AI-rewritten: This is a summary of an article from Tom’s Hardware, rewritten by AI (Qwen, running locally) to make it easier to read. The facts come from the original article – read it for the full story.

---

Gregory Price and his team at Meta have developed CRAM, a new memory compression model designed to avoid traditional swap layers. Unlike existing Linux options like ZRAM and Zswap, which function as swap-layer features, CRAM keeps compressed data entirely within RAM. The developers state that the primary performance bottleneck in current systems is not the compression process itself, but rather the fault handling and swap behavior associated with them.

CRAM utilizes a private NUMA node, described as essentially a "ghost CPU," to manage memory without pretending to be a block device. This approach allows Linux to use standard memory semantics for tasks like migration and ballooning. A key mechanism called the "Chicken Bit" prevents cascading failures, or what the presentation terms a "poison storm," when write operations outpace allocation capabilities.

Performance benchmarks show CRAM running at DRAM speeds during read-only access, achieving up to 452 times the speed of ZRAM on a logarithmic scale. In worst-case scenarios involving reads, CRAM performs roughly 489 million operations per second compared to ZRAM’s 1.1 million. Even with write operations enabled, CRAM remains significantly faster than ZRAM, offering a 5.4x speedup in cases where 20% of data is written. However, write performance drops sharply because the system must page fault and migrate folios back to the original NUMA domain to avoid corrupting data.