Meta Reportedly Unveils CRAM Memory Compression Solution, with Read-only Performance Approaching Native DRAM

CRAM Keeps Compressed Data in Memory Form, Bypassing the Swap Path

At the Linux Plumbers Conference 2026 (LPC 2026) recently held in Prague, Czech Republic, Meta has presented an experimental memory compression solution called CRAM (Compressed RAM). The goal of the solution is to keep compressed data in memory form, thereby reducing the software overhead associated with the traditional Swap path, rather than simply replacing ZRAM or zswap.

Currently, Linux systems generally use ZRAM and zswap to achieve memory capacity expansion, both of which rely on swap-layer logic. CRAM adopts a different design, in which compressed data can still maintain page table mappings and page cache state. It also supports Cacheline- and Byte-level access, so reading data does not require triggering software decompression through page faults as traditional solution does. This characteristic has been listed as the key difference between CRAM and ZRAM/zswap.

Reusing Existing Memory Management Mechanisms, CRAM's Read-only Performance Approaches Native DRAM

The core idea of CRAM is to provide hardware-compressed memory to Linux as a special type of NUMA memory, instead of emulating it as a block storage device. Its implementation relies on native Linux kernel mechanisms to build a dedicated private NUMA node that behaves like a virtualized shadow compute unit. This design allows the kernel to avoid recognizing it as a storage device, thereby seamlessly reusing all existing memory management functions, including page migration, memory ballooning, and other mechanisms.

This design targets the software path overhead in traditional memory compression solution. For read-only data, CRAM can read directly from compressed memory without first swapping the data into ordinary memory and then decompressing it. In this mode, CRAM's performance can approach native DRAM levels. It is reported that CRAM has completed actual testing and can achieve performance close to native DRAM in TAOBench and FIO benchmarks.

Read-only Performance Reaches over 400 Times that of ZRAM, Write Advantage Narrows but Remains Superior to ZRAM

Test data disclosed in the presentation shows that in read-only scenarios, CRAM can achieve approximately 489 million operations per second in the worst case, while ZRAM achieves about 1.1 million operations per second. Based on this set of data, the gap between the two is about 444 times, not 452 times; the official claim of a 452-fold improvement may require an ideal environment to be achieved.

However, when the workload includes write operations, CRAM's advantage narrows significantly. Even so, under a harsh scenario with an added 20% write load, folios must be migrated back to the original NUMA node through page faults, because compressed pages cannot be overwritten in place. CRAM still maintains about 5.4 times the operating efficiency of ZRAM. It should be noted that this figure should be understood as a specific benchmark result rather than a general performance conclusion.

It is reported that CRAM is currently still a tested kernel service prototype, not yet a mature feature that has entered the Linux mainline. Meta's focus in this public disclosure is to explore how to use existing Linux memory management mechanisms to support this new type of compressed memory, rather than to announce a product that has already completed standardization or commercial deployment.