PAR2 benchmarks: create, verify and repair

The raw numbers from a set of PAR2 rounds, without commentary. Seven PAR2 tools on the same fixtures: parfast, par2cmdline-turbo, par2j (MultiPar), ParPar, rarpar, par2cmdline and phpar2. Each round below states its machine, its fixture, its command line, the build it ran, how many repetitions it is a median of and the spread behind that median.

Where a section carries one comparison rather than all seven, the tool it names is the fastest of the others at that job, not the slowest. The slower tools stay in the tables at the depths they were run to, and the rounds that were not run, or were run once, or were thrown away, are declared as such in the status list rather than left out.

Every figure is a CLI timing read out of a log. Every repair leg was checked by SHA-256 of every member against a manifest taken before the damage, never by the tool's exit code. Nothing is extrapolated and nothing is fitted: a rung with no measurement is absent rather than estimated. The rounds ran unattended under automated harnesses, one round to a machine, each leg refusing to start on a busy box.

Table rows are depths, set sizes or machines, in order, never ranked by result; the tool columns hold the same order throughout, with parfast first because the ratio columns are taken against it. The bar charts in the seven-tool section are the one thing here ordered by result, shortest first. Every other chart is drawn from the rows of the table beside it and opens in seconds; the ratio view and the log axis are buttons.

The harnesses and every log behind these figures are published at the path parfast-bench/ on this site, so each number can be traced to the line it came from.

Published · 12 September 2026 · parfast 1.5.0-beta.1 vs par2cmdline-turbo 1.5.0 vs par2j 1.3.3.5

The machines and the rounds

Intel i5-10600KF (2020)

6 cores / 12 threads, 64 GB, Windows 11 desktop

Intel Core Ultra 9 386H (2026)

16 cores, 31 GB, Windows 11 laptop

Apple M3 Ultra (2025)

32 cores, 256 GB, macOS 27 desktop

Apple M5 Max (2026)

18 cores, 128 GB, macOS 27 laptop

Intel i5-10600KF (2020) - full range

6 cores / 12 threads, 64 GB - 23 members, 23 GiB, 100% redundancy, m to 32,177

Intel Core Ultra 9 386H (2026) - full range

16 cores, 31 GB laptop - 23 members, 23 GiB, 100% redundancy, m to 32,177

Apple M5 Max (2026) - full range

18 cores, 128 GB laptop - 23 members, 23 GiB, 100% redundancy, m to 32,177

Apple M3 Ultra (2025) - full range

32 cores, 256 GB - 23 members, 23 GiB, 100% redundancy, m to 32,177, turbo at -T23

Intel create sweep, with a memory-capped arm

Core Ultra 9, 31 GB: six sizes to 40 GiB, three parity levels, a third arm under -m2048. One run per cell. In the create section, with what the -m arm did and did not do.

25% and 30% redundancy

M3 Ultra, ten cells, three interleaved repetitions. The single-run inversions are gone; parfast's wall spread at 50-80 GiB is printed.

Memory floor - deep repair under a tight -m

Landed on the 31 GB laptop. -m8192 360 s, -m2048 352 s, -m512 1,372 s, all restoring 23/23.

Fast mode, per class of machine

The shipping build on the i5 (AVX2 class) and the Core Ultra 9 (GFNI class): eleven depths, three repetitions, an A/A control; nine half-step depths on three Apple parts. On by default everywhere. In the fast-mode section.

Fast mode capability

Engages on Apple silicon and on x86 with or without GFNI. See the section below for which machines have it on by default.

All seven PAR2 tools, side by side

Repairing a lost file

Intel i5-10600KF, 6 cores / 12 threads, 64 GB, Windows 11 - 10 GiB in 10 files, 15% parity, 13,990 source blocks - medians of 2 repetitions

One member of ten deleted: 1,399 blocks to rebuild.

Building the parity set

Intel i5-10600KF, 6 cores / 12 threads, 64 GB, Windows 11 - 10 GiB in 10 files, 15% parity, 13,990 source blocks - medians of 2 repetitions

Verifying an undamaged set

Intel i5-10600KF, 6 cores / 12 threads, 64 GB, Windows 11 - 10 GiB in 10 files, 15% parity, 13,990 source blocks - medians of 2 repetitions

blocks to rebuildcount	parfastseconds	par2cmdline-turboseconds	par2cmdlineseconds	phpar2seconds	par2j (MultiPar)seconds	rarparseconds	vs turbotimes	vs par2cmdlinetimes	vs phpar2times	vs par2j (MultiPar)times	vs rarpartimes
1,399	28	113	1130	1166	176	118	4x	40x	41x	6x	4x
2,098	32	159	1636	1795	239	165	5x	51x	56x	7x	5x
4,096	44	293	3572	4177	428	387	7x	81x	95x	10x	9x

Repair from one lost block to the format's ceiling

Intel i5-10600KF

cores 6 / 12 threads · ram 63.9 GB · os Microsoft Windows 11 Pro · set 23 files, 23.0 GiB · blocks 32177 · redundancy 100.0% · slice 768000 · reps 1 of 2 completed · argv r -q -t12 -T12 f.par2 · binary cae5a6f368 (8b0b67ec4d)

First pass. parfast ran every depth to 32,177; par2cmdline-turbo and par2j had not run the last depth when the machine was needed for the fast-mode gate, so the comparison stops at 24,576 and the 32,177 row shows parfast alone. Further repetitions follow the seven-tool round on this machine.

parfastpar2cmdline-turbopar2j (MultiPar)

blocks to rebuildcount	parfastseconds	par2cmdline-turboseconds	par2j (MultiPar)seconds	vs turbotimes	vs par2j (MultiPar)times
verify only	3.77	34.8	51.9	9.2x	13.8x
1	13.7	42.5	76.7	3.1x	5.6x
4	15.3	44.7	85.8	2.9x	5.6x
16	19.3	49.3	130	2.6x	6.7x
64	35.9	71.3	169	2.0x	4.7x
256	65.7	94.6	195	1.4x	3.0x
1,024	65.8	220	334	3.3x	5.1x
2,048	70.6	390	535	5.5x	7.6x
4,096	89.1	735	955	8.3x	10.7x
8,192	115	1442	1909	12.5x	16.6x
16,384	245	2964	4208	12.1x	17.2x
24,576	660	4446	7140	6.7x	10.8x
32,177	1135	-	-	-	-

Intel Core Ultra 9 386H

cores 16 · ram 31.4 GB · os Microsoft Windows 11 Home · set 23 files, 23.0 GiB · blocks 32177 · redundancy 100.0% · slice 768000 · reps 2 · argv r -q -t16 -T16 f.par2 · binary cae5a6f368 (8b0b67ec4d)

Sole round on that box for its whole window. Two earlier attempts overlapped another round on the same machine and are void, kept on disk and not shown.

parfastpar2cmdline-turbopar2j (MultiPar)

blocks to rebuildcount	parfastseconds	par2cmdline-turboseconds	par2j (MultiPar)seconds	vs turbotimes	vs par2j (MultiPar)times
verify only	7.64±16.0%	65.5±0.5%	75.6±1.0%	8.6x	9.9x
1	18.6±9.0%	76.1	110±9.4%	4.1x	5.9x
4	20.7±18.0%	77.3±18.9%	118±17.7%	3.7x	5.7x
16	27.9±58.2%	84.0±27.4%	159±15.8%	3.0x	5.7x
64	64.0±46.2%	110±40.2%	221±12.5%	1.7x	3.5x
256	47.2±61.9%	98.9±27.3%	245±15.4%	2.1x	5.2x
1,024	55.8	132	351	2.4x	6.3x
2,048	58.0	192	507	3.3x	8.8x
4,096	67.1	311	826	4.6x	12.3x
8,192	86.0	560	1523	6.5x	17.7x
16,384	231	1061	3150	4.6x	13.6x
24,576	346	1585	5125	4.6x	14.8x
32,177	612	2096	7027	3.4x	11.5x

Apple M5 Max

cores 18 · ram 128.0 GB · os macOS 27.0 · set 23 files, 23.0 GiB · blocks 32177 · redundancy 100.0% · slice 768000 · reps 2 · argv r -q -t18 -T18 f.par2 · binary 5d0707ee75 (8b0b67ec4d)

Started at 2% foreign CPU on an otherwise idle machine; every repetition agrees to within 5% bar one rung.

parfastpar2cmdline-turbo

blocks to rebuildcount	parfastseconds	par2cmdline-turboseconds	vs turbotimes
verify only	2.51±0.3%	28.1±0.2%	11.2x
1	4.62	31.5	6.8x
4	5.55	32.3	5.8x
16	5.51	33.1	6.0x
64	8.16	36.9	4.5x
256	9.27	47.2	5.1x
1,024	10.5	88.5	8.4x
2,048	12.7±14.0%	147±6.1%	11.6x
4,096	15.9	258	16.2x
8,192	20.6	491	23.9x
16,384	35.4	978	27.6x
24,576	68.1	1479	21.7x
32,177	134	1977	14.7x

Apple M3 Ultra

cores 32 · ram 256.0 GB · os macOS 27.0 · set 23 files, 23.0 GiB · blocks 32177 · redundancy 100.0% · slice 768000 · reps 2 · argv r -q -t32 -T23 f.par2 · binary 5d0707ee75 (8b0b67ec4d)

turbo given -T23, the thread count its file list allows rather than a flat -T16. Giving it that moves the deepest rung about 5% in par2cmdline-turbo's favour, and it is the fair setting.

parfastpar2cmdline-turbo

blocks to rebuildcount	parfastseconds	par2cmdline-turboseconds	vs turbotimes
verify only	1.55±0.3%	31.0±0.4%	20.0x
1	3.35±6.8%	35.0	10.4x
4	4.36	36.6	8.4x
16	4.65±8.4%	37.6	8.1x
64	6.17	41.2	6.7x
256	6.62	46.5	7.0x
1,024	7.46	66.8	9.0x
2,048	8.48	95.8	11.3x
4,096	10.9	154	14.2x
8,192	15.4	278	18.0x
16,384	27.6	534	19.4x
24,576	42.1	806	19.1x
32,177	109	1072	9.8x

Repair at 15% parity, four machines

Ratios of medians, par2cmdline-turbo over parfast. Ten 1 GiB files of distinct random data, 750 KiB slices, 13,990 source blocks, 2,098 recovery blocks, seeded scattered damage, two repetitions per rung, the same fixture shape and damage seeds on every box. Seconds, spreads and charts for every machine are behind the toggle.

machine	verify onlytimes	1blocks	64blocks	256blocks	512blocks	1,024blocks	1,600blocks	2,098blocks	100% parity, deepesttimes
Intel i5-10600KF	2.5x	2.0x	1.7x	1.6x	2.5x	3.5x	4.5x	5.3x	6.7x at 24,576
Intel Core Ultra 9 386H	2.1x	2.3x	2.3x	1.9x	1.4x	1.6x	2.1x	2.5x	3.4x
Apple M3 Ultra	2.3x	2.9x	2.8x	2.7x	3.4x	4.4x	5.3x	6.1x	9.8x
Apple M5 Max	2.4x	3.0x	2.3x	1.8x	2.6x	4.0x	5.7x	6.5x	14.7x

Show every machine's legs: seconds, spreads and charts

Intel i5-10600KF

cores 6 / 12 threads · ram 63.9 GB · os Microsoft Windows 11 Pro · set 10 files, 10.0 GiB · blocks 13990 · redundancy 14.996% · slice 768000 · reps 3 · argv r -q -t12 -T16 f.par2 · binary cae5a6f368 (8b0b67ec4d)

Three repetitions, started only after the machine had been idle for half an hour and under 3% foreign CPU for ten minutes.

parfastpar2cmdline-turbopar2j (MultiPar)

blocks to rebuildcount	parfastseconds	par2cmdline-turboseconds	par2j (MultiPar)seconds	vs turbotimes	vs par2j (MultiPar)times
verify only	1.65±0.7%	4.12±0.9%	9.17±3.8%	2.5x	5.6x
1	3.98±6.7%	7.87±5.8%	16.7±16.6%	2.0x	4.2x
16	7.22±78.9%	12.4±32.9%	39.3±43.2%	1.7x	5.4x
64	10.1	16.8±56.9%	50.0±18.4%	1.7x	5.0x
256	16.3	26.8	65.6	1.6x	4.0x
512	17.8	43.5	86.6	2.5x	4.9x
768	19.6	60.5	108	3.1x	5.5x
1,024	22.3	77.3	128	3.5x	5.7x
1,152	23.9	85.9	139	3.6x	5.8x
1,280	23.9	94.3	149	3.9x	6.2x
1,408	24.5	103	160	4.2x	6.5x
1,600	25.4	115	176	4.5x	7.0x
1,792	26.4	128	193	4.9x	7.3x
2,048	27.7	145	215	5.3x	7.8x
2,098	28.0±5.0%	148	217	5.3x	7.8x

Intel Core Ultra 9 386H

cores 16 · ram 31.4 GB · os Microsoft Windows 11 Home · set 10 files, 10.0 GiB · blocks 13990 · redundancy 14.996% · slice 768000 · reps 2 · argv r -q -t16 -T16 f.par2 · binary cae5a6f368

Sole round on that box for its whole window.

parfastpar2cmdline-turbopar2j (MultiPar)

blocks to rebuildcount	parfastseconds	par2cmdline-turboseconds	par2j (MultiPar)seconds	vs turbotimes	vs par2j (MultiPar)times
verify only	1.69±2.6%	3.55±4.5%	10.2±1.2%	2.1x	6.0x
1	4.11±43.5%	9.56±48.3%	23.3	2.3x	5.7x
16	8.78±17.3%	17.4	46.4±18.3%	2.0x	5.3x
64	9.89	22.8±14.4%	58.8	2.3x	5.9x
256	13.4±13.0%	25.2	71.4	1.9x	5.3x
512	22.6	31.3±5.7%	89.6	1.4x	4.0x
768	23.7	33.9	106	1.4x	4.5x
1,024	24.4	40.0	123	1.6x	5.0x
1,152	25.6	42.7	132	1.7x	5.2x
1,280	25.1	46.6	141	1.9x	5.6x
1,408	25.4±11.2%	49.4	150	1.9x	5.9x
1,600	25.8	54.8	163	2.1x	6.3x
1,792	27.1	60.6±7.5%	177	2.2x	6.5x
2,048	26.0	66.0	194	2.5x	7.5x
2,098	26.8	67.0	197	2.5x	7.3x

Apple M3 Ultra

cores 32 · ram 256.0 GB · os macOS 27.0 · set 10 files, 10.0 GiB · blocks 13990 · redundancy 14.996% · slice 768000 · reps 2 · argv r -q -t32 -T16 f.par2 · binary 0f859e0c35

Sole round on that box for its whole window.

parfastpar2cmdline-turbo

blocks to rebuildcount	parfastseconds	par2cmdline-turboseconds	vs turbotimes
verify only	1.49±0.1%	3.47±2.7%	2.3x
1	2.21	6.37±5.6%	2.9x
16	3.08	9.33±24.5%	3.0x
64	3.30±5.9%	9.40	2.8x
256	4.41	11.9	2.7x
512	4.61	15.7±22.9%	3.4x
640	4.75	16.5±6.6%	3.5x
704	4.88	17.6	3.6x
768	4.99	18.7	3.8x
896	5.00	20.6	4.1x
1,024	5.14	22.7±7.3%	4.4x
1,152	5.27	23.7±6.7%	4.5x
1,280	5.30	26.0	4.9x
1,408	5.45	27.2±5.7%	5.0x
1,600	5.63	29.9	5.3x
1,792	5.67	34.1±6.0%	6.0x
2,048	5.98	37.0±8.3%	6.2x
2,098	6.04	36.7	6.1x

Apple M5 Max

cores 18 · ram 128.0 GB · os macOS 27.0 · set 10 files, 10.0 GiB · blocks 13990 · redundancy 14.996% · slice 768000 · reps 2 · argv r -q -t18 -T16 f.par2 · binary 5d0707ee75

Sole round on that box for its whole window.

parfastpar2cmdline-turbo

blocks to rebuildcount	parfastseconds	par2cmdline-turboseconds	vs turbotimes
verify only	1.26±0.2%	3.04±2.7%	2.4x
1	1.96±6.5%	5.79	3.0x
16	2.85	7.07	2.5x
64	3.59±6.1%	8.23	2.3x
256	7.20±5.1%	12.7±5.4%	1.8x
512	7.54±5.6%	19.7	2.6x
640	7.61±5.9%	21.9	2.9x
704	7.79	24.1	3.1x
768	8.01	26.8	3.3x
896	8.03	29.6	3.7x
1,024	8.35±6.1%	33.8	4.0x
1,152	8.35±5.4%	36.7	4.4x
1,280	8.60±6.5%	40.1	4.7x
1,408	8.42	43.5	5.2x
1,600	8.51	48.7	5.7x
1,792	9.01	53.5	5.9x
2,048	9.18	60.5	6.6x
2,098	9.51	61.9	6.5x

Checking a download before repairing it

Six set sizes and shapes

box 6 cores / 12 threads, 64 GB, Windows 11 · drive sustained read 0.90 GB/s · reps 2 · seconds, median

Read the last two columns before quoting any number in this table. They are the reason it is here.

set	parfastseconds	turboseconds	par2jseconds	par2cmdlineseconds	phpar2seconds	rarparseconds	vs best rivaltimes	read rateGB/s	served from
1 GiB / 1 file	1.37	1.64	1.98	1.02	5.55	2.66	0.7x	0.78	from disk
10 GiB / 10 files	1.68	4.31	9.77	8.22	52.6	15.1	2.6x	6.39	from RAM
20 GiB / 20 files	3.67±7.7%	8.78	18.0	16.5	105	30.2	2.4x	5.85	from RAM
20 GiB / 1 file	23.5	32.2	39.3	18.9	105	53.3	0.8x	0.91	from disk
40 GiB / 40 files	6.52±38.5%	27.0±5.6%	41.9	48.4	265	60.9	4.1x	6.59	from RAM
80 GiB / 80 files	110±17.6%	128±12.2%	112±13.7%	149±9.7%	576	266±11.3%	1.0x	0.78	from disk

Two cases where parfast is slower

Building the parity set, at every size

Ten set sizes, three parity levels

box M3 Ultra, 32 cores, 256 GB · sizes 10-80 GiB · redundancy 10 / 15 / 20% · slice cap 32,768

The uncapped re-run: turbo at -T32, the box's core count, after the first pass had it at -T16. The capped log is kept and not shown.

set sizeGiB	redundancy%	source blockscount	parfastseconds	turboseconds	ratiotimes	parfast peakGB of RAM	turbo peakGB of RAM
10 GiB	10%	13,990	3.67	22.8	6.2x	11.6 GB	1.0 GB
10 GiB	15%	13,990	3.96	30.0	7.6x	12.2 GB	1.5 GB
10 GiB	20%	13,990	4.06	39.0	9.6x	12.9 GB	2.0 GB
15 GiB	10%	20,985	4.84	44.0	9.1x	17.2 GB	1.5 GB
15 GiB	15%	20,985	5.22	62.5	12.0x	18.2 GB	2.3 GB
15 GiB	20%	20,985	5.55	82.9	14.9x	19.2 GB	3.0 GB
20 GiB	10%	27,980	5.75	74.4	12.9x	22.9 GB	2.0 GB
20 GiB	15%	27,980	6.15	109	17.7x	24.2 GB	3.0 GB
20 GiB	20%	27,980	6.57	143	21.8x	25.3 GB	4.0 GB
23 GiB	10%	32,177	6.27	96.2	15.3x	26.3 GB	2.3 GB
23 GiB	15%	32,177	6.83	141	20.6x	27.7 GB	3.5 GB
23 GiB	20%	32,177	6.96	188	27.0x	29.0 GB	4.6 GB
30 GiB	10%	32,760	7.70	128	16.7x	34.1 GB	3.0 GB
30 GiB	15%	32,760	8.36	187	22.3x	35.9 GB	4.5 GB
30 GiB	20%	32,760	8.73	246	28.2x	37.6 GB	6.0 GB
40 GiB	10%	32,760	10.6	171	16.1x	45.1 GB	4.0 GB
40 GiB	15%	32,760	11.4	249	21.8x	47.4 GB	6.0 GB
40 GiB	20%	32,760	19.2	328	17.1x	42.3 GB	8.1 GB
50 GiB	10%	32,750	13.9	214	15.4x	56.1 GB	5.1 GB
50 GiB	15%	32,750	14.8	310	21.0x	59.0 GB	7.6 GB
50 GiB	20%	32,750	17.6	413	23.4x	52.3 GB	10.1 GB
60 GiB	10%	32,760	16.3	253	15.5x	67.2 GB	6.1 GB
60 GiB	15%	32,760	21.5	369	17.2x	62.1 GB	9.1 GB
60 GiB	20%	32,760	31.3	490	15.7x	62.4 GB	12.1 GB
70 GiB	10%	32,760	18.8	291	15.5x	78.2 GB	7.1 GB
70 GiB	15%	32,760	23.0	433	18.8x	72.0 GB	10.6 GB
70 GiB	20%	32,760	32.2	575	17.9x	72.4 GB	14.1 GB
80 GiB	10%	32,720	23.2	333	14.3x	81.6 GB	8.1 GB
80 GiB	15%	32,720	25.4	495	19.5x	82.1 GB	12.1 GB
80 GiB	20%	32,720	36.0	660	18.4x	82.4 GB	16.1 GB

The same sweep on a 31 GB laptop, with a memory-capped arm

Intel Core Ultra 9 386H, 16 cores, 31 GB · sizes 10-40 GiB · redundancy 10 / 15 / 20% · third arm: the identical create under -m2048

set sizeGiB	redundancy%	source blockscount	parfastseconds	parfast -m2048seconds	turboseconds	ratiotimes	parfast peakGB of RAM	-m2048 peakGB of RAM	turbo peakGB of RAM
10 GiB	10%	13,990	12.2	13.6	37.5	3.1x	11.2 GB	10.4 GB	1.0 GB
10 GiB	15%	13,990	13.5	15.7	53.9	4.0x	11.8 GB	10.4 GB	1.5 GB
10 GiB	20%	13,990	13.8	15.9	71.4	5.2x	12.3 GB	10.6 GB	2.0 GB
15 GiB	10%	20,985	12.8	14.5	79.6	6.2x	16.8 GB	15.4 GB	1.5 GB
15 GiB	15%	20,985	13.4	14.8	118	8.8x	17.6 GB	15.6 GB	2.3 GB
15 GiB	20%	20,985	14.1	16.4	157	11.1x	18.4 GB	15.8 GB	3.0 GB
20 GiB	10%	27,980	15.7	17.9	139	8.9x	22.3 GB	20.6 GB	2.0 GB
20 GiB	15%	27,980	17.4	18.1	209	12.0x	23.5 GB	20.8 GB	3.0 GB
20 GiB	20%	27,980	17.9	22.7	279	15.5x	21.0 GB	21.0 GB	3.9 GB
23 GiB	10%	32,177	31.1	20.0	184	5.9x	25.1 GB	23.6 GB	2.3 GB
23 GiB	15%	32,177	60.8	22.5	284	4.7x	25.0 GB	23.9 GB	3.5 GB
23 GiB	20%	32,177	22.4	36.4	373	16.6x	24.1 GB	24.1 GB	3.9 GB
30 GiB	10%	32,760	166	116	262	1.6x	25.9 GB	25.8 GB	3.0 GB
30 GiB	15%	32,760	133	181	416	3.1x	25.7 GB	25.6 GB	3.9 GB
30 GiB	20%	32,760	140	138	494	3.5x	25.6 GB	25.5 GB	3.9 GB
40 GiB	10%	32,760	204	260	369	1.8x	25.6 GB	25.6 GB	3.9 GB
40 GiB	15%	32,760	218	243	490	2.2x	25.6 GB	25.6 GB	3.9 GB
40 GiB	20%	32,760	221	2400	654	3.0x	25.5 GB	26.0 GB	3.9 GB

What going past 20% parity costs

20% parity

Apple M3 Ultra, 32 cores · median of 3

set sizeGiB	parity%	parfastseconds	turboseconds	ratiotimes
20 GiB	20%	6.51	142.6	21.9x
23 GiB	20%	7.08	186.7	26.4x
30 GiB	20%	8.66	246.3	28.4x
40 GiB	20%	13.70	328.7	24.0x
80 GiB	20%	27.90	675.8	24.2x	±28.8%

25% and 30% parity

Same box, same fixtures · median of 3, repetitions interleaved rather than run back to back

Medians of three interleaved repetitions. parfast's wall-clock spread at 50 and 80 GiB is printed beside those rows: its CPU-seconds are steady there and only the wall moves. turbo's spread never exceeds 6%.

set sizeGiB	parity%	parfastseconds	turboseconds	ratiotimes
10 GiB	25%	4.31	47.5	11.0x	±22.4%
10 GiB	30%	4.44	56.1	12.6x
23 GiB	25%	7.18	231.4	32.2x
23 GiB	30%	7.47	283.0	37.9x
30 GiB	25%	8.97	316.8	35.3x
30 GiB	30%	11.32	380.4	33.6x
50 GiB	25%	18.98	533.5	28.1x	±35.3%
50 GiB	30%	19.91	644.6	32.4x	±16.7%
80 GiB	25%	36.68	847.0	23.1x	±18.0%
80 GiB	30%	38.83	1008.1	26.0x	±44.9%

machine	vector feature that decides it	result	how it was determined
Apple M1 / M3 / M5, every variant	NEON, always sufficient	engages	measured, M3 Ultra
Intel and AMD x86 without GFNI	AVX2 nibble kernel	engages	measured, Intel i5-10600KF
Intel and AMD x86 with GFNI	the GFNI kernel	engages	measured, Intel Core Ultra 9 386H and AMD Ryzen 9800X3D
Older x86 with SSSE3 but no AVX2 - pre-2013 Atom, Celeron, NAS boxes	a 128-bit SSSE3 kernel	engages	derived from the source; no such part was available to test
32-bit ARM - the Raspberry Pi 2 / 3 / Zero 2 W build	no vector kernel on that architecture	does not engage	derived from the source

Protection against scattered article loss, by set size

How much you can lose and still recover

arithmetic

from MAX_INPUT_SLICES = 32,768 · article 768,000 bytes · knee 23.44 GiB

20% parity10% parity30% parity

How much parity you need to stay just as safe

arithmetic

same two constants · read as: parity = target x set size / 23.44 GiB

to hold 20%to hold 10%

set sizeGiB	slice sizeMB	source blockscount	articles you can lose at 20% parity% of articles	parity needed to hold 20%% of payload
5.00 GiB	0.73 MB	6,991	20.00%	20.0%
10.00 GiB	0.73 MB	13,982	20.00%	20.0%
15.00 GiB	0.73 MB	20,972	20.00%	20.0%
20.00 GiB	0.73 MB	27,963	20.00%	20.0%
23.44 GiB	0.73 MB	32,768	20.00%	20.0%
30.00 GiB	0.94 MB	32,768	15.62%	25.6%
50.00 GiB	1.56 MB	32,768	9.38%	42.7%
75.00 GiB	2.34 MB	32,768	6.25%	64.0%
100.00 GiB	3.12 MB	32,768	4.69%	85.3%
150.00 GiB	4.69 MB	32,768	3.12%	128%
200.00 GiB	6.25 MB	32,768	2.34%	171%
300.00 GiB	9.38 MB	32,768	1.56%	256%
500.00 GiB	15.62 MB	32,768	0.94%	427%
750.00 GiB	23.44 MB	32,768	0.62%	640%
1 TiB	31.25 MB	32,768	0.47%	853%

PAR2 benchmarks: create, verify and repair

The machines and the rounds

All seven PAR2 tools, side by side

Repairing a lost file

Building the parity set

Verifying an undamaged set

Repair from one lost block to the format's ceiling

Intel i5-10600KF

Intel Core Ultra 9 386H

Apple M5 Max

Apple M3 Ultra

Repair at 15% parity, four machines

Intel i5-10600KF

Intel Core Ultra 9 386H

Apple M3 Ultra

Apple M5 Max

Checking a download before repairing it

Six set sizes and shapes

Two cases where parfast is slower

Building the parity set, at every size

Ten set sizes, three parity levels

The same sweep on a 31 GB laptop, with a memory-capped arm

What going past 20% parity costs

20% parity

25% and 30% parity

Fast mode

Which machines it works on at all

Protection against scattered article loss, by set size

How much you can lose and still recover

How much parity you need to stay just as safe

How these were measured

Every repair is checked, never trusted

Every tool got the same command, and its best settings

Medians, and the spread behind them

How much the repetitions disagree

What is not here