Meta's Vistara Chip Cuts AI Infrastructure Costs by Repurposing Old Memory

Meta has unveiled a custom chip called Vistara to solve a major cost problem in data center expansion. The purpose-built ASIC uses the CXL 2.0 standard to let older DDR4 memory work inside servers designed for newer DDR5.

The strategy turns decommissioned DDR4 modules into valuable infrastructure. Companies like Meta are left with mountains of this functional memory after upgrading to DDR5-only platforms.

Vistara bridges the gap. Each chip supports two DDR4 channels and is integrated into Meta's new MemServer platform. A unit pairs a powerful AMD processor with native fast DDR5 memory and an additional, slower pool of CXL-attached DDR4.

This creates a tiered memory system. The operating system treats the slower DDR4 as a separate data pool, ideal for workloads like database caching or certain AI inference tasks that need large capacity but not maximum speed on every byte.

The result is a significant reduction in both the cost and waste of building out AI-capable infrastructure.