Performance¶
Below are measured performance for the Riva ASR, NLP, and TTS services on NVIDIA T4, V100 SXM2 16 GB, and NVIDIA A100 SXM4 40 GB GPUs. CPU specifications for each system can be found here:
ASR¶
The latency numbers below were measured using the streaming recognition mode, with the
BERT-based punctuation model enabled, a 4-gram language model, a decoder beam width of
128, and timestamps enabled. The Jasper, QuartzNet and CitriNet-1024 acoustic models were tested. The client and the
server used audio chunks of the same duration (100ms, 160ms, 800ms, 3200ms depending on
the server configuration). The Riva streaming client riva_streaming_asr_client
,
provided in the Riva client image, was used with the --simulate_realtime
flag to
simulate transcription from a microphone, where each stream was doing 5 iterations
over a sample audio file from the Librispeech dataset (1272-135031-0000.wav).
The command used was:
riva_streaming_asr_client \
--chunk_duration_ms=<chunk_duration> --simulate_realtime=true \
--automatic_punctuation=true --num_parallel_requests=<num_streams> \
--word_time_offsets=true --print_transcripts=false \
--interim_results=false --num_iterations=<5*num_streams> \
--audio_file=1272-135031-0000.wav --output_filename=/tmp/output.json
The riva_streaming_asr_client
returns latency measured in three different ways
after executing the benchmark task:
intermediate latency
: latency to return an intermediate transcript withis_final == false
final latency
: latency of messages return withis_final == true
latency
: the overall latency of all returned message types
The overall latency numbers are reported below.
NVIDIA A100 GPU¶
Streaming, low-latency¶
Acoustic model |
# of streams |
Chunk size (ms) |
Latency (ms) |
Throughput (RTFX) |
||||
---|---|---|---|---|---|---|---|---|
avg |
p50 |
p90 |
p95 |
p99 |
||||
quartznet |
1 |
100 |
13.775 |
13.088 |
17.046 |
20.146 |
20.146 |
0.99966 |
quartznet |
8 |
100 |
22.99 |
22.786 |
28.174 |
30.111 |
31.194 |
7.9956 |
quartznet |
16 |
100 |
35.618 |
35.717 |
43.401 |
43.649 |
45.561 |
15.985 |
quartznet |
32 |
100 |
50.736 |
52.46 |
57.71 |
59.6 |
60.75 |
31.961 |
quartznet |
48 |
100 |
64.695 |
65.642 |
76.823 |
78.963 |
84.639 |
47.92 |
quartznet |
64 |
100 |
75.197 |
77.22 |
91.216 |
92.174 |
96.638 |
63.876 |
quartznet |
96 |
100 |
107.64 |
106.95 |
149.25 |
156.29 |
163.78 |
95.767 |
quartznet |
128 |
100 |
124.94 |
131.37 |
165.57 |
168.86 |
172.62 |
127.63 |
jasper |
1 |
100 |
19.223 |
18.529 |
26.688 |
26.688 |
26.688 |
0.99963 |
jasper |
8 |
100 |
30.539 |
30.664 |
34.128 |
40.741 |
44.225 |
7.995 |
jasper |
16 |
100 |
48.232 |
48.636 |
53.826 |
56.712 |
57.008 |
15.984 |
jasper |
32 |
100 |
65.746 |
66.766 |
73.168 |
73.828 |
75.138 |
31.958 |
jasper |
48 |
100 |
80.328 |
82.805 |
93.649 |
94.126 |
97.669 |
47.92 |
jasper |
64 |
100 |
106.6 |
93.019 |
202.1 |
277.21 |
285.15 |
63.881 |
jasper |
96 |
100 |
148.85 |
167.47 |
182.81 |
184.51 |
187.58 |
95.739 |
jasper |
128 |
100 |
200.59 |
197.58 |
256.19 |
278.65 |
291.72 |
127.52 |
citrinet |
1 |
160 |
17.677 |
16.331 |
23.152 |
23.152 |
23.152 |
0.99967 |
citrinet |
8 |
160 |
34.511 |
34.28 |
38.447 |
41.862 |
41.922 |
7.9944 |
citrinet |
16 |
160 |
60.983 |
62.667 |
69.104 |
71.966 |
72.043 |
15.98 |
citrinet |
32 |
160 |
106.95 |
107.01 |
111.47 |
111.86 |
115.34 |
31.932 |
citrinet |
48 |
160 |
137.46 |
139.69 |
145.42 |
146.38 |
148.59 |
47.867 |
citrinet |
64 |
160 |
159.28 |
157.12 |
180.18 |
191.43 |
192.25 |
63.79 |
citrinet |
96 |
160 |
235.08 |
242.71 |
261.76 |
266.1 |
271.56 |
95.599 |
citrinet |
128 |
160 |
361.5 |
340.03 |
480.78 |
490.2 |
511 |
127.21 |
Streaming, high-throughput¶
Acoustic Model |
# of streams |
Chunk size (ms) |
Latency (ms) |
Throughput (RTFX) |
||||
---|---|---|---|---|---|---|---|---|
avg |
p50 |
p90 |
p95 |
p99 |
||||
quartznet |
1 |
800 |
15.743 |
14.814 |
25.57 |
25.57 |
25.57 |
0.99968 |
quartznet |
64 |
800 |
81.046 |
86.341 |
97.161 |
98.439 |
101.32 |
63.886 |
quartznet |
128 |
800 |
126.47 |
130.05 |
155.72 |
160.57 |
166.99 |
127.65 |
quartznet |
256 |
800 |
197.65 |
206.16 |
265.12 |
282.63 |
314.48 |
254.71 |
quartznet |
384 |
800 |
269.71 |
282.78 |
360.83 |
374.16 |
390.11 |
381.3 |
quartznet |
512 |
800 |
282.5 |
302.8 |
386.91 |
418.78 |
448.37 |
508.07 |
quartznet |
768 |
800 |
402.32 |
413.01 |
576.61 |
600.85 |
649.58 |
760.31 |
quartznet |
1024 |
800 |
514.61 |
510.91 |
718.04 |
773.47 |
872.36 |
1009.9 |
jasper |
1 |
800 |
21.802 |
21.758 |
31.508 |
31.508 |
31.508 |
0.99963 |
jasper |
64 |
800 |
108.88 |
109.01 |
137.25 |
138.69 |
142.34 |
63.864 |
jasper |
128 |
800 |
159.01 |
158.04 |
198.9 |
219.88 |
233.1 |
127.56 |
jasper |
256 |
800 |
224.17 |
231.63 |
287.34 |
296.27 |
304.76 |
254.64 |
jasper |
384 |
800 |
317.98 |
322.49 |
418.42 |
437.52 |
455.75 |
381.08 |
jasper |
512 |
800 |
397.81 |
402.84 |
547.15 |
581.55 |
637.14 |
507.09 |
jasper |
768 |
800 |
567.77 |
534.4 |
757.2 |
900.83 |
1100.6 |
757.87 |
jasper |
1024 |
800 |
1000.6 |
986.46 |
1388.4 |
1423.5 |
1512.9 |
1004.9 |
citrinet |
1 |
800 |
16.451 |
13.965 |
26.439 |
26.439 |
26.439 |
0.99971 |
citrinet |
64 |
800 |
111.57 |
112.32 |
132.88 |
139.58 |
142.24 |
63.819 |
citrinet |
128 |
800 |
170.44 |
180.26 |
202 |
209.28 |
212.95 |
127.49 |
citrinet |
256 |
800 |
250.8 |
251.68 |
311.61 |
324.22 |
332.85 |
254.29 |
citrinet |
384 |
800 |
347.34 |
371.71 |
429.55 |
437.91 |
448.47 |
380.63 |
citrinet |
512 |
800 |
426.17 |
436.69 |
531.7 |
554.93 |
578.55 |
506.7 |
citrinet |
768 |
800 |
912.08 |
802.33 |
1337.2 |
1355.9 |
1392.5 |
755.45 |
citrinet |
1024 |
800 |
1561.7 |
1519 |
2266.9 |
2334.7 |
2655.5 |
998.55 |
Offline¶
Acoustic model |
# of streams |
Chunk size (ms) |
Latency (ms) |
Throughput (RTFX) |
||||
---|---|---|---|---|---|---|---|---|
avg |
p50 |
p90 |
p95 |
p99 |
||||
quartznet |
1 |
1600 |
26.22 |
26.387 |
30.733 |
37.175 |
37.175 |
0.99957 |
quartznet |
256 |
1600 |
310.96 |
313.52 |
446.33 |
466.73 |
534.91 |
254.35 |
quartznet |
512 |
1600 |
461.33 |
479.58 |
632.47 |
684.36 |
766.31 |
506.68 |
quartznet |
768 |
1600 |
595.36 |
617.58 |
838.34 |
869.68 |
912.17 |
757.45 |
quartznet |
1024 |
1600 |
752.81 |
789.86 |
1053.1 |
1119.5 |
1212.9 |
1005.2 |
quartznet |
1280 |
1600 |
951.09 |
931.12 |
1317.6 |
1385 |
1954.4 |
1251.6 |
quartznet |
1512 |
1600 |
1189.7 |
1156.2 |
1801.4 |
1984.7 |
2323.6 |
1470.5 |
jasper |
1 |
1600 |
28.477 |
29.909 |
31.968 |
41.738 |
41.738 |
0.99956 |
jasper |
256 |
1600 |
390.12 |
385.22 |
507.51 |
556.41 |
592.78 |
253.94 |
jasper |
512 |
1600 |
617.12 |
632.95 |
804.37 |
822.53 |
887 |
505.1 |
jasper |
768 |
1600 |
840.32 |
841.29 |
1079.1 |
1112 |
1143.8 |
754.65 |
jasper |
1024 |
1600 |
1107.4 |
1073.3 |
1425 |
1723.7 |
2098.4 |
998.81 |
jasper |
1280 |
1600 |
1396 |
1344.9 |
2223.4 |
2288.1 |
2453.6 |
1241.4 |
jasper |
1512 |
1600 |
1814.3 |
1791.7 |
2695.5 |
2767.7 |
2861.2 |
1452.8 |
citrinet |
1 |
1600 |
16.525 |
14.823 |
23.402 |
23.402 |
23.402 |
0.9997 |
citrinet |
256 |
1600 |
284.78 |
311.47 |
345.06 |
358.33 |
376.56 |
254.26 |
citrinet |
512 |
1600 |
477.13 |
484.25 |
597.09 |
609.65 |
630.33 |
505.52 |
citrinet |
768 |
1600 |
649.29 |
667.79 |
825.69 |
843.51 |
869.35 |
755.88 |
citrinet |
1024 |
1600 |
827.25 |
871.89 |
1006.1 |
1062.2 |
1108.9 |
1002.9 |
citrinet |
1280 |
1600 |
993.9 |
1067.4 |
1231.1 |
1263.1 |
1288 |
1247.8 |
citrinet |
1512 |
1600 |
1217.5 |
1214.3 |
1553.1 |
1924.1 |
2475.2 |
1468.2 |
NVIDIA V100 GPU¶
Streaming, low-latency¶
Acoustic model |
# of streams |
Chunk size (ms) |
Latency (ms) |
Throughput (RTFX) |
||||
---|---|---|---|---|---|---|---|---|
avg |
p50 |
p90 |
p95 |
p99 |
||||
quartznet |
1 |
100 |
16.219 |
15.743 |
18.59 |
18.991 |
18.991 |
0.99962 |
quartznet |
8 |
100 |
28.638 |
27.698 |
40.32 |
41.184 |
43.96 |
7.9937 |
quartznet |
16 |
100 |
37.713 |
38.838 |
48.903 |
49.144 |
52.002 |
15.985 |
quartznet |
32 |
100 |
58.434 |
58.619 |
72.506 |
73.37 |
76.746 |
31.95 |
quartznet |
48 |
100 |
72.319 |
74.779 |
91.255 |
92.098 |
95.413 |
47.911 |
quartznet |
64 |
100 |
89.535 |
91.863 |
110.78 |
121.52 |
137.01 |
63.861 |
quartznet |
96 |
100 |
148 |
163.32 |
192.74 |
197.29 |
206.67 |
95.697 |
quartznet |
128 |
100 |
197.11 |
202.9 |
276.54 |
294.76 |
324.81 |
127.53 |
jasper |
1 |
100 |
29.375 |
28.314 |
37.363 |
37.363 |
37.363 |
0.99945 |
jasper |
8 |
100 |
41.848 |
42.036 |
55.031 |
61.506 |
61.754 |
7.993 |
jasper |
16 |
100 |
61.287 |
62.791 |
68.992 |
70.869 |
73.396 |
15.982 |
jasper |
32 |
100 |
106.26 |
98.485 |
156.7 |
157.2 |
159.56 |
31.931 |
jasper |
48 |
100 |
157.06 |
179.89 |
192.93 |
198.2 |
264.84 |
47.875 |
jasper |
64 |
100 |
319.66 |
304.95 |
396.29 |
400.8 |
478.1 |
63.649 |
jasper |
96 |
100 |
19208 |
20795 |
28822 |
28881 |
29014 |
79.351 |
jasper |
128 |
100 |
46510 |
49183 |
67122 |
67396 |
67557 |
66.463 |
citrinet |
1 |
160 |
19.539 |
18.898 |
22.244 |
22.244 |
22.244 |
0.99961 |
citrinet |
8 |
160 |
44.174 |
44.99 |
48.045 |
51.79 |
51.937 |
7.9926 |
citrinet |
16 |
160 |
82.064 |
75.79 |
143.32 |
148.02 |
148.08 |
15.976 |
citrinet |
32 |
160 |
130.32 |
128.85 |
140.86 |
142.08 |
145.73 |
31.919 |
citrinet |
48 |
160 |
175.12 |
175.14 |
188.77 |
194.13 |
234.44 |
47.826 |
citrinet |
64 |
160 |
250.29 |
264.11 |
280.87 |
285.19 |
285.73 |
63.716 |
citrinet |
96 |
160 |
408.13 |
417.58 |
582.75 |
614.63 |
635.04 |
95.402 |
citrinet |
128 |
160 |
723.41 |
783.02 |
1008.4 |
1015.8 |
1044.6 |
126.97 |
Streaming, high-throughput¶
Acoustic model |
# of streams |
Chunk size (ms) |
Latency (ms) |
Throughput (RTFX) |
||||
---|---|---|---|---|---|---|---|---|
avg |
p50 |
p90 |
p95 |
p99 |
||||
quartznet |
1 |
800 |
18.69 |
19.171 |
21.409 |
21.409 |
21.409 |
0.99962 |
quartznet |
64 |
800 |
107.39 |
112.99 |
133.36 |
135.93 |
137.04 |
63.828 |
quartznet |
128 |
800 |
155.8 |
162.67 |
202.49 |
212.42 |
225.76 |
127.55 |
quartznet |
256 |
800 |
261.27 |
271.26 |
339.65 |
359.18 |
373.54 |
254.35 |
quartznet |
384 |
800 |
345.09 |
355.63 |
448.95 |
465.11 |
506.95 |
380.59 |
quartznet |
512 |
800 |
427.9 |
431.01 |
564.88 |
592.34 |
641.33 |
506.45 |
quartznet |
768 |
800 |
682.67 |
679.89 |
932.53 |
978.61 |
1161.5 |
752.91 |
quartznet |
1024 |
800 |
1333.1 |
1380.9 |
1841.3 |
1933.2 |
2108.6 |
998.68 |
jasper |
1 |
800 |
31.131 |
33.332 |
37.633 |
37.633 |
37.633 |
0.99942 |
jasper |
64 |
800 |
160.4 |
158.69 |
193.9 |
205.84 |
215.41 |
63.803 |
jasper |
128 |
800 |
251.48 |
254.63 |
310.37 |
316.14 |
323.98 |
127.29 |
jasper |
256 |
800 |
429.14 |
426.38 |
519.52 |
533.69 |
559.53 |
253.75 |
jasper |
384 |
800 |
619.68 |
625.24 |
726.98 |
753.06 |
991.31 |
378.75 |
jasper |
512 |
800 |
1202.3 |
1405.2 |
1534.3 |
1562.1 |
1904.4 |
501.98 |
jasper |
768 |
800 |
13400 |
13849 |
20750 |
21581 |
21774 |
608.81 |
jasper |
1024 |
800 |
48565 |
54184 |
70625 |
72035 |
72329 |
465.68 |
citrinet |
1 |
800 |
17.44 |
17.189 |
19.707 |
19.707 |
19.707 |
0.99965 |
citrinet |
64 |
800 |
125.23 |
122.22 |
150.43 |
153.12 |
156.7 |
63.813 |
citrinet |
128 |
800 |
235.35 |
244.35 |
296.1 |
339.38 |
349.17 |
127.33 |
citrinet |
256 |
800 |
391.62 |
399.09 |
472.78 |
485.66 |
504.49 |
253.44 |
citrinet |
384 |
800 |
557.85 |
556.18 |
657.04 |
678.23 |
712.09 |
378.86 |
citrinet |
512 |
800 |
1224 |
1381.5 |
1533.5 |
1569.2 |
1913.1 |
501.52 |
citrinet |
768 |
800 |
5063.1 |
5236.3 |
6606.9 |
7179.4 |
7709.1 |
688.89 |
citrinet |
1024 |
800 |
19368 |
19916 |
30321 |
30436 |
30529 |
654.76 |
Offline¶
Acoustic model |
# of streams |
Chunk size (ms) |
Latency (ms) |
Throughput (RTFX) |
||||
---|---|---|---|---|---|---|---|---|
avg |
p50 |
p90 |
p95 |
p99 |
||||
quartznet |
1 |
1600 |
32.051 |
33.137 |
34.889 |
38 |
38 |
0.99944 |
quartznet |
256 |
1600 |
418.92 |
436.08 |
559.46 |
577.89 |
601.12 |
253.64 |
quartznet |
512 |
1600 |
719.14 |
736.58 |
999.11 |
1039.9 |
1130.8 |
503.17 |
quartznet |
768 |
1600 |
988.4 |
950.99 |
1420.8 |
1513 |
1617 |
749.3 |
quartznet |
1024 |
1600 |
1413.4 |
1466 |
1965.9 |
2235.4 |
2493.9 |
989.17 |
quartznet |
1280 |
1600 |
1871.1 |
2000.3 |
2644.3 |
2982.2 |
3118.8 |
1220.6 |
quartznet |
1512 |
1600 |
2206.1 |
2335.8 |
3166 |
3447.9 |
3695.4 |
1422.5 |
jasper |
1 |
1600 |
39.439 |
41.926 |
43.788 |
43.935 |
43.935 |
0.99935 |
jasper |
256 |
1600 |
734.73 |
738.24 |
884.44 |
927.17 |
987.07 |
252.21 |
jasper |
512 |
1600 |
1475.3 |
1378.1 |
2372.8 |
2431.2 |
2501.6 |
498 |
jasper |
768 |
1600 |
2427.6 |
2452.2 |
3414.7 |
3473.7 |
3575.9 |
729.88 |
jasper |
1024 |
1600 |
3954.7 |
4184.9 |
5899.5 |
6005.3 |
6765.2 |
953.3 |
jasper |
1280 |
1600 |
17678 |
17212 |
30763 |
35474 |
36963 |
996.97 |
jasper |
1512 |
1600 |
26000 |
26015 |
43323 |
45986 |
48253 |
988.53 |
citrinet |
1 |
1600 |
18.472 |
17.615 |
20.458 |
20.458 |
20.458 |
0.99963 |
citrinet |
256 |
1600 |
451.06 |
478.22 |
527.48 |
532.38 |
539.79 |
253.26 |
citrinet |
512 |
1600 |
807.31 |
866.29 |
984.91 |
1006.8 |
1030.9 |
501.78 |
citrinet |
768 |
1600 |
1139.2 |
1217.3 |
1390.5 |
1430.3 |
1477.9 |
746.5 |
citrinet |
1024 |
1600 |
2112.2 |
1914.5 |
3069 |
3186.8 |
3491.2 |
983.31 |
citrinet |
1280 |
1600 |
6875.1 |
7291.8 |
9279.1 |
9948.4 |
10842 |
1115.3 |
citrinet |
1512 |
1600 |
13901 |
13860 |
19986 |
20129 |
20849 |
1099.6 |
NVIDIA A30¶
Streaming, low-latency¶
Acoustic model |
# of streams |
chunk size (ms) |
Latency (ms) |
Throughput (RTFX) |
||||
---|---|---|---|---|---|---|---|---|
avg |
p50 |
p90 |
p95 |
p99 |
||||
quartznet |
1 |
100 |
23.012 |
21.484 |
26.527 |
35.573 |
35.573 |
0.99947 |
quartznet |
8 |
100 |
40.112 |
39.849 |
52.976 |
54.077 |
55.768 |
7.9913 |
quartznet |
16 |
100 |
53.306 |
55.101 |
63.018 |
63.96 |
66.982 |
15.979 |
quartznet |
32 |
100 |
74.77 |
77.611 |
86.416 |
90.268 |
91.86 |
31.944 |
quartznet |
48 |
100 |
92.743 |
95.436 |
111.61 |
116.51 |
124.5 |
47.895 |
quartznet |
64 |
100 |
113.18 |
115.12 |
145.84 |
154.07 |
174.84 |
63.838 |
quartznet |
96 |
100 |
141.52 |
148.03 |
176.71 |
180.35 |
195.07 |
95.706 |
quartznet |
128 |
100 |
164.22 |
178.01 |
209.67 |
221.58 |
241.87 |
127.5 |
jasper |
1 |
100 |
26.798 |
25.557 |
36.703 |
36.703 |
36.703 |
0.99947 |
jasper |
8 |
100 |
48.764 |
50.231 |
56.828 |
64.799 |
65.521 |
7.9916 |
jasper |
16 |
100 |
75.89 |
77.099 |
85.157 |
85.514 |
86.154 |
15.975 |
jasper |
32 |
100 |
89.519 |
87.85 |
99.081 |
99.469 |
118.22 |
31.942 |
jasper |
48 |
100 |
116.23 |
105.88 |
161.76 |
164.72 |
176.58 |
47.894 |
jasper |
64 |
100 |
160.6 |
180.7 |
196.33 |
201.24 |
217.37 |
63.829 |
jasper |
96 |
100 |
3475.8 |
3817.2 |
5894.9 |
5929.5 |
5994.8 |
93.12 |
jasper |
128 |
100 |
38441 |
44791 |
56350 |
56433 |
56467 |
73.177 |
citrinet |
1 |
160 |
27.388 |
25.202 |
36.378 |
36.378 |
36.378 |
0.9995 |
citrinet |
8 |
160 |
60.85 |
63.105 |
68.724 |
70.739 |
70.963 |
7.9893 |
citrinet |
16 |
160 |
99.186 |
102.57 |
106.95 |
108.31 |
111.2 |
15.967 |
citrinet |
32 |
160 |
150 |
152.07 |
159.92 |
160.37 |
164.37 |
31.898 |
citrinet |
48 |
160 |
177.62 |
176.17 |
216.78 |
218.33 |
230.31 |
47.827 |
citrinet |
64 |
160 |
231.88 |
246.42 |
260.82 |
264.49 |
270.19 |
63.722 |
citrinet |
96 |
160 |
385.64 |
412.34 |
477.62 |
486.61 |
500.4 |
95.411 |
citrinet |
128 |
160 |
615.73 |
673.92 |
838.69 |
862.88 |
900.76 |
127 |
Streaming, high-throughput¶
Acoustic model |
# of streams |
Chunk size (ms) |
Latency (ms) |
Throughput (RTFX) |
||||
---|---|---|---|---|---|---|---|---|
avg |
p50 |
p90 |
p95 |
p99 |
||||
quartznet |
1 |
800 |
26.547 |
25.805 |
37.417 |
37.417 |
37.417 |
0.9995 |
quartznet |
64 |
800 |
112.45 |
115.53 |
132.46 |
136.06 |
140.55 |
63.832 |
quartznet |
128 |
800 |
175.8 |
182.18 |
210.73 |
223.33 |
245.22 |
127.45 |
quartznet |
256 |
800 |
230.93 |
238.92 |
295.34 |
306.26 |
326.54 |
254.55 |
quartznet |
384 |
800 |
292.9 |
301.39 |
379.25 |
390.45 |
408.39 |
381.06 |
quartznet |
512 |
800 |
347.17 |
355.37 |
449.16 |
460.51 |
478.91 |
507.15 |
quartznet |
768 |
800 |
490.32 |
498.81 |
634.44 |
652.14 |
812.12 |
757.79 |
quartznet |
1024 |
800 |
831.5 |
781.33 |
1236.7 |
1274.9 |
1332 |
1005.3 |
jasper |
1 |
800 |
32.656 |
33.29 |
53.944 |
53.944 |
53.944 |
0.99947 |
jasper |
64 |
800 |
147.18 |
147.45 |
172.74 |
187.07 |
195.84 |
63.789 |
jasper |
128 |
800 |
211.23 |
209.42 |
258.8 |
288.3 |
305.71 |
127.43 |
jasper |
256 |
800 |
330 |
331.1 |
404.37 |
417.7 |
437.26 |
253.96 |
jasper |
384 |
800 |
448.81 |
450.9 |
542.06 |
554.14 |
580.03 |
379.93 |
jasper |
512 |
800 |
585.83 |
589.64 |
706.41 |
728.01 |
860.37 |
504.74 |
jasper |
768 |
800 |
1904.1 |
1875.4 |
2894.1 |
3055.6 |
3211.4 |
750.04 |
jasper |
1024 |
800 |
25507 |
35165 |
42603 |
44092 |
44443 |
601.03 |
citrinet |
1 |
800 |
23.94 |
21.378 |
34.312 |
34.312 |
34.312 |
0.99955 |
citrinet |
64 |
800 |
142.53 |
147.9 |
169.59 |
174.97 |
178.62 |
63.804 |
citrinet |
128 |
800 |
199.64 |
212.55 |
240.21 |
243.44 |
247.29 |
127.37 |
citrinet |
256 |
800 |
343.58 |
360.52 |
405.54 |
410.63 |
414.54 |
253.81 |
citrinet |
384 |
800 |
481.4 |
495.56 |
559.23 |
571.46 |
587.83 |
379.51 |
citrinet |
512 |
800 |
847.78 |
763.69 |
1280.3 |
1358.6 |
1378.5 |
503.55 |
citrinet |
768 |
800 |
2214.9 |
2090.3 |
3187.2 |
3395.2 |
3617.1 |
742.31 |
citrinet |
1024 |
800 |
14236 |
14517 |
22074 |
22161 |
22216 |
725.37 |
Offline¶
Acoustic model |
# of streams |
Chunk size (ms) |
Latency (ms) |
Throughput (RTFX) |
||||
---|---|---|---|---|---|---|---|---|
avg |
p50 |
p90 |
p95 |
p99 |
||||
quartznet |
1 |
1600 |
44.49 |
45.29 |
46.942 |
60.447 |
60.447 |
0.99925 |
quartznet |
256 |
1600 |
436.26 |
434.66 |
605.55 |
640.83 |
700.65 |
253.88 |
quartznet |
512 |
1600 |
660.24 |
635.61 |
966.28 |
1003.4 |
1054.1 |
505.12 |
quartznet |
768 |
1600 |
844.37 |
818.38 |
1239.9 |
1305.6 |
1418.8 |
753.56 |
quartznet |
1024 |
1600 |
1128.2 |
1087.3 |
1642.5 |
1784 |
2041.5 |
998.23 |
quartznet |
1280 |
1600 |
1428.9 |
1404.9 |
2032.5 |
2237.9 |
2363.1 |
1238.6 |
quartznet |
1512 |
1600 |
1785.9 |
1888.8 |
2555.6 |
2677.8 |
2810.1 |
1448.5 |
jasper |
1 |
1600 |
42.569 |
44.758 |
47.754 |
57.853 |
57.853 |
0.99932 |
jasper |
256 |
1600 |
588.18 |
581.99 |
708.42 |
728.89 |
759.12 |
253.13 |
jasper |
512 |
1600 |
1023.7 |
1033.8 |
1204.5 |
1236.5 |
1383.1 |
501.23 |
jasper |
768 |
1600 |
1759.2 |
1663.7 |
2649.2 |
2708.3 |
2752.6 |
741.65 |
jasper |
1024 |
1600 |
2446.9 |
2551.6 |
3506.5 |
3592.4 |
3712.4 |
971.2 |
jasper |
1280 |
1600 |
3700.4 |
3693.1 |
5732.5 |
5804.8 |
6500.5 |
1192.7 |
jasper |
1512 |
1600 |
12969 |
10839 |
24484 |
27178 |
30380 |
1288.9 |
citrinet |
1 |
1600 |
24.544 |
22.226 |
33.908 |
33.908 |
33.908 |
0.99954 |
citrinet |
256 |
1600 |
377.66 |
387.39 |
455.11 |
464.64 |
477.74 |
253.63 |
citrinet |
512 |
1600 |
714.22 |
732.03 |
823.01 |
837.19 |
850.16 |
503.68 |
citrinet |
768 |
1600 |
978.02 |
1051.6 |
1162.8 |
1178.3 |
1196.3 |
749.71 |
citrinet |
1024 |
1600 |
1408.8 |
1395.2 |
1800.7 |
2050.7 |
2365.8 |
992.91 |
citrinet |
1280 |
1600 |
3271.2 |
3044.7 |
5138.4 |
5403.9 |
5662.1 |
1223.4 |
citrinet |
1512 |
1600 |
9137.7 |
9228.6 |
13005 |
13105 |
13319 |
1212.5 |
NVIDIA T4¶
Streaming, low-latency¶
Acoustic model |
# of streams |
chunk size (ms) |
Latency (ms) |
Throughput (RTFX) |
||||
---|---|---|---|---|---|---|---|---|
avg |
p50 |
p90 |
p95 |
p99 |
||||
quartznet |
1 |
100 |
34.801 |
33.282 |
41.633 |
43.414 |
43.414 |
0.99922 |
quartznet |
8 |
100 |
59.16 |
58.793 |
72.267 |
79.687 |
81.207 |
7.9875 |
quartznet |
16 |
100 |
99.245 |
94.272 |
138.86 |
143.52 |
177.08 |
15.957 |
quartznet |
32 |
100 |
114.1 |
113.33 |
162.18 |
169.75 |
178.31 |
31.914 |
quartznet |
48 |
100 |
156.18 |
157.62 |
208.25 |
220.52 |
238.8 |
47.844 |
quartznet |
64 |
100 |
212.69 |
222.17 |
275.92 |
291.22 |
306.31 |
63.732 |
quartznet |
96 |
100 |
3850.5 |
4131.3 |
5927.7 |
6124.6 |
6234.4 |
87.748 |
quartznet |
128 |
100 |
14242 |
14185 |
24099 |
25082 |
25135 |
87.514 |
jasper |
1 |
100 |
63.836 |
64.246 |
72.547 |
72.547 |
72.547 |
0.99877 |
jasper |
8 |
100 |
88.517 |
90.074 |
101.94 |
135.44 |
135.95 |
7.9849 |
jasper |
16 |
100 |
141.71 |
146.86 |
181.78 |
190.8 |
191.21 |
15.958 |
jasper |
32 |
100 |
1550.5 |
1613.8 |
2251.7 |
2266.3 |
2320.6 |
31.771 |
jasper |
48 |
100 |
24754 |
27473 |
37153 |
37214 |
37601 |
36.12 |
jasper |
64 |
100 |
38824 |
46842 |
56015 |
56042 |
56223 |
37.552 |
jasper |
96 |
100 |
54981 |
56534 |
67697 |
69204 |
72110 |
38.728 |
jasper |
128 |
100 |
78679 |
85508 |
99784 |
101070 |
103300 |
39.162 |
citrinet |
1 |
160 |
46.805 |
45.001 |
54.768 |
54.768 |
54.768 |
0.99913 |
citrinet |
8 |
160 |
113.42 |
111.86 |
122.79 |
122.88 |
122.92 |
7.9819 |
citrinet |
16 |
160 |
144.7 |
142.91 |
164.55 |
168.15 |
194.16 |
15.96 |
citrinet |
32 |
160 |
270 |
286.08 |
299.6 |
299.85 |
305.03 |
31.852 |
citrinet |
48 |
160 |
448.73 |
470.56 |
547.27 |
549.06 |
636.78 |
47.695 |
citrinet |
64 |
160 |
4407.8 |
3847.9 |
8337.2 |
8342 |
8371.6 |
62.036 |
citrinet |
96 |
160 |
18316 |
16287 |
25966 |
25984 |
25996 |
64.945 |
citrinet |
128 |
160 |
35592 |
38581 |
53236 |
53258 |
53277 |
64.504 |
Streaming, high-throughput¶
Acoustic model |
# of streams |
Chunk size (ms) |
Latency (ms) |
Throughput (RTFX) |
||||
---|---|---|---|---|---|---|---|---|
avg |
p50 |
p90 |
p95 |
p99 |
||||
quartznet |
1 |
800 |
39.787 |
38.593 |
48.83 |
48.83 |
48.83 |
0.99926 |
quartznet |
64 |
800 |
218.91 |
220.37 |
261.84 |
282.89 |
298.67 |
63.704 |
quartznet |
128 |
800 |
277.12 |
280.97 |
329.95 |
343.71 |
363.26 |
127.14 |
quartznet |
256 |
800 |
408.6 |
413.78 |
499.96 |
512.58 |
541.7 |
253.38 |
quartznet |
384 |
800 |
587.63 |
600.25 |
723.74 |
749.9 |
788.41 |
378.35 |
quartznet |
512 |
800 |
951.12 |
830.73 |
1447.2 |
1475.7 |
1529.2 |
501.08 |
quartznet |
768 |
800 |
5059.1 |
5104.1 |
6927.2 |
7345.6 |
7845.1 |
695.97 |
quartznet |
1024 |
800 |
30971 |
43331 |
49598 |
50224 |
50660 |
563.6 |
jasper |
1 |
800 |
86.674 |
89.077 |
97.608 |
97.608 |
97.608 |
0.99845 |
jasper |
64 |
800 |
266.74 |
266.8 |
294.38 |
301.01 |
307.94 |
63.634 |
jasper |
128 |
800 |
455.42 |
461.97 |
510.03 |
516.28 |
529.67 |
126.78 |
jasper |
256 |
800 |
1552.4 |
1558.4 |
2347 |
2373.2 |
2458.4 |
250.4 |
jasper |
384 |
800 |
13581 |
14258 |
18098 |
18924 |
21047 |
297.66 |
jasper |
512 |
800 |
27412 |
28616 |
38468 |
39011 |
41410 |
298.23 |
jasper |
768 |
800 |
52793 |
57476 |
73666 |
75289 |
77714 |
297.88 |
jasper |
1024 |
800 |
83042 |
91887 |
116950 |
119320 |
120910 |
297.89 |
citrinet |
1 |
800 |
37.433 |
35.331 |
46.439 |
46.439 |
46.439 |
0.9993 |
citrinet |
64 |
800 |
241.68 |
247.81 |
271.41 |
274.04 |
275.09 |
63.657 |
citrinet |
128 |
800 |
424.03 |
433.7 |
493.04 |
520.81 |
534.62 |
126.79 |
citrinet |
256 |
800 |
1369.2 |
1497.7 |
1863.8 |
2114.7 |
2281.9 |
251.09 |
citrinet |
384 |
800 |
10020 |
10609 |
13375 |
13609 |
14362 |
307.57 |
citrinet |
512 |
800 |
23613 |
23310 |
36234 |
36344 |
36439 |
306.23 |
citrinet |
768 |
800 |
52285 |
58787 |
75664 |
76529 |
77375 |
306.91 |
citrinet |
1024 |
800 |
81081 |
92544 |
109610 |
115160 |
116940 |
307.48 |
Offline¶
Acoustic model |
# of streams |
Chunk size (ms) |
Latency (ms) |
Throughput (RTFX) |
||||
---|---|---|---|---|---|---|---|---|
avg |
p50 |
p90 |
p95 |
p99 |
||||
quartznet |
1 |
1600 |
60.939 |
61.636 |
64.237 |
74.117 |
74.117 |
0.99897 |
quartznet |
256 |
1600 |
809.5 |
816.96 |
994.01 |
1020.4 |
1054.3 |
251.57 |
quartznet |
512 |
1600 |
1431.8 |
1385.6 |
1748.1 |
2171.5 |
2459.6 |
494.84 |
quartznet |
768 |
1600 |
2512.7 |
2608.9 |
3429.6 |
3595.7 |
3750.9 |
724.14 |
quartznet |
1024 |
1600 |
4221.7 |
4241.3 |
6399.2 |
6762.4 |
7838.2 |
938.91 |
quartznet |
1280 |
1600 |
19563 |
19149 |
34639 |
37018 |
40548 |
932.52 |
quartznet |
1512 |
1600 |
27150 |
28335 |
46041 |
48533 |
53078 |
948.37 |
jasper |
1 |
1600 |
100.59 |
103.02 |
107.48 |
115.31 |
115.31 |
0.99826 |
jasper |
256 |
1600 |
1867.4 |
1687.9 |
2727.4 |
2755.7 |
2824 |
247.67 |
jasper |
512 |
1600 |
12465 |
10921 |
21519 |
25644 |
27407 |
448.37 |
jasper |
768 |
1600 |
32174 |
34305 |
52300 |
54010 |
58519 |
459.94 |
jasper |
1024 |
1600 |
51569 |
57061 |
81640 |
83689 |
88960 |
458.96 |
jasper |
1280 |
1600 |
71547 |
81632 |
111870 |
113440 |
120440 |
460.32 |
jasper |
1512 |
1600 |
88812 |
105260 |
137660 |
139270 |
146870 |
462.58 |
citrinet |
1 |
1600 |
38.012 |
35.651 |
47.481 |
47.481 |
47.481 |
0.9993 |
citrinet |
256 |
1600 |
888.45 |
896.11 |
1034.9 |
1083.5 |
1121.9 |
250.51 |
citrinet |
512 |
1600 |
4005.4 |
3846.1 |
6488.6 |
6636.4 |
7069.2 |
490.21 |
citrinet |
768 |
1600 |
19784 |
19917 |
28363 |
28513 |
29171 |
502.52 |
citrinet |
1024 |
1600 |
35765 |
35869 |
54278 |
55179 |
55298 |
504.51 |
citrinet |
1280 |
1600 |
51602 |
57014 |
74890 |
75890 |
77607 |
506.95 |
citrinet |
1512 |
1600 |
66732 |
74685 |
101050 |
102330 |
103250 |
506.53 |
NLP¶
Performance of the Riva named entity recognition (NER) service (using a BERT-base model, sequence length of 128) and the Riva Question Answering (QA) service (using a BERT-large model, sequence length of 384) was measured in Riva. Batch size 1 latency and maximum throughput were measured.
NVIDIA A100 GPU¶
Task |
# of streams |
Latency (ms) |
Throughput (seq/s) |
||||
---|---|---|---|---|---|---|---|
avg |
p50 |
p90 |
p95 |
p99 |
|||
NER |
1 |
2.76 |
2.66 |
3.07 |
3.12 |
4.14 |
359.722 |
NER |
256 |
79.2 |
79.2 |
103 |
115 |
120 |
3170.57 |
Q&A |
1 |
7.56 |
7.39 |
7.78 |
8.99 |
9.21 |
131.922 |
Q&A |
128 |
243 |
238 |
267 |
318 |
447 |
523.794 |
NVIDIA V100 GPU¶
Task |
# of streams |
Latency (ms) |
Throughput (seq/s) |
||||
---|---|---|---|---|---|---|---|
avg |
p50 |
p90 |
p95 |
p99 |
|||
NER |
1 |
4.51 |
4.48 |
4.65 |
4.72 |
4.83 |
219.756 |
NER |
256 |
133 |
138 |
144 |
146 |
150 |
1834.98 |
Q&A |
1 |
10.1 |
10.1 |
10.2 |
10.2 |
10.3 |
98.253 |
Q&A |
128 |
588 |
589 |
642 |
709 |
1010 |
217.025 |
NVIDIA T4¶
Task |
# of streams |
Latency (ms) |
Throughput (seq/s) |
||||
---|---|---|---|---|---|---|---|
avg |
p50 |
p90 |
p95 |
p99 |
|||
NER |
1 |
6.32 |
5.69 |
8.74 |
8.84 |
9.17 |
156.82 |
NER |
256 |
268 |
268 |
280 |
316 |
367 |
912.299 |
Q&A |
1 |
16.6 |
16 |
16.4 |
19.3 |
31.3 |
60.2562 |
Q&A |
128 |
1300 |
1300 |
1640 |
1730 |
2380 |
98.0043 |
TTS¶
Performance of the Riva text-to-speech (TTS) service was measured for different number of parallel streams. Each parallel stream performed 10 iterations over 10 input strings from the LJSpeech dataset. Latency to first audio chunk and latency between successive audio chunks and throughput were measured.
NVIDIA A100 GPU¶
Model |
# of streams |
Latency to first audio (s) |
Latency between audio chunks (s) |
Throughput (RTFX) |
||||||
---|---|---|---|---|---|---|---|---|---|---|
avg |
p90 |
p95 |
p99 |
avg |
p90 |
p95 |
p99 |
|||
Fastpitch-Hifigan |
1 |
0.023 |
0.026 |
0.026 |
0.027 |
0.005 |
0.006 |
0.006 |
0.007 |
107.21 |
Fastpitch-Hifigan |
4 |
0.047 |
0.060 |
0.064 |
0.093 |
0.011 |
0.013 |
0.013 |
0.014 |
216.39 |
Fastpitch-Hifigan |
6 |
0.064 |
0.083 |
0.085 |
0.098 |
0.012 |
0.014 |
0.014 |
0.015 |
266.74 |
Fastpitch-Hifigan |
8 |
0.086 |
0.114 |
0.142 |
0.191 |
0.016 |
0.017 |
0.054 |
0.057 |
261.71 |
Fastpitch-Hifigan |
10 |
0.120 |
0.187 |
0.228 |
0.284 |
0.023 |
0.056 |
0.058 |
0.061 |
213.50 |
Tacotron2-Waveglow |
1 |
0.042 |
0.044 |
0.045 |
0.047 |
0.022 |
0.023 |
0.023 |
0.025 |
33.227 |
Tacotron2-Waveglow |
4 |
0.247 |
0.359 |
0.373 |
0.430 |
0.027 |
0.043 |
0.049 |
0.057 |
57.417 |
Tacotron2-Waveglow |
6 |
0.345 |
0.460 |
0.479 |
0.533 |
0.034 |
0.058 |
0.065 |
0.076 |
64.378 |
Tacotron2-Waveglow |
8 |
0.443 |
0.577 |
0.615 |
0.667 |
0.040 |
0.066 |
0.073 |
0.088 |
70.005 |
Tacotron2-Waveglow |
10 |
0.534 |
0.667 |
0.716 |
0.900 |
0.041 |
0.070 |
0.080 |
0.093 |
73.183 |
NVIDIA V100 GPU¶
Model |
# of streams |
Latency to first audio (s) |
Latency between audio chunks (s) |
Throughput (RTFX) |
||||||
---|---|---|---|---|---|---|---|---|---|---|
avg |
p90 |
p95 |
p99 |
avg |
p90 |
p95 |
p99 |
|||
Fastpitch-Hifigan |
1 |
0.033 |
0.036 |
0.036 |
0.037 |
0.011 |
0.011 |
0.011 |
0.011 |
62.71 |
Fastpitch-Hifigan |
4 |
0.116 |
0.152 |
0.160 |
0.173 |
0.020 |
0.024 |
0.025 |
0.025 |
102.31 |
Fastpitch-Hifigan |
6 |
0.177 |
0.223 |
0.234 |
0.245 |
0.025 |
0.029 |
0.032 |
0.034 |
111.53 |
Fastpitch-Hifigan |
8 |
0.240 |
0.302 |
0.318 |
0.347 |
0.029 |
0.035 |
0.036 |
0.039 |
117.56 |
Fastpitch-Hifigan |
10 |
0.289 |
0.369 |
0.388 |
0.423 |
0.031 |
0.037 |
0.039 |
0.052 |
119.08 |
Tacotron2-Waveglow |
1 |
0.058 |
0.060 |
0.060 |
0.064 |
0.031 |
0.033 |
0.033 |
0.034 |
23.84 |
Tacotron2-Waveglow |
4 |
0.355 |
0.515 |
0.536 |
0.589 |
0.048 |
0.075 |
0.090 |
0.103 |
36.40 |
Tacotron2-Waveglow |
6 |
0.510 |
0.689 |
0.740 |
0.823 |
0.062 |
0.105 |
0.117 |
0.132 |
40.24 |
Tacotron2-Waveglow |
8 |
0.680 |
0.886 |
0.937 |
1.013 |
0.071 |
0.119 |
0.134 |
0.159 |
43.03 |
Tacotron2-Waveglow |
10 |
0.819 |
1.037 |
1.098 |
1.323 |
0.075 |
0.126 |
0.139 |
0.165 |
44.21 |
NVIDIA T4¶
Model |
# of streams |
Latency to first audio (s) |
Latency between audio chunks (s) |
Throughput (RTFX) |
||||||
---|---|---|---|---|---|---|---|---|---|---|
avg |
p90 |
p95 |
p99 |
avg |
p90 |
p95 |
p99 |
|||
Fastpitch-Hifigan |
1 |
0.070 |
0.070 |
0.070 |
0.110 |
0.020 |
0.020 |
0.020 |
0.020 |
32.53 |
Fastpitch-Hifigan |
4 |
0.450 |
0.590 |
0.610 |
0.660 |
0.030 |
0.040 |
0.040 |
0.050 |
38.49 |
Fastpitch-Hifigan |
6 |
0.670 |
0.920 |
0.980 |
1.030 |
0.040 |
0.060 |
0.060 |
0.060 |
39.94 |
Fastpitch-Hifigan |
8 |
0.900 |
1.230 |
1.300 |
1.440 |
0.050 |
0.070 |
0.080 |
0.080 |
40.73 |
Fastpitch-Hifigan |
10 |
1.070 |
1.460 |
1.540 |
1.610 |
0.050 |
0.080 |
0.080 |
0.090 |
40.99 |
Tacotron2-Waveglow |
1 |
0.110 |
0.110 |
0.110 |
0.130 |
0.050 |
0.050 |
0.050 |
0.050 |
14.57 |
Tacotron2-Waveglow |
4 |
0.640 |
0.910 |
0.990 |
1.100 |
0.100 |
0.170 |
0.190 |
0.220 |
18.57 |
Tacotron2-Waveglow |
6 |
0.970 |
1.330 |
1.410 |
1.540 |
0.140 |
0.230 |
0.270 |
0.320 |
19.26 |
Tacotron2-Waveglow |
8 |
1.370 |
1.810 |
1.890 |
2.050 |
0.170 |
0.290 |
0.330 |
0.410 |
19.69 |
Tacotron2-Waveglow |
10 |
1.700 |
2.200 |
2.320 |
2.570 |
0.180 |
0.300 |
0.340 |
0.420 |
19.97 |
When the server is under high load, requests might time out, as the server will not start inference for a new request until a previous request is completely generated so that inference slot can be freed. This is done to maximize throughput for the TTS service and allow for real-time interaction. NVIDIA does not recommend making more than 8-10 simultaneous requests with the models provided in Riva 1.0.0 beta.