Digital compute-in-memory chip claims 1.27 fJ/B/transition for NFA pattern matching

What happened

Researchers have unveiled a digital compute-in-memory (CIM) accelerator for evaluating non-deterministic finite automata (NFA), aimed squarely at high-speed pattern matching used in network intrusion detection. The work, published in the GLSVLSI ’25 proceedings, was fabricated in 22 nm FD‑SOI and, the authors report, delivers a peak throughput of 2,822 MB/s and a minimum-energy operating point that achieves 406 MB/s at 1.27 fJ per byte/transition. It has been reported that the team evaluated the design against real IDS workloads such as SNORT.

How it works

The design leans on digital CIM macros to perform state transitions inside memory arrays, reducing data movement — the usual energy tax in these pipelines. A clever twist: a CIM Bloom filter gates macro activity to enable opportunistic symbol skipping, lowering pointless computation when symbols can be ignored. In short: compute where the bits live, and skip what you don’t need. Bold, but not magic — it’s engineering tradeoffs and careful microarchitectural pruning.

Performance and validation

The paper lays out silicon measurements rather than just simulations, which matters. The reported 1.27 fJ/B/transition figure is striking; micro‑architects will raise an eyebrow. Throughput and energy numbers vary with operating point, but both speed and efficiency improved enough to make a practical case for on‑chip NFA acceleration. The authors used SNORT workloads for evaluation, which keeps the claims grounded in a known, real‑world application.

Why this matters

Pattern matching is a workhorse for security, monitoring, and even some text-processing tasks. So what? Reduced energy per transition changes the calculus for deploying specialized accelerators in edge boxes, smart NICs, or datacenter middleboxes. This fits a broader trend — pushing compute closer to memory to escape the von Neumann bottleneck. Will operators rip out general‑purpose CPUs and bolt in these CIM chips tomorrow? Probably not. But for anyone wrestling with power‑limited, always‑on inspection, this is the kind of efficiency that makes you sit up and take notice.

Sources: acm.org, Hacker News