Meta is developing a hardware‑offloaded compressed‑RAM system (CRAM) that presents compressed storage as normal RAM to the kernel, allowing cacheline/byte access and dramatically better performance than software solutions like ZRAM or Zswap. Benchmarks in the presentation claim near‑native DRAM speeds for reads and large speedups even when writes occur, because compression/decompression are handled by dedicated hardware and the memory is exposed via a private NUMA node.
— If adopted widely, CRAM could lower per‑workload DRAM needs, shift data‑center power/space economics, reduce pressure on DRAM markets, and change how constrained devices (phones, handhelds, gaming PCs) and servers manage memory.
EditorDavid
2026.10.10
100% relevant
Meta engineer Gregory Price presented CRAM at the Linux Plumbers Conference; the article cites measured speed comparisons versus ZRAM/Zswap (e.g., 452× read‑only case, 5.4× worst‑case for writes) and notes many kernel pieces are already mainline.
← Back to all ideas