I want you to use it on the PC side too.

For those who assemble the computer, they know that each generation can't be used in contrast, the DDR4 motherboard can't fit DDR5, and the DDR5 motherboard can't plug in DDR4, but most recently, Meta successfully found a way to "share" even if it's not directly plugged in the same slot.

.

Given the constant rise in global DRAM prices, even Meta-level companies began to look for cost-saving ways; instead of dropping DDR4 ramps from old servers that still worked, the company developed a new ASIC chip called Vistara to help new DDR5-based servers reuse DDR4 ramps through CXL or Compute Express Link technology.

.

Meta revealed this technology at ISCA 2026, explaining that data centers typically have a 3-5 year cycle before being retired or switched to a new platform. In fact, RAM modules often last much longer than that. Some cases can last around 7-10 years, causing many DDR4 rams that still perform well to be discarded with the old ones simply because they are not compatible with newer DDR5 platforms.

.

The centerpiece of this system is the Vistara chip, which acts as an intermediary, connects through PCIe Gen5 x16 and uses the CXL 2.0/1 .1 standard to allow the newer CPUs to access the DDR4 memory installed in a separate machine. Meta calls this system MemServer. In other words, it does not directly plug DDR4 into the DDR5 motherboard, but to install the old DDR4 ram into the auxiliary memory and then have the main server pull it through CXL.

.

In the Meta test, it uses 768GB of DDR5 memory as the main ram and adds 256GB of recycled DDR4 through the CXL system. These DDR4 can be seen as additional Memory Node. When the DDR5 area does not start enough, it automatically moves some data that is not frequently triggered on DDR4 through Transparent Page Placement or TPP software to reduce the impact of higher latency than DDR5.

.

Although DDR4 is slower than DDR5 in both bandwidth and Latency terms, for certain types of tasks such as AI Inference or Distributed Cache, having more RAM may be more important than using the fastest RAM alone. Meta indicates that internal testing can reduce the number of servers required for AI Inference by up to 25% and reduce the Latency of Distributed Cache by an average of around 29%.

.

As well as reducing costs during expensive ramps, Meta also sees Vistara as reducing electronic waste from discarding still-functional RAM modules, with the expectation of saving tens of millions of dollars in hardware costs, and in the future this idea could be capped up to the DDR6 era where both DDR5 and DDR4 could be reused if the same hardware is supported.

.

Source: HKEPC

# IT should know # IT News # Includes IT matters # IT