Twenty-four blocks
doing forty-one blocks of work

02The argument

We did not give you
more memory.

The NX5 has exactly as many physical blocks as the NX4 did. We could have added four and charged for them. Instead we spent the generation on the fold table.

Block folding finds pages that are byte-identical across two resident warez and keeps one copy. Most working sets are more repetitive than their authors think: shared library pages, shared constant tables, three Barrier probes that all carry the same handshake. On a mixed set of nineteen residents the NX5/X folds 24 blocks into 41.3 blocks of usable residency.

The eviction predictor does the other half. It keeps ninety seconds of history on which residents you re-enter, holds those, and pages out the ones you loaded and never touched.

Past twenty-two concurrent residents the fold table thrashes and the Mk V’s flatter design pulls ahead. The crossover is charted on the benchmarks page with the figures under it.

Bench 2 · NITROBENCH 4 harness24 °C ± 0.5, still air30-minute run, 12 runs per figure

Sustained figures only

Peak burst on the NX5/X is 1,760 MB/s and lasts under two seconds. Every other throughput figure on this site is a thirty-minute unbroken run at 24 °C with the case closed, averaged over the last twenty-five minutes, so throttling is inside the number. The /X does throttle, and the 1,180 includes it.

Hotfloor tells you why

nitro fold --explain prints the decision for every page in your resident set: folded, refused, and the reason it was refused.

The SDK is free. It ships the benchmark harness we publish our own figures with, and a block-budget linter that fails your build when a set will not fit the target deck.