, the current trend is to repurpose "obsolete" DDR4 memory—Previously discarded DDR4 memory from servers is being repurposed for AI servers using new technology.
Not long ago, we reported that Meta developed Vistara technology, redeploying previously upgraded and discarded DDR4 memory onto current AI server platforms.
The technology they use involves connecting DDR4 to platforms that previously required native DDR5 bus support via CXL controller chips. This allows millions of retired DDR4 modules to be reused instead of being sold as scrap.
This approach offers many benefits. Servers typically have a lifespan of 4-5 years, while DDR4 memory can last 10-12 years, allowing for extended reuse after redeployment.Significantly reduces costs, cuts down on the demand for DDR5 memory, and helps curb the rampant price hikes of DDR5.
There are downsides, too. DDR4's bandwidth and latency certainly can't match DDR5, impacting performance, but not all AI applications require top-tier performance. Many workloads benefit more from large-capacity memory, so expanding DDR4 memory capacity by 2-2.5 times via the CXL interface is quite helpful.
