Interesting detail on the algorithm but seems to completely miss that if you care about non-streaming performance, there are parallel versions of xz and gzip (pxzip encodes compatible metadata about the breakup points so that while xz can still decompress it, pxzip can use as many cores as you let it have instead.) Great for disk-image OS installers (the reason I was benchmarking it in the first place - but this was about 5 years back, I don't know if those have gotten upstreamed...)