Outputs
What the benchmarks print and save, and the page bmb_report makes of it.
In the terminal
A benchmark prints where the result came from, then one row per point as it is measured: here dgemm over three sizes, at one and two threads. These runs were made when this page was built, by the machine that built it.
$ bmb_dgemm -m 256:1024 -t 1,2 # blas-microbenchmark 1.5.0-dev # backend: openblas (OpenBLAS 0.3.26 NO_LAPACKE DYNAMIC_ARCH NO_AFFINITY Cooperlake MAX_THREADS=64) # verified: every point, against a reference, before it was timed # layout: row-major # cpu: AMD EPYC 9V74 80-Core Processor (4 logical CPUs, 1 NUMA node) # caches: L1d 32K, L1i 32K, L2 1M, L3 32M # os: Linux 6.17.0-1022-azure x86_64 # date: 2026-10-12T09:48:01Z # routine: dgemm Thread count Matrix dim1 (M=K) Matrix dim2 (N) time [s] GFLOP/s 1 256 256 0.000658143 50.983 1 512 512 0.005228642 51.339 1 1024 1024 0.040134619 53.507 2 256 256 0.000352101 95.298 2 512 512 0.002579224 104.076 2 1024 1024 0.020220410 106.204
CSV
With -o, the same results go to a file as well; -s adds the mean, the standard deviation, the slowest sample and the batch size. The comment lines say where the result came from; skip them when reading the file.
$ bmb_ddot -v 1024:8192 -s -o ddot.csv $ cat ddot.csv # blas-microbenchmark 1.5.0-dev # backend: openblas (OpenBLAS 0.3.26 NO_LAPACKE DYNAMIC_ARCH NO_AFFINITY Cooperlake MAX_THREADS=64) # verified: every point, against a reference, before it was timed # cpu: AMD EPYC 9V74 80-Core Processor (4 logical CPUs, 1 NUMA node) # caches: L1d 32K, L1i 32K, L2 1M, L3 32M # os: Linux 6.17.0-1022-azure x86_64 # date: 2026-10-12T09:48:02Z # routine: ddot thread_count,vector_size,time_s,gflops,gbytes_per_s,mean_s,stddev_s,max_s,calls_per_sample,median_s 1,1024,0.000000109,18.752850,150.022801,0.000000112,0.000000007,0.000000131,914,0.000000110 1,2048,0.000000267,15.346063,122.768504,0.000000271,0.000000011,0.000000303,373,0.000000268 1,4096,0.000000625,13.099028,104.792220,0.000000636,0.000000028,0.000000721,159,0.000000627 1,8192,0.000001232,13.295194,106.361552,0.000001282,0.000000106,0.000001586,80,0.000001237
JSON
The JSON holds the same as fields, and is what bmb_report(1) reads. Here with a label, to tell two runs on the same machine apart.
$ bmb_dgemv -m 512 -M 256:512 -l 'turbo off' -o dgemv.json
$ cat dgemv.json
{
"version": "1.5.0-dev",
"backend": "openblas",
"blas": "OpenBLAS 0.3.26 NO_LAPACKE DYNAMIC_ARCH NO_AFFINITY Cooperlake MAX_THREADS=64",
"label": "turbo off",
"verified": true,
"layout": "row",
"date": "2026-10-12T09:48:02Z",
"machine": {
"cpu": "AMD EPYC 9V74 80-Core Processor",
"logical_cpus": 4,
"numa_nodes": 1,
"caches": [
{"level": 1, "type": "data", "size_bytes": 32768},
{"level": 1, "type": "instruction", "size_bytes": 32768},
{"level": 2, "type": "unified", "size_bytes": 1048576},
{"level": 3, "type": "unified", "size_bytes": 33554432}
],
"os": "Linux 6.17.0-1022-azure x86_64"
},
"routine": "dgemv",
"dim1_label": "Matrix dim1 (M)",
"dim2_label": "Matrix dim2 (N)",
"results": [
{
"thread_count": 1,
"dim1": 512,
"dim2": 256,
"time_s": 0.000013223,
"gflops": 19.825350,
"gbytes_per_s": 80.075829
},
{
"thread_count": 1,
"dim1": 512,
"dim2": 512,
"time_s": 0.000026109,
"gflops": 20.081123,
"gbytes_per_s": 80.795143
}
]
}
The report
bmb_report results/*.json > report.html makes one page of all of them. These are from results measured on a laptop (Intel Core Ultra 7 155U, WSL2), with OpenBLAS 0.3.26 and BLIS 0.9.0: they show what the page draws, not which library is faster.
Summary
The best figure for each routine, and where every series came from.
Charts
Hovering a chart gives the exact values.
Repeated runs
The same ddot command run four times with each library, each run into its own file: a band shows how far the runs spread, and the comparison tells a difference within that noise (a hollow point) from one beyond it (see Repeated runs and noise).
Raw data
Every row, sortable by any column.
