Meta kernel engineer Gregory Price has presented CRAM, a Linux system designed to let compressed memory work more like ordinary RAM. If its reported benchmark results hold up, it could help computers use CXL memory expansion without paying the usual speed penalty of software decompression and page faults.
Watch Desk analysis
What happened
CRAM is designed for CXL memory expanders with hardware compression. HardwareBusters reports that compressed pages can remain mapped and readable at near-DRAM speed, without triggering read-access faults or requiring the CPU to decompress data. That differs from software-based approaches such as ZRAM and Zswap.
The report says CRAM uses read-only mappings to handle data that does not compress, alongside Linux memory-tiering features and a private NUMA node state. It also reports near-native DRAM performance in TAOBench and FIO benchmarks. Read the HardwareBusters report.
Why it matters
Memory capacity is increasingly important for workloads that need to keep large amounts of data close to the processor. CRAM’s reported approach aims to make extra, compressed memory more useful without asking the CPU to do the decompression work each time data is read. That could matter to operators expanding memory capacity, though the supplied report does not give benchmark figures or enough test detail to judge how broadly the results apply.
Our read
The interesting idea is not simply “compress memory”, which Linux already does in other ways. It is keeping compressed data accessible without the familiar fault-and-decompress detour. That is a promising direction, not yet a reason to rewrite anyone’s memory plan: the practical test is how CRAM performs across workloads and hardware beyond the reported benchmarks.
What to watch
- Whether CRAM becomes available in the Linux kernel, and on what timeline.
- Which CXL hardware and compression implementations it supports.
- Benchmark results and comparisons across a wider range of workloads.
- How it handles data that does not compress well.
Discussion spark: If CRAM’s reported performance holds across more workloads, would you trust compressed CXL memory for production systems, or wait for broader testing first?
Sources and evidence
- Meta’s CRAM Wants Compressed RAM to Behave Like Real Memory, Not Swap (6 October 2026, 12:14 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.