Benchmarks

Clause compared with Erlang/OTP on one machine: the same Erlang programs, compiled by erlc and by clau -O2, time their kernels after a warm-up call. These numbers change with every optimization.

Geometric mean of kernel times, Clause -O2 against the BEAM JIT (OTP): 1.49× faster on average
Kernels where Clause is faster: 31 / 31
Geometric mean on 2026-10-12, the first recorded run: 20.2×

Kernel times

Median of 3 runs in milliseconds. The bar shows Clause's time divided by OTP's on a logarithmic scale.

Clause fasterOTP faster🔺 Runs varied by more than half of the median.

Kernel	OTP JIT	Clause -O2	Clause ÷ OTP
Data generation
bigint_factorialComputes the factorial of 3,000 with big integers, 330 times	1,620.17	212.76	0.13×
binary_appendAppends 10,000 32-bit integers to a growing binary, 23,000 times	1,080.93	802.57	0.74×
list_of_tuplesBuilds a list of 200,000 small tuples, 260 times	1,467.60	1,128.44	0.77×
list_to_tupleConverts 1,000-element lists to tuples, 170,000 times	1,586.38	1,278.69	0.81×
map_putInserts 2,000 integer keys into a map one by one, 4,900 times	1,605.84	1,519.51	0.95×
string_buildBuilds 20,000 strings from integers and atoms, 1,000 times	1,044.38	720.73	0.69×
Data processing
binary_matchSums 20,000 32-bit integers matched out of a binary, 49,000 times	1,272.63	1,071.47	0.84×
comprehensionList comprehension filtering and doubling 20,000 integers, 15,000 times	1,407.08	893.85	0.64×
int_to_listConverts 20,000 integers to strings and back, 2,000 times	908.70	412.79	0.45×
map_foldlMaps, then folds a list of 20,000 integers, 4,500 times	1,500.77	1,113.04	0.74×
map_updateIncrements every value of a 500-key map, 24,000 times	1,151.28	669.97	0.58×
small_mapUpdates a small state map 80 million times	1,486.44	1,061.83	0.71×
sortSorts a list of 20,000 random integers, 600 times	1,112.37	466.55	0.42×
term_to_binaryEncodes and decodes a large nested term, 4,300 times	1,433.29	1,071.95	0.75×
File I/O
large_readReads a 16 MB file in 64 KB chunks, 370 times	1,444.15	1,007.93	0.70×
large_writeWrites a 16 MB file in 64 KB chunks, 200 times	1,019.80	989.09	0.97×
read_fileReads a whole 16 MB file in one call, 330 times	1,485.52	1,483.03	1.00×
read_lineReads a 20,000-line text file line by line, 52 times	1,522.07	1,467.28	0.96×
small_filesWrites, reads back and deletes 500 small files, 11 times	1,209.75	1,116.40	0.92×
Garbage collection
explicit_collectForces garbage collection with 10,000 live tuples, 5,100 times	1,547.98	639.85	0.41×
grow_shrinkGrows the heap with a 200,000-element list, then drops it, 300 times	1,041.41	861.75	0.83×
long_lived_churnKeeps a sliding window of 10,000 live elements among garbage, 370 times	1,351.27	1,095.82	0.81×
short_processesSpawns 5,000 short-lived processes that each make garbage, 69 times	1,667.28	1,037.98	0.62×
Libraries
qlcRuns the qlc transform over 27 queries and 3 modules, 11 times	1,873.72	1,714.66	0.92×
Message passing
fan_in100 processes each send 1,000 messages to one, 25 times	1,402.68	391.78	0.28×
large_messageSends a 100,000-element list to another process and back, 560 times	1,063.94	789.04	0.74×
ping_pongBounces one message between two processes 2 million times	1,504.97	904.20	0.60×
ringPasses a token around a 1,000-process ring 3,050 times	1,438.21	1,099.66	0.76×
Loopback sockets
tcp_connectOpens 900 TCP connections over loopback, echoing 25 messages of 1 KB on each	1,418.85	1,047.85	0.74×
tcp_echoEchoes 30,000 messages of 1 KB over loopback TCP	1,455.62	1,136.06	0.78×
udp_round_tripBounces a 1 KB datagram 27,900 times over loopback UDP	1,458.07	1,196.82	0.82×

Whole programs

Wall time in seconds, startup included, median of the runs.

Program	OTP JIT	Clause -O2	Clause ÷ OTP
bench_datagenBuilds lists, tuples, maps, binaries and big integers in bulk	8.54	5.72	0.67×
bench_dataprocSorting, folds, map updates, binary matching and conversions	10.45	6.76	0.65×
bench_fileSmall and large file writes and reads, line reads	6.95	6.24	0.90×
bench_gcShort-lived processes, heap growth and shrinking, live data churn	5.83	3.69	0.63×
bench_libraryThe qlc parse transform at run time, as qlc:q/1 uses it	2.30	1.89	0.82×
bench_msgPing-pong, fan-in, a process ring and large message copies	5.69	3.28	0.58×
bench_socketTCP echo, TCP connects and UDP round trips over loopback	5.99	4.57	0.76×

Progress

Geometric mean of Clause's time divided by OTP's on each recorded run day. Runs before 2026-10-21 were not pinned to a CPU set, so part of the change on that day comes from the pinning.

Date	Commit	Clause ÷ OTP
2026-10-12	a7bc0c9b	20.2×
2026-10-13	8a244d9c	16.9×
2026-10-14	55f42a9d	16.9×
2026-10-15	6ed3851c	2.73×
2026-10-16	97295c84	1.95×
2026-10-17	a8c3f8f6	1.38×
2026-10-18	336caea7	1.41×
2026-10-20	7d7a2f2e	1.00×
2026-10-21	71ec52f8	0.78×
2026-10-23	dea68508	0.69×
2026-10-24	af105043	0.65×
2026-10-26	d6eafa72	0.71×
2026-10-30	f4e5f3e0	0.67×

Executable sizes

Size of the benchmark programs as Clause builds them at each optimization level, in megabytes.

Program	-O0	-O2	-O2 --lto
bench_datagen	9.90 MB	9.82 MB	9.82 MB
bench_dataproc	10.42 MB	10.01 MB	10.01 MB
bench_file	11.34 MB	10.59 MB	10.59 MB
bench_gc	9.89 MB	9.81 MB	9.81 MB
bench_library	10.24 MB	9.94 MB	9.95 MB
bench_msg	9.91 MB	9.82 MB	9.82 MB
bench_socket	11.58 MB	10.64 MB	10.65 MB

-O2 sizes of the last 6 benchmark runs, oldest first, one Clause commit each. The table below adds the total of all bench_* programs and the size of the clau compiler.

2026-10-207d7a2f2e
2026-10-2171ec52f8
2026-10-23dea68508
2026-10-24af105043
2026-10-26d6eafa72
2026-10-30f4e5f3e0

Date	Commit	All bench_* programs	clau
2026-10-20	7d7a2f2e	38.59 MB	59.50 MB
2026-10-21	71ec52f8	38.78 MB	59.54 MB
2026-10-23	dea68508	38.79 MB	59.54 MB
2026-10-24	af105043	38.81 MB	59.65 MB
2026-10-26	d6eafa72	62.21 MB	59.98 MB
2026-10-30	f4e5f3e0	70.63 MB	60.75 MB

Setup

Machine: Intel64 Family 6 Model 183 Stepping 1, GenuineIntel, 32 logical CPUs, Windows 11 (10.0.26200)
CPU set: Performance cores 0-15, 16 of 32 logical CPUs
Erlang/OTP: 29, ERTS 17.1
Clause commit: f4e5f3e0
Run: 2026-10-30

OTP and Clause run on the same CPUs: every benchmark program is pinned to the performance cores of the processor, so on a hybrid CPU no kernel lands on a slower efficiency core. Both keep one scheduler online per CPU of that set, unbound within it.

To run the benchmarks yourself, use make benchmark or make-benchmark.bat in the Clause repository; the benchmark notes describe the programs and options. Benchmark notes