# Meta reuses DDR4 RAM via CXL ASIC to boost AI server capacity

**Published:** 2026-07-04T14:50:46.964Z  
**Topic:** Meta%5C  
**Sentiment:** neutral  
**Publisher:** TrendWatcher — https://www.trendwatcher.in/article/d18e5230-e64f-41a6-9ac3-af7fb3891931

Meta adds 256 GB DDR4 to 1 TB AI servers using Vistara CXL ASIC, cutting server count 25% amid DDR5 shortages.

Meta’s new “MemServer” combines 768 GB DDR5‑6400 with 256 GB DDR4‑2400, reaching a terabyte of memory per box by bridging the older RAM through a custom CXL 2.0 ASIC called Vistara【1】. The move lets Meta sidestep the global DDR5 shortage and reduces the number of AI inference servers needed by up to 25%【3】.

| At a glance | |
|---|---|
| Server memory mix | 768 GB DDR5 + 256 GB DDR4 |
| Total capacity | 1 TB per MemServer |
| DDR5 bandwidth | 614 GB/s |
| DDR4 bandwidth | 76 GB/s |
| Server count reduction | up to 25% |

## Architecture and performance trade‑offs  
Meta’s MemServers run on AMD’s Epyc “Turin” CPUs, which officially support DDR5 only. Vistara’s ASIC links two 72‑bit DDR4 channels to the host via a PCIe Gen5 x16 CXL 2.0/1.1 interface, supporting up to 256 GB per chip and 64 GB per DIMM, though the current design uses 32 GB DIMMs—the largest capacity available for reuse【1】. This configuration delivers a local DDR5 peak bandwidth of 614 GB/s, while the DDR4 tier provides just 76 GB/s and roughly double the idle latency, meaning the slower tier contributes only about one‑tenth the performance of the DDR5 pool【1】.

Meta argues that AI workloads often leave large portions of memory idle, so only a small fraction of pages need fast access. By treating the DDR4 pool as a separate NUMA node, the system keeps hot data in DDR5 and relegates cold pages to DDR4, limiting the impact of the lower bandwidth and higher latency【3】. The company reports that this “disaggregated” approach cuts AI inference server counts by up to 25% and reduces job‑restart and fragmentation overhead by 33%【3】.

## Market context and competitive implications  
The DDR5 shortage and rising DRAM prices have forced hyperscalers to explore unconventional solutions. Meta’s recycling of retired DDR4 modules avoids the “RAM tax” of buying new memory and reduces electronic waste, a claim the company highlights as near‑zero‑cost expansion【2】. While commercial CXL products typically bundle controllers with fresh DRAM, Vistara’s design enables reuse of existing DDR4 inventories at scale, a capability that could appeal to other large cloud providers facing similar supply constraints【2】.

Other industry players are also experimenting with memory disaggregation. Nvidia’s NVLink and the emerging Ultra Accelerator Link (UAL) consortium, which includes AMD, AWS, Google, Microsoft, and Meta, aim to connect accelerators across hardware vendors, suggesting a broader shift toward flexible interconnects that can accommodate heterogeneous memory pools【2】.

## What to watch
- **Vistara rollout timeline** – Meta has not disclosed a production schedule; monitoring announcements will reveal how quickly the design moves from prototype to data‑center deployment.  
- **Adoption by other hyperscalers** – Competitors’ responses, such as similar CXL‑based memory reuse strategies, could signal whether the approach becomes a wider industry practice.  
- **DDR5 supply trends** – Changes in DDR5 availability and pricing will affect the economic calculus of mixing DDR4 and DDR5 in future server designs.

Meta’s hybrid memory architecture shows that, even for a company with deep pockets, the scarcity of next‑gen DRAM can drive innovative hardware‑software co‑design. Whether the performance trade‑offs remain acceptable outside hyperscale AI workloads will determine if recycled DDR4 via CXL becomes a mainstream solution.

## Sources
1. PC Magazine — [Even Meta Is Using DDR4 to Get Around DDR5 Memory Shortages](https://ca.pcmag.com/ai/16664/even-meta-is-using-ddr4-to-get-around-ddr5-memory-shortages)
2. TechRadar — [DDR4 memory gets a second life as Meta fights soaring server RAM prices](https://www.techradar.com/pro/near-zero-cost-memory-expansion-through-recycling-meta-will-reuse-terabytes-worth-of-ddr4-memory-using-cxl-tech-and-avoid-paying-the-ram-tax)
3. TechSpot — [Meta is using old DDR4 memory in DDR5-only AI servers to save on hardware costs](https://www.techspot.com/news/112977-meta-using-old-ddr4-memory-ddr5-only-ai.html)

---
Cite as: TrendWatcher, "Meta reuses DDR4 RAM via CXL ASIC to boost AI server capacity", https://www.trendwatcher.in/article/d18e5230-e64f-41a6-9ac3-af7fb3891931