Interleave width across the benchmark machines
The generic row kernels (gemv, trmv and the forward substitution on a
row-major matrix) sum W rows per pass, one accumulator per row (#749, #772).
These sweeps were built with the earlier default, W = 6 for float and double,
measured only on an i7-12700K (#775). This page measures W on every machine of
the #737 platform campaign (#780). The per-ISA defaults below come from it, and
so does the table run-time selection uses (#778).
How it was measured
Section titled “How it was measured”- Tool:
bench_interleave_rows --reps 21, n = 4096. It’s built in its own build directory, the same generic build as the platform session’snativevariant: no BLAS, noMTL5_NATIVE_FAST_GEMM, no-march. That means SSE2 on x86-64, NEON on AArch64, and/O2under MSVC. - Where: it runs on the session’s first CPU, under the session’s clock
policy, after its capability card’s runs (
INTERLEAVE=1in the machine profiles), and only on a machine that passed the session’s preflight. - Repeats: three runs per compiler in each session, and two sessions per machine, the second after a cold boot. Those are the same sessions as the machines’ capability cards (benchmark systems). Each cell below is the median of the 6 runs. Each run’s value is already a median of 21.
- Compilers: GCC and Clang on Linux, MSVC and clang-cl on Windows. The Jetson has no Clang.
- Why the median over runs: each run’s fastest width is listed too, but the median is what counts, because the first run of a sweep often favours a narrower width than the runs after it.
Findings
Section titled “Findings”-
x86-64: W = 8 is the better default. Over every x86-64 machine, compiler and kernel, W = 8 averages 1.5% slower than each kernel’s fastest width, 10% at worst (Clang on the i7-12700K, f32 gemv). W = 6 averages 4.1% slower, and 31% at worst: clang-cl’s f64
trmv, slower at W = 6 than at W = 4 on both Ryzens and in both sessions. On the two Zen 4 machines, under both compilers, W = 8 is the fastest width of every kernel or within 0.2% of it. The Ivy Bridge Xeon is the exception: it prefers 4 for f64 and loses at most 6% at W = 8. -
AArch64: W = 6 is not safe; W = 4 is the better default. With Clang (on the Arduino Ventuno Q), W = 4 is the fastest width in every kernel, and W = 6 is 18-24% slower than it. With GCC, gemv agrees: f64 and f32 are fastest at 4 on both AArch64 machines. On the Jetson, GCC’s f64
trmvand forward substitution are fastest at 2, and lose 12% at 4. -
GCC on AArch64 had a separate problem, now fixed (#795). In these sessions row-major f32
trmvand the forward substitution take 5.6-7.2x longer at W = 4 than at W = 2, and 1.4-4.5x longer at W = 6 to 12, on both machines (GCC 11 and 13). Clang on the same machine does not. The cause was thekeep_scalarbarrier from #776, which was on for GCC on AArch64 although it had been measured only on x86-64. An A/B on both boards (median of 3 runs,--reps 21):GCC, f32 W = 4, with the barrier W = 4, without it W = 2, before Jetson Orin, trmv50.04 ms 6.31 ms 7.29 ms Jetson Orin, forward substitution 39.45 ms 5.58 ms 7.08 ms Ventuno Q, trmv31.66 ms 3.12 ms 4.75 ms Ventuno Q, forward substitution 26.22 ms 2.99 ms 4.70 ms Jetson Orin, gemv8.23 ms 8.34 ms Ventuno Q, gemv5.51 ms 9.14 ms So on AArch64 the triangular kernels no longer take the barrier, and W = 4 is their fastest width, 13-36% faster than the 2 they were held at.
gemvkeeps it: it is neutral on the Jetson and 1.66x faster with it on the Ventuno. The tables below are from the 2026-11-11 sessions, with the barrier.
Defaults per ISA class
Section titled “Defaults per ISA class”These are the built defaults (include/mtl/interface/interleave_rows.hpp).
Every target caps the width at half its vector registers. Targets other than
x86-64 and AArch64 are assumed to have 8, so they get 4: the measured 6 that
was the default before #772 is capped there, as it is on 32-bit x86.
| ISA class | float | double | worst loss against the fastest width |
|---|---|---|---|
| x86-64 (SSE2, generic build) | 8 | 8 | 10% (i7-12700K, Clang, f32 gemv); 6% on the Xeon |
| AArch64 (NEON, generic build) | 4 | 4 | 0% under Clang; 12% under GCC (Jetson, f64 trmv and forward substitution) |
- x86-64 with AVX-512: none of these machines has it. The Zen 4 machines support AVX-512, but the generic build doesn’t use it, so that class is still unmeasured.
- Run-time selection (#778): it recovers the rest, for the CPUs it knows; see the next section.
Run-time selection (#778)
Section titled “Run-time selection (#778)”With MTL5_INTERLEAVE_RUNTIME, float and double compile each of the three
kernels at W = 2, 4, 6, 8 and 12. On first use they take the width a table
lists for the CPU and the compiler, and the compile-time default where it
lists nothing (include/mtl/interface/interleave_runtime.hpp). It is off by
default: a file that calls all three kernels for both types took 0.5-0.9 s
longer to compile on the i7 (1.2 s without), and their code grew from
24-29 KB to 62-78 KB.
The table is each kernel’s fastest width in the sweeps above, for the cores those machines have:
| core | identified by | GCC | Clang | MSVC, clang-cl |
|---|---|---|---|---|
| Alder Lake (i7-12700K) | Intel family 6, model 0x97, 0x9A | f64 12, 12, 12; f32 6, 8, 8 | f64 8; f32 6 | |
| Ivy Bridge (Xeon E5-2420 v2) | Intel family 6, model 0x3A, 0x3E | f64 4, 6, 4; f32 8, 4, 4 | f64 4; f32 6 | |
| Zen 4 (both Ryzens) | AMD family 0x19, model 0x60-0x7F | 8 | ||
| Cortex-A78AE (Jetson Orin) | Arm part 0xd42 | f64 4, 2, 2; f32 4 | ||
| Cortex-A78C (Ventuno Q) | Arm part 0xd4b | 4 | 4 |
The three numbers are gemv, trmv and the forward substitution. On
AArch64 the part comes from /proc/cpuinfo (Linux only). On the Ventuno that
is the A78C, not the A55, which is where the sweeps ran. The Jetson’s f32
trmv and forward substitution are 4 since #795. The sweeps were a generic
build, so an x86-64 build with AVX uses no entry.
How close it comes, measured with the table in place
(bench_interleave_rows, same method as above, one session, median of the
runs): the worst kernel, as time at the chosen width over time at the
fastest, against today’s fixed default.
| machine, compiler | runs | run-time choice, worst | fixed default, worst |
|---|---|---|---|
| i7-12700K, GCC | 6 | 1.007 (f64 gemv) | 1.044 (f64 trmv) |
| i7-12700K, Clang | 3 | 1.007 (f64 trmv) | 1.094 (f32 gemv) |
| Jetson Orin, GCC | 3 | 1.000 | 1.128 (f64 trmv) |
| Ventuno Q, GCC | 3 | 1.000 | 1.000 |
| Ventuno Q, Clang | 3 | 1.000 | 1.000 |
The Xeon and the two Ryzens weren’t reachable for this. Their entries are the sweeps’ fastest widths, so in those sweeps the choice is the fastest by construction. Against the fixed default that is up to 6% (the Xeon) and 0.2% (the Ryzens). They are still to be measured with the table in place.
f64 gemv on the i7 under GCC is bandwidth-bound at n = 4096. Widths 6 to 16
are within about 5% of each other there, and run to run the fastest moves
among 8, 12 and 16.
The dispatch is a switch on a cached width per call.
bench_interleave_dispatch times the generic gemv built without and with
run-time selection, at the same width, out of line as a caller’s call is.
At n = 64 the cost is under 1% on the i7 (GCC and Clang), the Jetson and the
Ventuno under Clang, and 1-3.4% on the Ventuno under GCC (f64 the most). At
n = 256 it is at most 1.6% (Clang on the i7). At n = 16 it is 1-2 ns on the
i7, and inside the run-to-run spread on the boards.
The table’s widths are for n = 4096. At small n they need not be the fastest: on the i7 under GCC, the chosen 12 (f64) and 6 (f32) are 3-4% slower than 8 at n = 64 and 2-3% at n = 256. Under Clang the chosen f32 6 is 8% faster than 8 at n = 64 and 256, and level with it at n = 16.
Summary
Section titled “Summary”Per machine and compiler: the fastest width of each kernel, and the worst cost of a fixed width against the fastest (time at that width / time at the fastest; 1.000 is no loss).
| machine | ISA | compiler | fastest (f64 gemv, trmv, trisolve / f32 …) | W=4 | W=6 | W=8 | W=6 / W=4 |
|---|---|---|---|---|---|---|---|
| arduino-ventuno-q | AArch64 | clang++ | 4, 4, 4 / 4, 4, 4 | 1.000 | 1.237 | 1.472 | 1.237 |
| arduino-ventuno-q | AArch64 | g++ | 4, 4, 4 / 4, 2, 2 | 6.654 | 4.408 | 3.164 | 1.252 |
| i7-12700k | x86-64 | clang++ | 8, 8, 8 / 6, 6, 6 | 1.124 | 1.052 | 1.100 | 0.957 |
| i7-12700k | x86-64 | g++ | 12, 12, 12 / 6, 8, 8 | 1.088 | 1.041 | 1.029 | 0.984 |
| jetson-orin | AArch64 | g++ | 4, 2, 2 / 4, 2, 2 | 7.186 | 3.976 | 4.532 | 1.582 |
| ryzen-7-8845hs | x86-64 | clang-cl | 8, 8, 6 / 8, 8, 8 | 1.189 | 1.310 | 1.002 | 1.149 |
| ryzen-7-8845hs | x86-64 | msvc | 8, 8, 8 / 8, 8, 8 | 1.323 | 1.078 | 1.000 | 0.976 |
| ryzen-9-8945hs | x86-64 | clang-cl | 8, 8, 8 / 8, 8, 8 | 1.216 | 1.315 | 1.000 | 1.137 |
| ryzen-9-8945hs | x86-64 | msvc | 8, 8, 8 / 8, 8, 8 | 1.330 | 1.059 | 1.000 | 0.988 |
| xeon-e5-2420v2 | x86-64 | clang++ | 4, 4, 4 / 6, 6, 6 | 1.092 | 1.039 | 1.063 | 1.039 |
| xeon-e5-2420v2 | x86-64 | g++ | 4, 6, 4 / 8, 4, 4 | 1.038 | 1.026 | 1.046 | 1.019 |
A fixed default per ISA class
Section titled “A fixed default per ISA class”Time at a fixed width / time at the fastest width, over every machine, compiler, kernel and type of the class: geometric mean / worst case.
| ISA | W=2 | W=4 | W=6 | W=8 | W=12 |
|---|---|---|---|---|---|
| x86-64 | 1.628 / 2.683 | 1.102 / 1.330 | 1.041 / 1.315 | 1.015 / 1.100 | 1.518 / 3.119 |
| AArch64 | 1.242 / 1.708 | 1.523 / 7.186 | 1.500 / 4.408 | 1.640 / 4.532 | 2.053 / 3.799 |
Per machine
Section titled “Per machine”arduino-ventuno-q — clang++
Section titled “arduino-ventuno-q — clang++”0xd4b; Ubuntu clang version 18.1.3 (1ubuntu1); vector registers 32; built default W = 6 (f64), 6 (f32); median of 6 runs (2 sessions), each the median of 21. Data: benchmarks/data/arduino-ventuno-q/interleave_rows/.
| kernel | W=1 | W=2 | W=4 | W=6 | W=8 | W=12 | W=16 | fastest | W=6 / fastest | W=6 / W=4 | per-run fastest |
|---|---|---|---|---|---|---|---|---|---|---|---|
| gemv f64 | 17.749 | 10.629 | 8.254 | 9.840 | 10.425 | 12.238 | 14.228 | 4 | 1.192 | 1.192 | 4: 6 |
| trmv f64 | 10.225 | 5.800 | 4.523 | 5.381 | 5.768 | 6.620 | 10.535 | 4 | 1.190 | 1.190 | 4: 6 |
| trisolve f64 | 10.282 | 5.877 | 4.483 | 5.300 | 5.620 | 8.893 | 10.459 | 4 | 1.182 | 1.182 | 4: 6 |
| gemv f32 | 16.730 | 9.563 | 6.389 | 7.544 | 9.406 | 11.408 | 14.054 | 4 | 1.181 | 1.181 | 4: 6 |
| trmv f32 | 8.852 | 5.042 | 3.290 | 4.059 | 4.429 | 5.426 | 9.395 | 4 | 1.234 | 1.234 | 4: 6 |
| trisolve f32 | 8.729 | 4.962 | 3.179 | 3.934 | 4.313 | 8.601 | 9.318 | 4 | 1.237 | 1.237 | 4: 6 |
arduino-ventuno-q — g++
Section titled “arduino-ventuno-q — g++”0xd4b; g++ (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0; vector registers 32; built default W = 6 (f64), 6 (f32); median of 6 runs (2 sessions), each the median of 21. Data: benchmarks/data/arduino-ventuno-q/interleave_rows/.
| kernel | W=1 | W=2 | W=4 | W=6 | W=8 | W=12 | W=16 | fastest | W=6 / fastest | W=6 / W=4 | per-run fastest |
|---|---|---|---|---|---|---|---|---|---|---|---|
| gemv f64 | 18.844 | 10.431 | 8.104 | 9.587 | 9.992 | 12.108 | 14.022 | 4 | 1.183 | 1.183 | 4: 6 |
| trmv f64 | 10.479 | 5.710 | 4.449 | 5.124 | 5.359 | 6.266 | 6.877 | 4 | 1.152 | 1.152 | 4: 6 |
| trisolve f64 | 10.442 | 5.777 | 4.393 | 5.118 | 5.292 | 6.106 | 6.708 | 4 | 1.165 | 1.165 | 4: 6 |
| gemv f32 | 16.659 | 9.005 | 5.556 | 6.958 | 7.777 | 9.213 | 11.011 | 4 | 1.252 | 1.252 | 4: 6 |
| trmv f32 | 8.976 | 4.761 | 31.675 | 6.787 | 7.572 | 7.675 | 8.175 | 2 | 1.426 | 0.214 | 2: 6 |
| trisolve f32 | 8.755 | 4.712 | 26.197 | 20.768 | 14.909 | 15.145 | 17.940 | 2 | 4.408 | 0.793 | 2: 6 |
i7-12700k — clang++
Section titled “i7-12700k — clang++”12th Gen Intel(R) Core(TM) i7-12700K; Ubuntu clang version 21.1.8 (6ubuntu1); vector registers 16; built default W = 6 (f64), 6 (f32); median of 6 runs (2 sessions), each the median of 21. Data: benchmarks/data/i7-12700k/interleave_rows/.
| kernel | W=1 | W=2 | W=4 | W=6 | W=8 | W=12 | W=16 | fastest | W=6 / fastest | W=6 / W=4 | per-run fastest |
|---|---|---|---|---|---|---|---|---|---|---|---|
| gemv f64 | 8.338 | 5.115 | 4.002 | 3.817 | 3.705 | 3.751 | 3.845 | 8 | 1.030 | 0.954 | 8: 6 |
| trmv f64 | 4.346 | 2.725 | 2.125 | 1.993 | 1.921 | 1.951 | 2.694 | 8 | 1.037 | 0.938 | 8: 4, 12: 2 |
| trisolve f64 | 4.304 | 2.710 | 2.085 | 1.996 | 1.897 | 2.242 | 2.749 | 8 | 1.052 | 0.957 | 8: 6 |
| gemv f32 | 6.990 | 3.940 | 2.910 | 2.625 | 2.887 | 2.876 | 2.970 | 6 | 1.000 | 0.902 | 6: 6 |
| trmv f32 | 3.759 | 2.110 | 1.589 | 1.416 | 1.518 | 1.508 | 2.292 | 6 | 1.000 | 0.891 | 6: 6 |
| trisolve f32 | 3.728 | 2.059 | 1.542 | 1.372 | 1.474 | 1.906 | 2.359 | 6 | 1.000 | 0.889 | 6: 6 |
i7-12700k — g++
Section titled “i7-12700k — g++”12th Gen Intel(R) Core(TM) i7-12700K; g++ (Ubuntu 15.2.0-16ubuntu1) 15.2.0; vector registers 16; built default W = 6 (f64), 6 (f32); median of 6 runs (2 sessions), each the median of 21. Data: benchmarks/data/i7-12700k/interleave_rows/.
| kernel | W=1 | W=2 | W=4 | W=6 | W=8 | W=12 | W=16 | fastest | W=6 / fastest | W=6 / W=4 | per-run fastest |
|---|---|---|---|---|---|---|---|---|---|---|---|
| gemv f64 | 8.373 | 5.096 | 3.817 | 3.708 | 3.633 | 3.562 | 3.647 | 12 | 1.041 | 0.971 | 12: 5, 16: 1 |
| trmv f64 | 4.331 | 2.665 | 2.107 | 2.010 | 2.010 | 1.953 | 2.054 | 12 | 1.029 | 0.954 | 12: 6 |
| trisolve f64 | 4.309 | 2.645 | 2.079 | 1.976 | 1.933 | 1.910 | 2.045 | 12 | 1.034 | 0.950 | 12: 6 |
| gemv f32 | 6.968 | 7.333 | 2.779 | 2.734 | 2.745 | 2.769 | 3.510 | 6 | 1.000 | 0.984 | 6: 3, 8: 1, 12: 2 |
| trmv f32 | 3.734 | 2.052 | 1.544 | 1.472 | 1.441 | 1.458 | 1.840 | 8 | 1.022 | 0.953 | 8: 6 |
| trisolve f32 | 3.694 | 2.016 | 1.481 | 1.432 | 1.413 | 1.421 | 1.768 | 8 | 1.013 | 0.967 | 8: 5, 12: 1 |
jetson-orin — g++
Section titled “jetson-orin — g++”ARMv8 Processor rev 1 (v8l); g++ (Ubuntu 11.4.0-1ubuntu1~22.04.3) 11.4.0; vector registers 32; built default W = 6 (f64), 6 (f32); median of 6 runs (2 sessions), each the median of 21. Data: benchmarks/data/jetson-orin/interleave_rows/.
| kernel | W=1 | W=2 | W=4 | W=6 | W=8 | W=12 | W=16 | fastest | W=6 / fastest | W=6 / W=4 | per-run fastest |
|---|---|---|---|---|---|---|---|---|---|---|---|
| gemv f64 | 25.134 | 13.599 | 12.888 | 19.942 | 24.796 | 41.006 | 40.270 | 4 | 1.547 | 1.547 | 4: 6 |
| trmv f64 | 13.853 | 7.976 | 8.976 | 12.552 | 14.895 | 21.994 | 21.157 | 2 | 1.574 | 1.398 | 2: 5, 4: 1 |
| trisolve f64 | 13.549 | 7.800 | 8.710 | 12.431 | 15.082 | 22.457 | 21.260 | 2 | 1.594 | 1.427 | 2: 5, 4: 1 |
| gemv f32 | 24.912 | 13.213 | 7.735 | 12.236 | 14.143 | 18.181 | 23.286 | 4 | 1.582 | 1.582 | 4: 6 |
| trmv f32 | 13.188 | 6.926 | 49.775 | 11.046 | 12.965 | 14.921 | 17.179 | 2 | 1.595 | 0.222 | 2: 6 |
| trisolve f32 | 13.027 | 6.800 | 39.427 | 27.035 | 30.816 | 25.833 | 28.998 | 2 | 3.976 | 0.686 | 2: 6 |
ryzen-7-8845hs — clang-cl
Section titled “ryzen-7-8845hs — clang-cl”AMD Ryzen 7 8845HS w/ Radeon 780M Graphics; Clang 22.1.3 with MSVC-like command-line; vector registers 16; built default W = 6 (f64), 6 (f32); median of 6 runs (2 sessions), each the median of 21. Data: benchmarks/data/ryzen-7-8845hs/interleave_rows/.
| kernel | W=1 | W=2 | W=4 | W=6 | W=8 | W=12 | W=16 | fastest | W=6 / fastest | W=6 / W=4 | per-run fastest |
|---|---|---|---|---|---|---|---|---|---|---|---|
| gemv f64 | 10.704 | 5.841 | 3.489 | 3.206 | 3.119 | 6.969 | 7.739 | 8 | 1.028 | 0.919 | 6: 2, 8: 4 |
| trmv f64 | 5.734 | 3.212 | 1.971 | 2.265 | 1.728 | 3.738 | 3.986 | 8 | 1.310 | 1.149 | 8: 6 |
| trisolve f64 | 5.745 | 3.236 | 1.925 | 1.762 | 1.765 | 3.715 | 3.929 | 6 | 1.000 | 0.915 | 6: 2, 8: 4 |
| gemv f32 | 10.095 | 5.453 | 2.997 | 2.738 | 2.520 | 7.824 | 9.236 | 8 | 1.087 | 0.914 | 6: 2, 8: 4 |
| trmv f32 | 5.428 | 2.894 | 1.614 | 1.448 | 1.439 | 3.975 | 4.825 | 8 | 1.006 | 0.897 | 6: 2, 8: 4 |
| trisolve f32 | 5.408 | 2.911 | 1.700 | 1.481 | 1.431 | 3.925 | 4.713 | 8 | 1.035 | 0.871 | 6: 2, 8: 4 |
ryzen-7-8845hs — msvc
Section titled “ryzen-7-8845hs — msvc”AMD Ryzen 7 8845HS w/ Radeon 780M Graphics; MSVC 19.52.37029.0; vector registers 16; built default W = 6 (f64), 6 (f32); median of 6 runs (2 sessions), each the median of 21. Data: benchmarks/data/ryzen-7-8845hs/interleave_rows/.
| kernel | W=1 | W=2 | W=4 | W=6 | W=8 | W=12 | W=16 | fastest | W=6 / fastest | W=6 / W=4 | per-run fastest |
|---|---|---|---|---|---|---|---|---|---|---|---|
| gemv f64 | 10.698 | 5.836 | 3.619 | 3.172 | 3.046 | 4.176 | 8.704 | 8 | 1.041 | 0.877 | 8: 6 |
| trmv f64 | 5.753 | 3.180 | 1.929 | 1.804 | 1.710 | 2.575 | 4.574 | 8 | 1.055 | 0.935 | 6: 1, 8: 5 |
| trisolve f64 | 5.814 | 3.224 | 1.933 | 1.785 | 1.697 | 2.549 | 4.504 | 8 | 1.052 | 0.923 | 8: 6 |
| gemv f32 | 10.087 | 5.391 | 3.046 | 2.482 | 2.302 | 4.345 | 9.029 | 8 | 1.078 | 0.815 | 6: 1, 8: 5 |
| trmv f32 | 5.432 | 2.832 | 1.558 | 1.520 | 1.480 | 2.489 | 4.731 | 8 | 1.027 | 0.976 | 6: 1, 8: 5 |
| trisolve f32 | 5.389 | 2.864 | 1.621 | 1.567 | 1.493 | 2.474 | 4.663 | 8 | 1.049 | 0.966 | 6: 1, 8: 5 |
ryzen-9-8945hs — clang-cl
Section titled “ryzen-9-8945hs — clang-cl”AMD Ryzen 9 8945HS w/ Radeon 780M Graphics; Clang 22.1.3 with MSVC-like command-line; vector registers 16; built default W = 6 (f64), 6 (f32); median of 6 runs (2 sessions), each the median of 21. Data: benchmarks/data/ryzen-9-8945hs/interleave_rows/.
| kernel | W=1 | W=2 | W=4 | W=6 | W=8 | W=12 | W=16 | fastest | W=6 / fastest | W=6 / W=4 | per-run fastest |
|---|---|---|---|---|---|---|---|---|---|---|---|
| gemv f64 | 10.586 | 5.928 | 3.595 | 3.247 | 3.119 | 6.877 | 7.694 | 8 | 1.041 | 0.903 | 8: 6 |
| trmv f64 | 5.694 | 3.229 | 2.002 | 2.277 | 1.731 | 3.752 | 3.994 | 8 | 1.315 | 1.137 | 8: 6 |
| trisolve f64 | 5.697 | 3.224 | 1.939 | 1.776 | 1.741 | 3.665 | 3.963 | 8 | 1.020 | 0.916 | 8: 6 |
| gemv f32 | 9.951 | 5.425 | 2.988 | 2.665 | 2.457 | 7.663 | 9.101 | 8 | 1.085 | 0.892 | 8: 6 |
| trmv f32 | 6.704 | 3.550 | 1.895 | 1.716 | 1.647 | 4.951 | 6.056 | 8 | 1.043 | 0.906 | 8: 6 |
| trisolve f32 | 6.674 | 3.547 | 1.996 | 1.772 | 1.644 | 4.931 | 5.927 | 8 | 1.078 | 0.888 | 8: 6 |
ryzen-9-8945hs — msvc
Section titled “ryzen-9-8945hs — msvc”AMD Ryzen 9 8945HS w/ Radeon 780M Graphics; MSVC 19.52.37029.0; vector registers 16; built default W = 6 (f64), 6 (f32); median of 6 runs (2 sessions), each the median of 21. Data: benchmarks/data/ryzen-9-8945hs/interleave_rows/.
| kernel | W=1 | W=2 | W=4 | W=6 | W=8 | W=12 | W=16 | fastest | W=6 / fastest | W=6 / W=4 | per-run fastest |
|---|---|---|---|---|---|---|---|---|---|---|---|
| gemv f64 | 10.596 | 5.864 | 3.606 | 3.189 | 3.066 | 4.147 | 8.617 | 8 | 1.040 | 0.884 | 8: 6 |
| trmv f64 | 5.752 | 3.201 | 1.972 | 1.797 | 1.721 | 2.583 | 4.561 | 8 | 1.044 | 0.911 | 8: 6 |
| trisolve f64 | 5.775 | 3.217 | 1.966 | 1.785 | 1.696 | 2.552 | 4.457 | 8 | 1.053 | 0.908 | 8: 6 |
| gemv f32 | 10.043 | 5.300 | 2.986 | 2.379 | 2.245 | 4.290 | 8.878 | 8 | 1.059 | 0.797 | 8: 6 |
| trmv f32 | 5.351 | 2.808 | 1.880 | 1.857 | 1.769 | 3.107 | 5.933 | 8 | 1.049 | 0.988 | 4: 1, 6: 1, 8: 4 |
| trisolve f32 | 6.700 | 3.521 | 1.940 | 1.869 | 1.792 | 3.082 | 5.856 | 8 | 1.043 | 0.964 | 8: 6 |
xeon-e5-2420v2 — clang++
Section titled “xeon-e5-2420v2 — clang++”Intel(R) Xeon(R) CPU E5-2420 v2 @ 2.20GHz; Ubuntu clang version 21.1.8 (6ubuntu1); vector registers 16; built default W = 6 (f64), 6 (f32); median of 6 runs (2 sessions), each the median of 21. Data: benchmarks/data/xeon-e5-2420v2/interleave_rows/.
| kernel | W=1 | W=2 | W=4 | W=6 | W=8 | W=12 | W=16 | fastest | W=6 / fastest | W=6 / W=4 | per-run fastest |
|---|---|---|---|---|---|---|---|---|---|---|---|
| gemv f64 | 19.770 | 13.851 | 13.125 | 13.496 | 13.701 | 15.127 | 17.516 | 4 | 1.028 | 1.028 | 4: 6 |
| trmv f64 | 10.511 | 7.554 | 7.171 | 7.288 | 7.361 | 8.212 | 15.052 | 4 | 1.016 | 1.016 | 4: 4, 6: 2 |
| trisolve f64 | 10.198 | 7.312 | 7.049 | 7.321 | 7.274 | 10.136 | 15.041 | 4 | 1.039 | 1.039 | 4: 6 |
| gemv f32 | 19.439 | 10.723 | 9.085 | 8.521 | 8.698 | 10.037 | 11.502 | 6 | 1.000 | 0.938 | 6: 5, 8: 1 |
| trmv f32 | 10.194 | 5.627 | 4.750 | 4.463 | 4.664 | 5.307 | 10.866 | 6 | 1.000 | 0.940 | 6: 5, 8: 1 |
| trisolve f32 | 9.957 | 5.442 | 4.637 | 4.248 | 4.515 | 8.652 | 11.117 | 6 | 1.000 | 0.916 | 6: 6 |
xeon-e5-2420v2 — g++
Section titled “xeon-e5-2420v2 — g++”Intel(R) Xeon(R) CPU E5-2420 v2 @ 2.20GHz; g++ (Ubuntu 15.2.0-16ubuntu1) 15.2.0; vector registers 16; built default W = 6 (f64), 6 (f32); median of 6 runs (2 sessions), each the median of 21. Data: benchmarks/data/xeon-e5-2420v2/interleave_rows/.
| kernel | W=1 | W=2 | W=4 | W=6 | W=8 | W=12 | W=16 | fastest | W=6 / fastest | W=6 / W=4 | per-run fastest |
|---|---|---|---|---|---|---|---|---|---|---|---|
| gemv f64 | 20.675 | 13.618 | 13.177 | 13.296 | 13.676 | 13.981 | 15.346 | 4 | 1.009 | 1.009 | 4: 5, 6: 1 |
| trmv f64 | 10.344 | 7.333 | 7.067 | 7.050 | 7.270 | 7.674 | 9.014 | 6 | 1.000 | 0.998 | 4: 3, 6: 3 |
| trisolve f64 | 10.091 | 7.210 | 6.854 | 6.931 | 7.169 | 7.363 | 8.803 | 4 | 1.011 | 1.011 | 4: 5, 6: 1 |
| gemv f32 | 19.093 | 19.755 | 8.752 | 8.648 | 8.431 | 10.623 | 14.854 | 8 | 1.026 | 0.988 | 4: 1, 8: 5 |
| trmv f32 | 9.998 | 5.471 | 4.471 | 4.523 | 4.627 | 5.483 | 7.484 | 4 | 1.012 | 1.012 | 4: 4, 6: 2 |
| trisolve f32 | 10.029 | 5.353 | 4.283 | 4.364 | 4.406 | 5.237 | 7.024 | 4 | 1.019 | 1.019 | 4: 6 |
Reproducing
Section titled “Reproducing”Run a platform session with INTERLEAVE=1: --interleave in
benchmarks/remote_session.sh, or $env:INTERLEAVE = '1' for a Windows
profile. Publish it with benchmarks/publish_platform.cmake. Then:
python3 benchmarks/interleave_table.py import <machine> <session stamp> <results dir>python3 benchmarks/interleave_table.py tableThe first command copies the sweeps to
benchmarks/data/<machine>/interleave_rows/<stamp>/: one CSV per compiler,
with its .sysinfo sidecar, and the raw outputs. The second rewrites the
generated part of this page.