Skip to content

Switch.md

Latest commit

 

History

History
37 lines (30 loc) · 1.54 KB

Switch.md

File metadata and controls

37 lines (30 loc) · 1.54 KB
Raw

Switch NSZ benchmark

rom-converto versus nsz (https://github.com/nicoboss/nsz) on NSP and XCI inputs. Both tools produce zstd-compressed .ncz files inside a PFS0 (NSZ) or HFS0 (XCZ) outer container. Full uninterpreted tables, hosts, and methodology:

Summary

Delta is the rom-converto / nsz ratio; below 1.00x means rom-converto is faster.

Operation Windows (Ryzen 9 5900X) macOS (Apple M4)
NSP compress (L18 solid) 0.58x (1.71x faster) 0.72x (1.40x faster)
NSP decompress (large input) 0.47x (2.12x faster) 0.74x (1.35x faster)
NSP decompress (308 MB) 0.35x (2.85x faster) 1.07x
XCI compress (L18 block) 0.59x (1.69x faster) 0.80x (1.25x faster)
XCI decompress (block) 0.30x (3.39x faster) 0.93x (1.07x faster)

Interpretation

  • NSP compress wins on both hosts and produces 16.6 % smaller output on the shared 308 MB input, from EnableLongDistanceMatching = true being on by default (nsz hides the same option behind an off-by-default -L flag).
  • Decompress gains grow with core count: the 24-thread desktop reaches 2.1x to 3.4x where the 10-core M4 sits at 1.0x to 1.35x. Block-mode XCZ decompress parallelises per block and benefits the most.
  • The cross-tool round trip (rom-converto compress, then nsz -D, then SHA-256 against the source) passed for every input on both hosts. XCI compressed sizes match nsz's to within framing noise (-1,604 B on Windows, -990 B on macOS).