git/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp synced 2026-04-23 07:11:30 +02:00
Code Issues Packages Projects Releases Wiki Activity

benches : update

Browse Source
This commit is contained in:
Georgi Gerganov 2026-03-18 22:34:51 +02:00
parent f5b477ab09
commit 4bbce1e5b2
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
3 changed files with 127 additions and 120 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

41
examples/bench/bench.cpp
View File

@ -85,33 +85,38 @@ static int whisper_bench_full(const whisper_params & params) {
fprintf(stderr, "error: failed to set mel: %d\n", ret);
return 3;
}
// heat encoder
if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
fprintf(stderr, "error: failed to encode: %d\n", ret);
return 4;
}
whisper_token tokens[512];
memset(tokens, 0, sizeof(tokens));
// prompt heat
if (int ret = whisper_decode(ctx, tokens, 256, 0, params.n_threads) != 0) {
fprintf(stderr, "error: failed to decode: %d\n", ret);
return 4;
}
// TODO: need 2 loops because of the current graph capture logic in the CUDA backend
// https://github.com/ggml-org/llama.cpp/pull/19754
for (int h = 0; h < 2; ++h) {
// heat encoder
if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
fprintf(stderr, "error: failed to encode: %d\n", ret);
return 4;
}
// text-generation heat
for (int i = 0; i < 256; i++) {
if (int ret = whisper_decode(ctx, tokens, 1, i, params.n_threads) != 0) {
// prompt heat
if (int ret = whisper_decode(ctx, tokens, 256, 0, params.n_threads) != 0) {
fprintf(stderr, "error: failed to decode: %d\n", ret);
return 4;
}
}
// batched heat
if (int ret = whisper_decode(ctx, tokens, 5, 0, params.n_threads) != 0) {
fprintf(stderr, "error: failed to decode: %d\n", ret);
return 4;
// text-generation heat
for (int i = 0; i < 256; i++) {
if (int ret = whisper_decode(ctx, tokens, 1, i, params.n_threads) != 0) {
fprintf(stderr, "error: failed to decode: %d\n", ret);
return 4;
}
}
// batched heat
if (int ret = whisper_decode(ctx, tokens, 5, 0, params.n_threads) != 0) {
fprintf(stderr, "error: failed to decode: %d\n", ret);
return 4;
}
}
whisper_reset_timings(ctx);

196
scripts/bench-all-gg.txt
View File

@ -111,61 +111,61 @@ make -j && ./scripts/bench-all.sh 1 1 0
| CPU | Config | Model | Th | FA | Enc. | Dec. | Bch5 | PP | Commit |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| M2 ULTRA | METAL | tiny | 1 | 0 | 8.80 | 1.13 | 0.28 | 0.01 | 47af2fb7 |
| M2 ULTRA | METAL | tiny-q5_0 | 1 | 0 | 9.34 | 1.09 | 0.28 | 0.01 | 47af2fb7 |
| M2 ULTRA | METAL | tiny-q5_1 | 1 | 0 | 9.29 | 1.09 | 0.29 | 0.01 | 47af2fb7 |
| M2 ULTRA | METAL | tiny-q8_0 | 1 | 0 | 9.00 | 1.12 | 0.28 | 0.01 | 47af2fb7 |
| M2 ULTRA | METAL | base | 1 | 0 | 15.92 | 1.60 | 0.43 | 0.02 | 47af2fb7 |
| M2 ULTRA | METAL | base-q5_0 | 1 | 0 | 17.01 | 1.53 | 0.43 | 0.02 | 47af2fb7 |
| M2 ULTRA | METAL | base-q5_1 | 1 | 0 | 17.02 | 1.53 | 0.44 | 0.02 | 47af2fb7 |
| M2 ULTRA | METAL | base-q8_0 | 1 | 0 | 16.25 | 1.55 | 0.43 | 0.02 | 47af2fb7 |
| M2 ULTRA | METAL | small | 1 | 0 | 47.83 | 3.09 | 0.91 | 0.05 | 47af2fb7 |
| M2 ULTRA | METAL | small-q5_0 | 1 | 0 | 52.85 | 2.98 | 0.94 | 0.06 | 47af2fb7 |
| M2 ULTRA | METAL | small-q5_1 | 1 | 0 | 52.92 | 2.97 | 0.94 | 0.06 | 47af2fb7 |
| M2 ULTRA | METAL | small-q8_0 | 1 | 0 | 49.05 | 2.89 | 0.90 | 0.06 | 47af2fb7 |
| M2 ULTRA | METAL | medium | 1 | 0 | 127.98 | 6.62 | 2.05 | 0.12 | 47af2fb7 |
| M2 ULTRA | METAL | medium-q5_0 | 1 | 0 | 145.42 | 6.09 | 2.12 | 0.14 | 47af2fb7 |
| M2 ULTRA | METAL | medium-q5_1 | 1 | 0 | 145.16 | 6.08 | 2.14 | 0.14 | 47af2fb7 |
| M2 ULTRA | METAL | medium-q8_0 | 1 | 0 | 132.72 | 6.10 | 2.07 | 0.13 | 47af2fb7 |
| M2 ULTRA | METAL | medium-dis | 1 | 0 | 115.09 | 0.91 | 0.25 | 0.02 | 47af2fb7 |
| M2 ULTRA | METAL | large-v2 | 1 | 0 | 243.69 | 9.68 | 3.14 | 0.22 | 47af2fb7 |
| M2 ULTRA | METAL | large-v2-q5_0 | 1 | 0 | 280.38 | 8.95 | 3.18 | 0.25 | 47af2fb7 |
| M2 ULTRA | METAL | large-v2-q5_1 | 1 | 0 | 279.76 | 8.92 | 3.18 | 0.25 | 47af2fb7 |
| M2 ULTRA | METAL | large-v2-q8_0 | 1 | 0 | 254.55 | 9.35 | 3.04 | 0.23 | 47af2fb7 |
| M2 ULTRA | METAL | large-v2-dis | 1 | 0 | 219.23 | 1.01 | 0.28 | 0.02 | 47af2fb7 |
| M2 ULTRA | METAL | large-v3-turbo | 1 | 0 | 220.57 | 1.55 | 0.46 | 0.03 | 47af2fb7 |
| M2 ULTRA | METAL | large-v3-turbo-q5_0 | 1 | 0 | 253.03 | 1.40 | 0.47 | 0.04 | 47af2fb7 |
| M2 ULTRA | METAL | large-v3-turbo-q8_0 | 1 | 0 | 229.82 | 1.43 | 0.45 | 0.04 | 47af2fb7 |
| M2 ULTRA | METAL | tiny | 1 | 0 | 8.57 | 1.12 | 0.27 | 0.01 | f5b477ab |
| M2 ULTRA | METAL | tiny-q5_0 | 1 | 0 | 9.17 | 1.10 | 0.28 | 0.01 | f5b477ab |
| M2 ULTRA | METAL | tiny-q5_1 | 1 | 0 | 9.16 | 1.09 | 0.28 | 0.01 | f5b477ab |
| M2 ULTRA | METAL | tiny-q8_0 | 1 | 0 | 8.81 | 1.12 | 0.27 | 0.01 | f5b477ab |
| M2 ULTRA | METAL | base | 1 | 0 | 15.60 | 1.61 | 0.41 | 0.02 | f5b477ab |
| M2 ULTRA | METAL | base-q5_0 | 1 | 0 | 16.75 | 1.54 | 0.42 | 0.02 | f5b477ab |
| M2 ULTRA | METAL | base-q5_1 | 1 | 0 | 16.64 | 1.54 | 0.43 | 0.02 | f5b477ab |
| M2 ULTRA | METAL | base-q8_0 | 1 | 0 | 16.09 | 1.55 | 0.41 | 0.02 | f5b477ab |
| M2 ULTRA | METAL | small | 1 | 0 | 46.74 | 3.13 | 0.89 | 0.05 | f5b477ab |
| M2 ULTRA | METAL | small-q5_0 | 1 | 0 | 51.57 | 3.03 | 0.91 | 0.06 | f5b477ab |
| M2 ULTRA | METAL | small-q5_1 | 1 | 0 | 51.85 | 3.03 | 0.92 | 0.06 | f5b477ab |
| M2 ULTRA | METAL | small-q8_0 | 1 | 0 | 48.34 | 3.01 | 0.89 | 0.06 | f5b477ab |
| M2 ULTRA | METAL | medium | 1 | 0 | 125.82 | 6.46 | 2.01 | 0.12 | f5b477ab |
| M2 ULTRA | METAL | medium-q5_0 | 1 | 0 | 143.44 | 5.97 | 2.07 | 0.14 | f5b477ab |
| M2 ULTRA | METAL | medium-q5_1 | 1 | 0 | 143.41 | 5.97 | 2.09 | 0.14 | f5b477ab |
| M2 ULTRA | METAL | medium-q8_0 | 1 | 0 | 131.23 | 6.30 | 2.01 | 0.13 | f5b477ab |
| M2 ULTRA | METAL | medium-dis | 1 | 0 | 114.07 | 0.90 | 0.25 | 0.02 | f5b477ab |
| M2 ULTRA | METAL | large-v2 | 1 | 0 | 240.73 | 9.46 | 3.21 | 0.21 | f5b477ab |
| M2 ULTRA | METAL | large-v2-q5_0 | 1 | 0 | 276.56 | 8.62 | 3.16 | 0.25 | f5b477ab |
| M2 ULTRA | METAL | large-v2-q5_1 | 1 | 0 | 275.90 | 8.98 | 3.16 | 0.25 | f5b477ab |
| M2 ULTRA | METAL | large-v2-q8_0 | 1 | 0 | 251.00 | 9.10 | 3.02 | 0.22 | f5b477ab |
| M2 ULTRA | METAL | large-v2-dis | 1 | 0 | 217.43 | 1.01 | 0.28 | 0.02 | f5b477ab |
| M2 ULTRA | METAL | large-v3-turbo | 1 | 0 | 218.39 | 1.55 | 0.47 | 0.03 | f5b477ab |
| M2 ULTRA | METAL | large-v3-turbo-q5_0 | 1 | 0 | 249.41 | 1.39 | 0.47 | 0.04 | f5b477ab |
| M2 ULTRA | METAL | large-v3-turbo-q8_0 | 1 | 0 | 227.54 | 1.43 | 0.45 | 0.03 | f5b477ab |
make -j && ./scripts/bench-all.sh 1 1 1
| CPU | Config | Model | Th | FA | Enc. | Dec. | Bch5 | PP | Commit |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| M2 ULTRA | METAL | tiny | 1 | 1 | 6.19 | 0.93 | 0.21 | 0.01 | 47af2fb7 |
| M2 ULTRA | METAL | tiny-q5_0 | 1 | 1 | 6.64 | 0.89 | 0.22 | 0.01 | 47af2fb7 |
| M2 ULTRA | METAL | tiny-q5_1 | 1 | 1 | 6.65 | 0.91 | 0.23 | 0.01 | 47af2fb7 |
| M2 ULTRA | METAL | tiny-q8_0 | 1 | 1 | 6.26 | 0.93 | 0.22 | 0.01 | 47af2fb7 |
| M2 ULTRA | METAL | base | 1 | 1 | 10.89 | 1.31 | 0.32 | 0.02 | 47af2fb7 |
| M2 ULTRA | METAL | base-q5_0 | 1 | 1 | 12.10 | 1.22 | 0.33 | 0.02 | 47af2fb7 |
| M2 ULTRA | METAL | base-q5_1 | 1 | 1 | 12.05 | 1.22 | 0.33 | 0.02 | 47af2fb7 |
| M2 ULTRA | METAL | base-q8_0 | 1 | 1 | 11.24 | 1.24 | 0.32 | 0.02 | 47af2fb7 |
| M2 ULTRA | METAL | small | 1 | 1 | 32.06 | 2.41 | 0.64 | 0.04 | 47af2fb7 |
| M2 ULTRA | METAL | small-q5_0 | 1 | 1 | 37.20 | 2.32 | 0.67 | 0.04 | 47af2fb7 |
| M2 ULTRA | METAL | small-q5_1 | 1 | 1 | 37.13 | 2.30 | 0.67 | 0.04 | 47af2fb7 |
| M2 ULTRA | METAL | small-q8_0 | 1 | 1 | 33.63 | 2.28 | 0.64 | 0.04 | 47af2fb7 |
| M2 ULTRA | METAL | medium | 1 | 1 | 89.22 | 5.14 | 1.46 | 0.09 | 47af2fb7 |
| M2 ULTRA | METAL | medium-q5_0 | 1 | 1 | 106.82 | 4.83 | 1.49 | 0.11 | 47af2fb7 |
| M2 ULTRA | METAL | medium-q5_1 | 1 | 1 | 106.60 | 4.88 | 1.50 | 0.11 | 47af2fb7 |
| M2 ULTRA | METAL | medium-q8_0 | 1 | 1 | 94.48 | 4.93 | 1.43 | 0.09 | 47af2fb7 |
| M2 ULTRA | METAL | medium-dis | 1 | 1 | 77.85 | 0.80 | 0.20 | 0.01 | 47af2fb7 |
| M2 ULTRA | METAL | large-v2 | 1 | 1 | 170.73 | 7.50 | 2.12 | 0.16 | 47af2fb7 |
| M2 ULTRA | METAL | large-v2-q5_0 | 1 | 1 | 206.46 | 7.05 | 2.17 | 0.20 | 47af2fb7 |
| M2 ULTRA | METAL | large-v2-q5_1 | 1 | 1 | 206.15 | 7.10 | 2.19 | 0.20 | 47af2fb7 |
| M2 ULTRA | METAL | large-v2-q8_0 | 1 | 1 | 180.31 | 6.90 | 2.10 | 0.17 | 47af2fb7 |
| M2 ULTRA | METAL | large-v2-dis | 1 | 1 | 147.44 | 0.90 | 0.22 | 0.02 | 47af2fb7 |
| M2 ULTRA | METAL | large-v3-turbo | 1 | 1 | 148.79 | 1.30 | 0.34 | 0.03 | 47af2fb7 |
| M2 ULTRA | METAL | large-v3-turbo-q5_0 | 1 | 1 | 180.34 | 1.14 | 0.35 | 0.03 | 47af2fb7 |
| M2 ULTRA | METAL | large-v3-turbo-q8_0 | 1 | 1 | 158.04 | 1.18 | 0.33 | 0.03 | 47af2fb7 |
| M2 ULTRA | METAL | tiny | 1 | 1 | 6.06 | 0.96 | 0.22 | 0.01 | f5b477ab |
| M2 ULTRA | METAL | tiny-q5_0 | 1 | 1 | 6.51 | 0.93 | 0.22 | 0.01 | f5b477ab |
| M2 ULTRA | METAL | tiny-q5_1 | 1 | 1 | 6.47 | 0.93 | 0.23 | 0.01 | f5b477ab |
| M2 ULTRA | METAL | tiny-q8_0 | 1 | 1 | 6.16 | 0.94 | 0.21 | 0.01 | f5b477ab |
| M2 ULTRA | METAL | base | 1 | 1 | 10.63 | 1.37 | 0.32 | 0.01 | f5b477ab |
| M2 ULTRA | METAL | base-q5_0 | 1 | 1 | 11.75 | 1.27 | 0.33 | 0.02 | f5b477ab |
| M2 ULTRA | METAL | base-q5_1 | 1 | 1 | 11.73 | 1.25 | 0.33 | 0.02 | f5b477ab |
| M2 ULTRA | METAL | base-q8_0 | 1 | 1 | 11.17 | 1.28 | 0.32 | 0.02 | f5b477ab |
| M2 ULTRA | METAL | small | 1 | 1 | 31.74 | 2.55 | 0.67 | 0.04 | f5b477ab |
| M2 ULTRA | METAL | small-q5_0 | 1 | 1 | 36.21 | 2.47 | 0.69 | 0.04 | f5b477ab |
| M2 ULTRA | METAL | small-q5_1 | 1 | 1 | 36.22 | 2.47 | 0.70 | 0.04 | f5b477ab |
| M2 ULTRA | METAL | small-q8_0 | 1 | 1 | 32.73 | 2.45 | 0.66 | 0.04 | f5b477ab |
| M2 ULTRA | METAL | medium | 1 | 1 | 86.94 | 5.21 | 1.49 | 0.09 | f5b477ab |
| M2 ULTRA | METAL | medium-q5_0 | 1 | 1 | 104.31 | 4.93 | 1.51 | 0.10 | f5b477ab |
| M2 ULTRA | METAL | medium-q5_1 | 1 | 1 | 104.09 | 4.98 | 1.51 | 0.10 | f5b477ab |
| M2 ULTRA | METAL | medium-q8_0 | 1 | 1 | 92.13 | 5.06 | 1.45 | 0.09 | f5b477ab |
| M2 ULTRA | METAL | medium-dis | 1 | 1 | 76.67 | 0.81 | 0.20 | 0.01 | f5b477ab |
| M2 ULTRA | METAL | large-v2 | 1 | 1 | 167.66 | 7.56 | 2.25 | 0.16 | f5b477ab |
| M2 ULTRA | METAL | large-v2-q5_0 | 1 | 1 | 203.09 | 7.13 | 2.29 | 0.20 | f5b477ab |
| M2 ULTRA | METAL | large-v2-q5_1 | 1 | 1 | 202.53 | 7.12 | 2.29 | 0.20 | f5b477ab |
| M2 ULTRA | METAL | large-v2-q8_0 | 1 | 1 | 177.48 | 6.94 | 2.18 | 0.17 | f5b477ab |
| M2 ULTRA | METAL | large-v2-dis | 1 | 1 | 145.61 | 0.91 | 0.23 | 0.02 | f5b477ab |
| M2 ULTRA | METAL | large-v3-turbo | 1 | 1 | 146.95 | 1.33 | 0.36 | 0.03 | f5b477ab |
| M2 ULTRA | METAL | large-v3-turbo-q5_0 | 1 | 1 | 178.57 | 1.17 | 0.36 | 0.03 | f5b477ab |
| M2 ULTRA | METAL | large-v3-turbo-q8_0 | 1 | 1 | 156.19 | 1.21 | 0.34 | 0.03 | f5b477ab |
## M4 Max
@ -268,35 +268,35 @@ make -j && ./scripts/bench-all.sh 1 1 0
| GPU | Config | Model | Th | FA | Enc. | Dec. | Bch5 | PP | Commit |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| RTX 5090 | CUDA | tiny | 1 | 0 | 2.12 | 0.51 | 0.13 | 0.00 | 47af2fb7 |
| RTX 5090 | CUDA | tiny-q8_0 | 1 | 0 | 2.50 | 0.52 | 0.14 | 0.01 | 47af2fb7 |
| RTX 5090 | CUDA | base | 1 | 0 | 3.74 | 0.76 | 0.19 | 0.01 | 47af2fb7 |
| RTX 5090 | CUDA | base-q8_0 | 1 | 0 | 4.38 | 0.74 | 0.20 | 0.01 | 47af2fb7 |
| RTX 5090 | CUDA | small | 1 | 0 | 11.25 | 1.46 | 0.39 | 0.02 | 47af2fb7 |
| RTX 5090 | CUDA | small-q8_0 | 1 | 0 | 12.70 | 1.58 | 0.41 | 0.02 | 47af2fb7 |
| RTX 5090 | CUDA | medium | 1 | 0 | 31.16 | 3.07 | 0.80 | 0.04 | 47af2fb7 |
| RTX 5090 | CUDA | medium-q8_0 | 1 | 0 | 32.50 | 3.23 | 0.83 | 0.05 | 47af2fb7 |
| RTX 5090 | CUDA | large-v2 | 1 | 0 | 50.04 | 4.59 | 1.15 | 0.05 | 47af2fb7 |
| RTX 5090 | CUDA | large-v2-q8_0 | 1 | 0 | 52.17 | 4.38 | 1.14 | 0.07 | 47af2fb7 |
| RTX 5090 | CUDA | large-v3-turbo | 1 | 0 | 46.88 | 0.70 | 0.17 | 0.01 | 47af2fb7 |
| RTX 5090 | CUDA | large-v3-turbo-q8_0 | 1 | 0 | 48.49 | 0.64 | 0.16 | 0.01 | 47af2fb7 |
| RTX 5090 | CUDA | tiny | 1 | 0 | 2.20 | 0.51 | 0.13 | 0.01 | f5b477ab |
| RTX 5090 | CUDA | tiny-q8_0 | 1 | 0 | 2.35 | 0.52 | 0.14 | 0.01 | f5b477ab |
| RTX 5090 | CUDA | base | 1 | 0 | 3.97 | 0.77 | 0.20 | 0.01 | f5b477ab |
| RTX 5090 | CUDA | base-q8_0 | 1 | 0 | 4.20 | 0.73 | 0.20 | 0.01 | f5b477ab |
| RTX 5090 | CUDA | small | 1 | 0 | 11.87 | 1.48 | 0.40 | 0.02 | f5b477ab |
| RTX 5090 | CUDA | small-q8_0 | 1 | 0 | 12.40 | 1.59 | 0.42 | 0.02 | f5b477ab |
| RTX 5090 | CUDA | medium | 1 | 0 | 32.63 | 3.11 | 0.82 | 0.04 | f5b477ab |
| RTX 5090 | CUDA | medium-q8_0 | 1 | 0 | 31.80 | 3.23 | 0.84 | 0.05 | f5b477ab |
| RTX 5090 | CUDA | large-v2 | 1 | 0 | 52.22 | 4.66 | 1.18 | 0.06 | f5b477ab |
| RTX 5090 | CUDA | large-v2-q8_0 | 1 | 0 | 51.11 | 4.37 | 1.15 | 0.07 | f5b477ab |
| RTX 5090 | CUDA | large-v3-turbo | 1 | 0 | 48.72 | 0.70 | 0.18 | 0.01 | f5b477ab |
| RTX 5090 | CUDA | large-v3-turbo-q8_0 | 1 | 0 | 47.81 | 0.64 | 0.16 | 0.01 | f5b477ab |
make -j && ./scripts/bench-all.sh 1 1 1
| GPU | Config | Model | Th | FA | Enc. | Dec. | Bch5 | PP | Commit |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| RTX 5090 | CUDA | tiny | 1 | 1 | 1.42 | 0.44 | 0.11 | 0.00 | 47af2fb7 |
| RTX 5090 | CUDA | tiny-q8_0 | 1 | 1 | 1.83 | 0.45 | 0.12 | 0.01 | 47af2fb7 |
| RTX 5090 | CUDA | base | 1 | 1 | 2.21 | 0.65 | 0.16 | 0.01 | 47af2fb7 |
| RTX 5090 | CUDA | base-q8_0 | 1 | 1 | 2.85 | 0.62 | 0.17 | 0.01 | 47af2fb7 |
| RTX 5090 | CUDA | small | 1 | 1 | 5.11 | 1.23 | 0.32 | 0.01 | 47af2fb7 |
| RTX 5090 | CUDA | small-q8_0 | 1 | 1 | 6.50 | 1.35 | 0.34 | 0.02 | 47af2fb7 |
| RTX 5090 | CUDA | medium | 1 | 1 | 14.01 | 2.57 | 0.64 | 0.03 | 47af2fb7 |
| RTX 5090 | CUDA | medium-q8_0 | 1 | 1 | 15.34 | 2.72 | 0.67 | 0.04 | 47af2fb7 |
| RTX 5090 | CUDA | large-v2 | 1 | 1 | 21.70 | 3.96 | 0.97 | 0.04 | 47af2fb7 |
| RTX 5090 | CUDA | large-v2-q8_0 | 1 | 1 | 23.57 | 3.70 | 0.94 | 0.05 | 47af2fb7 |
| RTX 5090 | CUDA | large-v3-turbo | 1 | 1 | 18.61 | 0.62 | 0.15 | 0.01 | 47af2fb7 |
| RTX 5090 | CUDA | large-v3-turbo-q8_0 | 1 | 1 | 20.10 | 0.56 | 0.14 | 0.01 | 47af2fb7 |
| RTX 5090 | CUDA | tiny | 1 | 1 | 1.37 | 0.44 | 0.11 | 0.00 | f5b477ab |
| RTX 5090 | CUDA | tiny-q8_0 | 1 | 1 | 1.48 | 0.44 | 0.12 | 0.01 | f5b477ab |
| RTX 5090 | CUDA | base | 1 | 1 | 2.34 | 0.66 | 0.16 | 0.01 | f5b477ab |
| RTX 5090 | CUDA | base-q8_0 | 1 | 1 | 2.51 | 0.62 | 0.17 | 0.01 | f5b477ab |
| RTX 5090 | CUDA | small | 1 | 1 | 5.53 | 1.23 | 0.32 | 0.01 | f5b477ab |
| RTX 5090 | CUDA | small-q8_0 | 1 | 1 | 5.88 | 1.35 | 0.33 | 0.02 | f5b477ab |
| RTX 5090 | CUDA | medium | 1 | 1 | 15.09 | 2.55 | 0.65 | 0.03 | f5b477ab |
| RTX 5090 | CUDA | medium-q8_0 | 1 | 1 | 14.06 | 2.72 | 0.67 | 0.03 | f5b477ab |
| RTX 5090 | CUDA | large-v2 | 1 | 1 | 23.24 | 3.94 | 0.97 | 0.04 | f5b477ab |
| RTX 5090 | CUDA | large-v2-q8_0 | 1 | 1 | 22.00 | 3.68 | 0.93 | 0.05 | f5b477ab |
| RTX 5090 | CUDA | large-v3-turbo | 1 | 1 | 19.81 | 0.62 | 0.15 | 0.01 | f5b477ab |
| RTX 5090 | CUDA | large-v3-turbo-q8_0 | 1 | 1 | 18.62 | 0.56 | 0.14 | 0.01 | f5b477ab |
# DGX Spark
@ -305,35 +305,35 @@ make -j && ./scripts/bench-all.sh 1 1 0
| GPU | Config | Model | Th | FA | Enc. | Dec. | Bch5 | PP | Commit |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| DGX Spk. | CUDA | tiny | 1 | 0 | 9.42 | 0.85 | 0.22 | 0.01 | 47af2fb7 |
| DGX Spk. | CUDA | tiny-q8_0 | 1 | 0 | 9.69 | 0.81 | 0.20 | 0.01 | 47af2fb7 |
| DGX Spk. | CUDA | base | 1 | 0 | 18.81 | 1.36 | 0.33 | 0.02 | 47af2fb7 |
| DGX Spk. | CUDA | base-q8_0 | 1 | 0 | 18.11 | 1.20 | 0.30 | 0.02 | 47af2fb7 |
| DGX Spk. | CUDA | small | 1 | 0 | 59.83 | 3.01 | 0.74 | 0.04 | 47af2fb7 |
| DGX Spk. | CUDA | small-q8_0 | 1 | 0 | 59.12 | 2.66 | 0.67 | 0.05 | 47af2fb7 |
| DGX Spk. | CUDA | medium | 1 | 0 | 163.73 | 7.53 | 1.70 | 0.12 | 47af2fb7 |
| DGX Spk. | CUDA | medium-q8_0 | 1 | 0 | 157.54 | 5.98 | 1.48 | 0.13 | 47af2fb7 |
| DGX Spk. | CUDA | large-v2 | 1 | 0 | 279.83 | 12.26 | 2.77 | 0.21 | 47af2fb7 |
| DGX Spk. | CUDA | large-v2-q8_0 | 1 | 0 | 273.05 | 9.31 | 2.33 | 0.22 | 47af2fb7 |
| DGX Spk. | CUDA | large-v3-turbo | 1 | 0 | 271.11 | 2.06 | 0.47 | 0.03 | 47af2fb7 |
| DGX Spk. | CUDA | large-v3-turbo-q8_0 | 1 | 0 | 262.69 | 1.49 | 0.36 | 0.03 | 47af2fb7 |
| DGX Spk. | CUDA | tiny | 1 | 0 | 9.00 | 0.85 | 0.14 | 0.01 | f5b477ab |
| DGX Spk. | CUDA | tiny-q8_0 | 1 | 0 | 8.86 | 0.83 | 0.12 | 0.01 | f5b477ab |
| DGX Spk. | CUDA | base | 1 | 0 | 18.48 | 1.38 | 0.22 | 0.02 | f5b477ab |
| DGX Spk. | CUDA | base-q8_0 | 1 | 0 | 17.28 | 1.22 | 0.19 | 0.02 | f5b477ab |
| DGX Spk. | CUDA | small | 1 | 0 | 56.43 | 3.01 | 0.51 | 0.04 | f5b477ab |
| DGX Spk. | CUDA | small-q8_0 | 1 | 0 | 55.70 | 2.68 | 0.44 | 0.04 | f5b477ab |
| DGX Spk. | CUDA | medium | 1 | 0 | 160.20 | 7.52 | 1.25 | 0.11 | f5b477ab |
| DGX Spk. | CUDA | medium-q8_0 | 1 | 0 | 150.84 | 6.01 | 1.01 | 0.12 | f5b477ab |
| DGX Spk. | CUDA | large-v2 | 1 | 0 | 276.42 | 12.29 | 2.16 | 0.20 | f5b477ab |
| DGX Spk. | CUDA | large-v2-q8_0 | 1 | 0 | 264.92 | 9.32 | 1.67 | 0.20 | f5b477ab |
| DGX Spk. | CUDA | large-v3-turbo | 1 | 0 | 264.90 | 2.03 | 0.37 | 0.03 | f5b477ab |
| DGX Spk. | CUDA | large-v3-turbo-q8_0 | 1 | 0 | 253.56 | 1.48 | 0.27 | 0.03 | f5b477ab |
make -j && ./scripts/bench-all.sh 1 1 1
| GPU | Config | Model | Th | FA | Enc. | Dec. | Bch5 | PP | Commit |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| DGX Spk. | CUDA | tiny | 1 | 1 | 2.89 | 0.76 | 0.19 | 0.01 | 47af2fb7 |
| DGX Spk. | CUDA | tiny-q8_0 | 1 | 1 | 3.06 | 0.72 | 0.17 | 0.01 | 47af2fb7 |
| DGX Spk. | CUDA | base | 1 | 1 | 5.37 | 1.23 | 0.29 | 0.01 | 47af2fb7 |
| DGX Spk. | CUDA | base-q8_0 | 1 | 1 | 4.70 | 1.07 | 0.26 | 0.01 | 47af2fb7 |
| DGX Spk. | CUDA | small | 1 | 1 | 17.70 | 2.73 | 0.66 | 0.02 | 47af2fb7 |
| DGX Spk. | CUDA | small-q8_0 | 1 | 1 | 16.77 | 2.38 | 0.58 | 0.03 | 47af2fb7 |
| DGX Spk. | CUDA | medium | 1 | 1 | 56.22 | 6.98 | 1.53 | 0.06 | 47af2fb7 |
| DGX Spk. | CUDA | medium-q8_0 | 1 | 1 | 46.39 | 5.46 | 1.28 | 0.07 | 47af2fb7 |
| DGX Spk. | CUDA | large-v2 | 1 | 1 | 100.33 | 11.59 | 2.53 | 0.09 | 47af2fb7 |
| DGX Spk. | CUDA | large-v2-q8_0 | 1 | 1 | 97.28 | 8.60 | 2.10 | 0.10 | 47af2fb7 |
| DGX Spk. | CUDA | large-v3-turbo | 1 | 1 | 92.59 | 2.00 | 0.44 | 0.02 | 47af2fb7 |
| DGX Spk. | CUDA | large-v3-turbo-q8_0 | 1 | 1 | 85.96 | 1.40 | 0.33 | 0.02 | 47af2fb7 |
| DGX Spk. | CUDA | tiny | 1 | 1 | 2.63 | 0.76 | 0.13 | 0.01 | f5b477ab |
| DGX Spk. | CUDA | tiny-q8_0 | 1 | 1 | 2.46 | 0.73 | 0.11 | 0.01 | f5b477ab |
| DGX Spk. | CUDA | base | 1 | 1 | 4.96 | 1.24 | 0.20 | 0.01 | f5b477ab |
| DGX Spk. | CUDA | base-q8_0 | 1 | 1 | 4.23 | 1.08 | 0.17 | 0.01 | f5b477ab |
| DGX Spk. | CUDA | small | 1 | 1 | 16.26 | 2.73 | 0.47 | 0.02 | f5b477ab |
| DGX Spk. | CUDA | small-q8_0 | 1 | 1 | 14.94 | 2.38 | 0.39 | 0.02 | f5b477ab |
| DGX Spk. | CUDA | medium | 1 | 1 | 51.81 | 6.94 | 1.22 | 0.05 | f5b477ab |
| DGX Spk. | CUDA | medium-q8_0 | 1 | 1 | 41.51 | 5.44 | 0.93 | 0.05 | f5b477ab |
| DGX Spk. | CUDA | large-v2 | 1 | 1 | 98.54 | 11.53 | 2.05 | 0.08 | f5b477ab |
| DGX Spk. | CUDA | large-v2-q8_0 | 1 | 1 | 91.61 | 8.49 | 1.55 | 0.08 | f5b477ab |
| DGX Spk. | CUDA | large-v3-turbo | 1 | 1 | 87.20 | 1.94 | 0.36 | 0.02 | f5b477ab |
| DGX Spk. | CUDA | large-v3-turbo-q8_0 | 1 | 1 | 80.28 | 1.38 | 0.26 | 0.01 | f5b477ab |
# V100

10
scripts/bench-all.sh
View File

@ -100,12 +100,14 @@ for model in "${models[@]}"; do
if [[ $system_info == *"CUDA = 1"* ]]; then
config="$config CUDA"
elif [[ $system_info == *"CUDA : ARCHS"* ]]; then
config="$config CUDA"
fi
if [[ $system_info == *"METAL = 1"* ]]; then
config="$config METAL"
elif [[ $system_info == *"Metal : EMBED_LIBRARY = 1"* ]]; then
config="$config METAL"
if [[ $system_info == *"MTL = 1"* ]]; then
config="$config MTL"
elif [[ $system_info == *"MTL : EMBED_LIBRARY = 1"* ]]; then
config="$config MTL"
fi
commit=$(git rev-parse --short HEAD)