A new Linux feature called CRAM claims read speeds up to 445 times faster than the kernel's existing ZRAM compression.
Meta engineer Gregory Price and his team built CRAM and presented it at the Linux Plumbers Conference in Prague. Unlike zswap and ZRAM, which treat compressed memory as a swap-layer trick, CRAM keeps compressed data in RAM itself, managed through a private NUMA node so the kernel can still use normal memory tools like migration and ballooning. In the worst-case read benchmark shown in Price's slides, CRAM hit 489 million operations per second versus ZRAM's 1.1 million, a gap of roughly 445x. Enable writes and the advantage drops to 5.4x in the worst tested case, because writing to compressed data means first migrating it out of its compressed state.
That write penalty is the real story here, not the headline number. ZRAM and zswap are everywhere, from Meta's server fleets down to the Steam Deck, and most distributions turn one on by default. A compression layer that's merely 5x faster on mixed workloads, instead of 445x on cherry-picked reads, is still the more useful number for anyone deciding whether this is worth shipping.
CRAM also hasn't solved a harder problem: knowing how much usable memory you actually have when some of it compresses better than the rest. Price's team added a safety valve, nicknamed the "Chicken Bit," to stop the system from cascading into what the slides call a "poison storm" when allocations can't keep up. Until that gets sorted, CRAM is a promising conference talk, not a kernel patch.