Performance and Memory for NVIDIA NeMo Retriever Embedding NIM#

To benchmark the performance of NVIDIA NeMo Retriever Embedding NIM, you can use the AIPerf tool.

To run a performance benchmark with representative input data, create a JSONL file where each line contains a {"texts": [...]} object, as shown in the following example.

Example#

Create a file named embeddings.jsonl that contains the following content.

{"texts": ["What was the first car ever driven?"]}
{"texts": ["Who served as the 5th President of the United States of America?"]}
{"texts": ["Is the Sydney Opera House located in Australia?"]}
{"texts": ["In what state did they film Shrek 2?"]}

Install the verified AIPerf version in a Python environment that can reach the NIM endpoint.

pip install aiperf==0.9.0

Run the following command to benchmark the embedding endpoint by using AIPerf.

aiperf profile \
    --model nvidia/nemotron-3-embed-1b \
    --endpoint-type nim_embeddings \
    --url localhost:8000 \
    --input-file embeddings.jsonl \
    --custom-dataset-type single_turn \
    --extra-inputs input_type:query \
    --extra-inputs truncate:END \
    --request-count 20 \
    --concurrency 5

For the full set of command line options, refer to the AIPerf command line options documentation.

About the Measurements#

The following sections report the memory footprint and the benchmark results for each model. All latency measurements are reported in milliseconds.

The previous AIPerf example benchmarks a custom JSONL workload. The benchmark tables use a separate synthetic AIPerf benchmark matrix for consistent GPU comparisons. For text rows, batch size is embedding inputs per request and throughput is embedding inputs per second. For image rows, batch size is images per request, image size is the generated JPEG resolution, and throughput is images per second.

Nemotron 3 Embed 1B#

Memory Footprint#

The following table contains the memory footprint data for Nemotron 3 Embed 1B.

Compute Capability

Precision

Approximate GPU Memory Size (GiB)

12.0

nvfp4

2.34

10.0

nvfp4

2.4

9.0

bf16

3.9

8.9

bf16

3.72

8.6

bf16

3.55

8.0

bf16

3.72

NVIDIA RTX PRO 6000 Blackwell Server Edition#

The following table contains the performance data for Nemotron 3 Embed 1B on NVIDIA RTX PRO 6000 Blackwell Server Edition.

Precision

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

NVFP4

20

1

1

2.9

2.9

2.9

2.9

222.9

NVFP4

20

1

3

3.5

3.6

4.0

4.5

570.2

NVFP4

20

1

5

5.1

4.4

5.8

13.0

742.9

NVFP4

20

1

7

4.3

4.5

5.0

5.1

1123.8

NVFP4

20

1

9

5.6

5.1

5.5

13.5

1206.0

NVFP4

20

1

11

6.2

4.8

13.6

13.8

1345.3

NVFP4

20

1

13

11.2

13.7

14.3

14.3

936.1

NVFP4

20

1

15

9.6

12.4

14.4

14.5

1185.6

NVFP4

300

64

1

63.6

63.7

63.9

64.0

906.6

NVFP4

300

64

3

156.0

157.2

157.7

157.8

1168.3

NVFP4

300

64

5

261.9

266.5

267.1

267.5

1167.7

NVFP4

512

64

1

118.1

118.1

118.5

118.7

511.4

NVFP4

512

64

3

317.8

320.7

321.5

321.7

585.8

NVFP4

512

64

5

529.8

540.0

541.1

541.4

584.6

BF16

20

1

1

13.5

13.5

13.5

13.5

66.1

BF16

20

1

3

13.8

13.8

13.9

14.0

189.8

BF16

20

1

5

13.8

13.8

13.9

14.0

323.6

BF16

20

1

7

14.2

14.1

14.5

14.5

421.6

BF16

20

1

9

14.7

14.7

14.9

15.0

510.0

BF16

20

1

11

15.3

15.3

15.6

15.7

591.9

BF16

20

1

13

15.6

15.6

15.9

16.0

729.1

BF16

20

1

15

16.3

16.4

17.0

17.1

792.6

BF16

300

64

1

202.8

202.3

204.1

206.5

305.5

BF16

300

64

3

544.3

548.5

553.0

554.0

344.9

BF16

300

64

5

903.2

920.4

925.3

925.9

344.7

BF16

512

64

1

517.9

515.3

527.5

528.7

121.9

BF16

512

64

3

1449.8

1463.3

1477.5

1482.5

130.5

BF16

512

64

5

2397.6

2443.2

2462.1

2467.2

130.5

NVIDIA GB200#

The following table contains the performance data for Nemotron 3 Embed 1B on NVIDIA GB200.

Precision

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

NVFP4

20

1

1

1.7

1.7

1.7

1.7

452.5

NVFP4

20

1

3

2.4

2.4

2.4

2.5

1006.3

NVFP4

20

1

5

2.9

2.4

2.5

3.6

1433.7

NVFP4

20

1

7

3.8

2.4

12.0

12.1

1557.6

NVFP4

20

1

9

3.5

2.6

3.5

12.4

2103.4

NVFP4

20

1

11

3.8

2.7

12.6

12.8

2356.3

NVFP4

20

1

13

2.8

2.7

4.2

4.3

3540.1

NVFP4

20

1

15

4.5

2.9

13.1

13.2

2764.9

NVFP4

300

64

1

36.8

36.8

37.3

38.0

1500.6

NVFP4

300

64

3

69.2

69.3

70.5

70.6

2510.3

NVFP4

300

64

5

118.4

120.1

121.0

121.4

2512.4

NVFP4

512

64

1

59.4

59.4

60.1

60.2

980.0

NVFP4

512

64

3

127.9

128.7

129.6

129.9

1414.0

NVFP4

512

64

5

215.1

218.3

219.5

219.7

1413.7

BF16

20

1

1

12.0

12.1

12.1

12.1

79.6

BF16

20

1

3

12.3

12.2

12.3

12.7

228.8

BF16

20

1

5

12.5

12.4

12.8

13.1

380.5

BF16

20

1

7

12.6

12.6

12.7

13.1

503.0

BF16

20

1

9

12.8

12.7

13.3

13.4

619.1

BF16

20

1

11

13.7

13.1

16.8

16.9

701.0

BF16

20

1

13

13.7

13.3

15.7

15.7

862.3

BF16

20

1

15

14.4

13.7

17.5

17.6

942.3

BF16

300

64

1

136.5

137.2

141.0

142.3

450.1

BF16

300

64

3

330.3

333.3

341.7

344.1

565.3

BF16

300

64

5

549.0

558.9

572.4

574.7

565.1

BF16

512

64

1

345.8

350.4

356.2

358.3

182.1

BF16

512

64

3

912.8

923.8

942.9

946.2

206.8

BF16

512

64

5

1510.6

1540.2

1571.9

1574.0

206.8

NVIDIA H100 80GB HBM3#

The following table contains the performance data for Nemotron 3 Embed 1B on NVIDIA H100 80GB HBM3.

Precision

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

BF16

20

1

1

3.3

3.2

3.3

3.3

202.8

BF16

20

1

3

3.4

3.3

3.6

3.8

591.6

BF16

20

1

5

6.6

6.6

6.9

7.0

594.2

BF16

20

1

7

9.8

9.9

10.1

10.1

594.4

BF16

20

1

9

12.9

13.3

13.5

13.6

593.1

BF16

20

1

11

16.0

16.6

16.8

16.9

593.9

BF16

20

1

13

19.0

19.9

20.1

20.5

592.3

BF16

20

1

15

21.9

23.2

23.4

23.5

592.1

BF16

300

64

1

292.9

292.8

294.9

295.1

213.2

BF16

300

64

3

702.8

706.1

710.8

711.9

268.4

BF16

300

64

5

1179.9

1180.1

1185.8

1189.8

269.1

BF16

512

64

1

499.4

499.6

509.5

513.1

126.2

BF16

512

64

3

1277.2

1284.9

1302.2

1306.3

148.2

BF16

512

64

5

2146.0

2143.2

2171.2

2174.5

148.5

NVIDIA H100 NVL#

The following table contains the performance data for Nemotron 3 Embed 1B on NVIDIA H100 NVL.

Precision

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

BF16

20

1

1

3.2

3.2

3.2

3.3

208.1

BF16

20

1

3

3.3

3.3

3.6

3.7

590.6

BF16

20

1

5

6.4

6.4

6.6

6.7

612.4

BF16

20

1

7

9.5

9.6

9.8

9.8

612.2

BF16

20

1

9

12.5

12.8

13.1

13.2

612.5

BF16

20

1

11

15.5

16.1

16.3

16.4

611.6

BF16

20

1

13

18.4

19.3

19.5

19.6

612.3

BF16

20

1

15

21.3

22.5

22.8

22.9

611.0

BF16

300

64

1

303.3

303.3

304.9

305.3

206.4

BF16

300

64

3

747.4

751.7

755.8

756.9

252.8

BF16

300

64

5

1254.0

1255.5

1263.7

1264.9

253.4

BF16

512

64

1

533.4

532.7

545.2

547.5

118.5

BF16

512

64

3

1401.8

1409.8

1428.6

1434.9

135.3

BF16

512

64

5

2355.3

2351.4

2382.5

2385.2

135.4

NVIDIA A100 SXM4 80GB#

The following table contains the performance data for Nemotron 3 Embed 1B on NVIDIA A100 SXM4 80GB.

Precision

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

BF16

20

1

1

3.2

3.3

3.3

3.3

252.1

BF16

20

1

3

6.8

6.6

7.5

7.5

388.8

BF16

20

1

5

12.5

12.9

13.0

13.1

367.6

BF16

20

1

7

16.9

16.5

18.5

18.6

379.3

BF16

20

1

9

22.5

23.8

24.0

24.1

367.0

BF16

20

1

11

26.1

26.1

29.5

29.6

384.5

BF16

20

1

13

32.4

34.7

35.0

35.2

365.0

BF16

20

1

15

35.0

36.1

40.6

40.8

387.2

BF16

300

64

1

510.2

510.1

513.8

514.9

123.2

BF16

300

64

3

1412.2

1420.5

1428.5

1429.9

134.1

BF16

300

64

5

2374.8

2376.4

2387.5

2388.2

134.2

BF16

512

64

1

913.2

912.1

937.1

942.0

69.3

BF16

512

64

3

2550.6

2563.2

2608.4

2612.8

74.5

BF16

512

64

5

4285.0

4280.4

4344.7

4348.0

74.5

NVIDIA L40S#

The following table contains the performance data for Nemotron 3 Embed 1B on NVIDIA L40S.

Precision

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

BF16

20

1

1

4.8

4.8

4.9

4.9

156.4

BF16

20

1

3

8.7

8.7

8.9

9.1

293.6

BF16

20

1

5

15.1

15.3

15.6

15.7

293.8

BF16

20

1

7

21.6

22.1

22.4

22.5

293.6

BF16

20

1

9

27.8

28.8

29.1

29.2

293.7

BF16

20

1

11

34.1

35.7

36.0

36.0

293.9

BF16

20

1

13

40.2

42.4

42.6

42.7

293.6

BF16

20

1

15

46.0

49.0

49.3

49.4

293.6

BF16

300

64

1

493.7

493.8

495.9

496.2

127.9

BF16

300

64

3

1344.4

1353.0

1356.4

1358.2

141.1

BF16

300

64

5

2271.3

2272.4

2276.9

2278.5

140.4

BF16

512

64

1

796.9

795.1

820.0

822.4

79.6

BF16

512

64

3

2240.7

2255.1

2287.4

2292.2

84.8

BF16

512

64

5

3765.1

3766.5

3813.4

3817.6

84.8

NVIDIA A10G#

The following table contains the performance data for Nemotron 3 Embed 1B on NVIDIA A10G.

Precision

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

BF16

20

1

1

5.8

5.8

5.8

5.8

148.5

BF16

20

1

3

13.9

14.0

14.4

14.6

196.5

BF16

20

1

5

23.7

24.1

24.4

24.4

196.6

BF16

20

1

7

33.5

34.4

34.7

34.7

196.7

BF16

20

1

9

42.8

44.4

44.9

45.0

196.3

BF16

20

1

11

52.0

54.6

55.0

55.1

196.0

BF16

20

1

13

61.0

64.7

65.0

65.1

196.4

BF16

20

1

15

69.7

74.8

75.2

75.3

196.1

BF16

300

64

1

1148.5

1147.6

1152.8

1157.4

55.2

BF16

300

64

3

3306.7

3326.1

3342.6

3343.9

57.6

BF16

300

64

5

5546.9

5547.5

5553.8

5556.8

57.6

BF16

512

64

1

2113.1

2114.9

2179.9

2194.4

30.1

BF16

512

64

3

6117.3

6157.5

6257.9

6276.2

31.1

BF16

512

64

5

10270.1

10277.2

10429.7

10455.3

31.1

Llama Nemotron Embed VL 1B v2#

Memory Footprint#

The following table contains the memory footprint data for Llama Nemotron Embed VL 1B v2.

Compute Capability

Precision

Approximate GPU Memory Size (GiB)

12.0

fp8

7.96

10.0

fp8

9.04

9.0

fp8

8.87

8.9

fp8

8.21

8.9

fp16

5.26

8.6

fp16

5.33

8.0

fp16

5.49

NVIDIA RTX PRO 6000 Blackwell Server Edition#

The following table contains the performance data for Llama Nemotron Embed VL 1B v2 on NVIDIA RTX PRO 6000 Blackwell Server Edition.

Precision

Input Type

Image Size

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

Mixed

text

20

1

1

3.8

3.8

3.9

3.9

183.8

Mixed

text

20

1

3

5.8

5.7

6.0

6.1

407.0

Mixed

text

20

1

5

10.6

10.7

10.9

11.0

403.2

Mixed

text

20

1

7

15.3

15.6

15.9

16.1

401.5

Mixed

text

20

1

9

20.5

21.2

21.4

21.5

393.4

Mixed

text

20

1

11

25.1

26.2

26.4

26.5

393.2

Mixed

text

20

1

13

29.6

31.2

31.5

31.6

392.9

Mixed

text

20

1

15

34.0

36.2

36.5

36.6

392.7

Mixed

text

300

64

1

260.6

260.5

261.0

261.2

239.4

Mixed

text

300

64

3

752.0

759.2

760.3

760.5

250.5

Mixed

text

300

64

5

1245.2

1270.1

1271.1

1271.3

250.5

Mixed

text

512

64

1

401.9

401.3

411.2

412.8

156.6

Mixed

text

512

64

3

1170.1

1180.9

1194.8

1201.3

161.5

Mixed

text

512

64

5

1932.8

1972.3

1988.9

1990.8

161.6

Mixed

image passage

336x336 JPEG

1

1

12.3

12.3

12.5

12.7

81.1

Mixed

image passage

336x336 JPEG

1

2

19.2

19.2

19.4

19.5

104.4

Mixed

image passage

336x336 JPEG

1

3

29.5

29.7

29.9

30.0

102.5

Mixed

image passage

336x336 JPEG

1

5

49.8

50.6

50.8

50.9

103.3

Mixed

image passage

336x336 JPEG

1

8

79.5

82.1

82.3

82.5

108.7

Mixed

image passage

336x336 JPEG

4

1

11.1

11.1

11.1

11.2

90.1

Mixed

image passage

336x336 JPEG

4

2

20.4

20.5

20.6

20.6

98.3

Mixed

image passage

336x336 JPEG

4

3

30.6

30.9

30.9

30.9

99.2

Mixed

image passage

336x336 JPEG

4

5

50.9

51.8

51.9

51.9

102.1

Mixed

image passage

336x336 JPEG

4

8

80.5

83.3

83.4

83.4

108.4

Mixed

image passage

336x336 JPEG

8

1

10.9

10.9

10.9

10.9

91.5

Mixed

image passage

336x336 JPEG

8

2

20.6

20.6

20.7

20.8

97.7

Mixed

image passage

336x336 JPEG

8

3

30.8

31.1

31.1

31.2

98.6

Mixed

image passage

336x336 JPEG

8

5

51.1

52.1

52.2

52.2

101.7

Mixed

image passage

336x336 JPEG

8

8

80.8

83.5

83.7

84.5

108.3

Mixed

image passage

512x512 JPEG

1

1

12.2

12.2

12.4

12.4

81.7

Mixed

image passage

512x512 JPEG

1

2

19.1

19.2

19.3

19.4

104.7

Mixed

image passage

512x512 JPEG

1

3

29.5

29.7

30.0

30.3

102.5

Mixed

image passage

512x512 JPEG

1

5

49.7

50.6

50.8

50.8

103.9

Mixed

image passage

512x512 JPEG

1

8

79.4

82.1

82.2

82.3

108.6

Mixed

image passage

512x512 JPEG

4

1

11.2

11.2

11.3

11.3

89.1

Mixed

image passage

512x512 JPEG

4

2

20.3

20.4

20.5

20.5

98.7

Mixed

image passage

512x512 JPEG

4

3

30.6

30.9

30.9

30.9

99.3

Mixed

image passage

512x512 JPEG

4

5

50.9

51.8

51.9

51.9

102.1

Mixed

image passage

512x512 JPEG

4

8

80.4

83.3

83.3

83.3

108.3

Mixed

image passage

512x512 JPEG

8

1

11.1

11.1

11.1

11.1

90.4

Mixed

image passage

512x512 JPEG

8

2

20.5

20.6

20.6

20.6

97.8

Mixed

image passage

512x512 JPEG

8

3

30.8

31.1

31.1

31.1

98.6

Mixed

image passage

512x512 JPEG

8

5

51.0

52.0

52.1

52.1

101.8

Mixed

image passage

512x512 JPEG

8

8

80.7

83.6

83.7

83.7

108.2

Mixed

image passage

1024x1024 JPEG

1

1

12.7

12.7

13.0

13.0

78.5

Mixed

image passage

1024x1024 JPEG

1

2

19.0

19.0

19.2

19.3

105.6

Mixed

image passage

1024x1024 JPEG

1

3

29.3

29.6

29.8

29.8

103.1

Mixed

image passage

1024x1024 JPEG

1

5

49.7

50.5

50.7

50.8

103.7

Mixed

image passage

1024x1024 JPEG

1

8

79.3

81.9

82.2

82.2

108.3

Mixed

image passage

1024x1024 JPEG

4

1

11.8

11.8

11.9

12.0

84.8

Mixed

image passage

1024x1024 JPEG

4

2

20.3

20.3

20.3

20.4

99.0

Mixed

image passage

1024x1024 JPEG

4

3

30.5

30.8

30.8

30.8

99.4

Mixed

image passage

1024x1024 JPEG

4

5

50.8

51.7

52.0

52.4

101.8

Mixed

image passage

1024x1024 JPEG

4

8

80.3

83.1

83.2

83.2

108.0

Mixed

image passage

1024x1024 JPEG

8

1

11.7

11.7

11.8

11.8

85.5

Mixed

image passage

1024x1024 JPEG

8

2

20.4

20.5

20.5

20.5

98.4

Mixed

image passage

1024x1024 JPEG

8

3

30.7

30.9

31.0

31.1

98.8

Mixed

image passage

1024x1024 JPEG

8

5

51.0

51.9

52.0

52.0

101.6

Mixed

image passage

1024x1024 JPEG

8

8

80.5

83.3

83.4

83.4

107.9

Mixed

image passage

2048x2048 JPEG

1

1

14.4

14.4

15.0

15.0

69.6

Mixed

image passage

2048x2048 JPEG

1

2

18.7

18.6

18.9

18.9

107.4

Mixed

image passage

2048x2048 JPEG

1

3

29.0

29.2

29.5

29.5

104.0

Mixed

image passage

2048x2048 JPEG

1

5

49.3

50.1

50.4

50.5

103.9

Mixed

image passage

2048x2048 JPEG

1

8

79.3

81.6

82.1

82.5

107.4

Mixed

image passage

2048x2048 JPEG

4

1

13.5

13.5

13.9

14.0

74.3

Mixed

image passage

2048x2048 JPEG

4

2

20.0

20.0

20.1

20.1

100.3

Mixed

image passage

2048x2048 JPEG

4

3

30.3

30.5

30.5

30.6

99.8

Mixed

image passage

2048x2048 JPEG

4

5

50.5

51.3

51.5

51.5

101.8

Mixed

image passage

2048x2048 JPEG

4

8

80.2

82.8

82.9

82.9

107.2

Mixed

image passage

2048x2048 JPEG

8

1

13.1

13.1

13.3

13.4

76.4

Mixed

image passage

2048x2048 JPEG

8

2

20.0

20.1

20.1

20.2

100.0

Mixed

image passage

2048x2048 JPEG

8

3

30.3

30.5

30.6

30.7

99.8

Mixed

image passage

2048x2048 JPEG

8

5

50.7

51.6

51.7

52.0

101.6

Mixed

image passage

2048x2048 JPEG

8

8

80.3

82.9

83.1

83.1

106.9

NVIDIA B200#

The following table contains the performance data for Llama Nemotron Embed VL 1B v2 on NVIDIA B200.

Precision

Input Type

Image Size

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP8

text

20

1

1

2.9

2.9

3.0

3.0

220.2

FP8

text

20

1

3

3.9

4.3

4.9

4.9

530.0

FP8

text

20

1

5

4.6

4.7

6.6

6.9

793.7

FP8

text

20

1

7

4.9

4.7

9.1

9.4

1042.5

FP8

text

20

1

9

5.0

4.6

6.7

11.0

1315.1

FP8

text

20

1

11

5.4

5.2

7.6

9.5

1476.3

FP8

text

20

1

13

5.2

4.4

10.8

11.1

1871.1

FP8

text

20

1

15

6.1

5.5

11.6

11.8

1786.1

FP8

text

300

64

1

58.9

58.6

60.5

60.9

971.2

FP8

text

300

64

3

154.6

155.8

159.4

161.0

1177.6

FP8

text

300

64

5

261.1

266.3

270.2

270.9

1170.5

FP8

text

512

64

1

96.4

96.4

97.3

97.5

618.8

FP8

text

512

64

3

267.0

269.6

271.0

271.5

693.8

FP8

text

512

64

5

445.3

453.7

457.5

457.8

693.2

FP8

image passage

336x336 JPEG

1

1

15.3

15.3

15.4

15.4

65.2

FP8

image passage

336x336 JPEG

1

2

23.7

22.0

27.1

37.3

87.0

FP8

image passage

336x336 JPEG

1

3

33.8

34.0

34.2

34.3

89.5

FP8

image passage

336x336 JPEG

1

5

56.8

57.7

58.0

58.1

90.5

FP8

image passage

336x336 JPEG

1

8

90.4

93.4

93.7

93.8

94.6

FP8

image passage

336x336 JPEG

4

1

12.9

12.9

12.9

12.9

77.7

FP8

image passage

336x336 JPEG

4

2

23.1

23.2

23.3

23.3

86.8

FP8

image passage

336x336 JPEG

4

3

36.2

35.1

35.2

35.4

85.7

FP8

image passage

336x336 JPEG

4

5

57.9

59.0

59.0

59.0

89.7

FP8

image passage

336x336 JPEG

4

8

91.4

94.6

94.7

94.7

95.3

FP8

image passage

336x336 JPEG

8

1

12.5

12.5

12.5

12.5

80.2

FP8

image passage

336x336 JPEG

8

2

23.4

23.5

23.5

23.5

86.0

FP8

image passage

336x336 JPEG

8

3

35.0

35.3

35.4

35.4

86.8

FP8

image passage

336x336 JPEG

8

5

58.6

59.1

59.1

63.2

88.9

FP8

image passage

336x336 JPEG

8

8

91.5

94.8

94.8

94.8

95.5

FP8

image passage

512x512 JPEG

1

1

15.5

15.5

15.6

15.6

64.6

FP8

image passage

512x512 JPEG

1

2

22.1

22.1

22.3

22.5

90.6

FP8

image passage

512x512 JPEG

1

3

34.0

34.2

34.5

34.6

88.8

FP8

image passage

512x512 JPEG

1

5

59.9

58.0

59.1

76.1

87.5

FP8

image passage

512x512 JPEG

1

8

90.7

93.6

94.0

94.1

94.3

FP8

image passage

512x512 JPEG

4

1

13.0

13.0

13.1

13.1

76.8

FP8

image passage

512x512 JPEG

4

2

23.2

23.3

23.3

23.3

86.6

FP8

image passage

512x512 JPEG

4

3

34.9

35.2

35.2

35.2

87.0

FP8

image passage

512x512 JPEG

4

5

60.1

59.0

59.1

79.0

87.6

FP8

image passage

512x512 JPEG

4

8

92.4

94.8

96.4

105.1

94.5

FP8

image passage

512x512 JPEG

8

1

12.7

12.6

12.7

12.7

78.8

FP8

image passage

512x512 JPEG

8

2

23.8

23.5

23.5

23.5

84.9

FP8

image passage

512x512 JPEG

8

3

35.4

35.4

35.5

39.0

85.9

FP8

image passage

512x512 JPEG

8

5

58.1

59.2

59.2

59.3

89.5

FP8

image passage

512x512 JPEG

8

8

91.7

95.0

95.1

95.2

95.2

FP8

image passage

1024x1024 JPEG

1

1

52.8

52.7

53.0

53.0

19.0

FP8

image passage

1024x1024 JPEG

1

2

95.9

95.6

95.9

96.2

21.0

FP8

image passage

1024x1024 JPEG

1

3

144.7

144.4

144.8

150.3

21.0

FP8

image passage

1024x1024 JPEG

1

5

237.9

242.2

244.5

244.9

21.8

FP8

image passage

1024x1024 JPEG

1

8

374.9

388.0

388.5

388.7

23.2

FP8

image passage

1024x1024 JPEG

4

1

50.5

50.3

50.4

50.5

19.8

FP8

image passage

1024x1024 JPEG

4

2

96.8

96.8

98.0

98.1

20.8

FP8

image passage

1024x1024 JPEG

4

3

145.9

145.6

146.9

156.8

20.9

FP8

image passage

1024x1024 JPEG

4

5

240.7

243.0

250.3

257.5

21.6

FP8

image passage

1024x1024 JPEG

4

8

380.2

389.4

399.9

428.6

23.1

FP8

image passage

1024x1024 JPEG

8

1

50.2

49.9

50.0

50.4

19.9

FP8

image passage

1024x1024 JPEG

8

2

97.2

97.0

98.4

98.8

20.7

FP8

image passage

1024x1024 JPEG

8

3

145.1

145.7

146.1

151.0

21.0

FP8

image passage

1024x1024 JPEG

8

5

240.2

243.3

248.7

251.2

21.7

FP8

image passage

1024x1024 JPEG

8

8

378.5

389.7

395.7

408.4

23.2

FP8

image passage

2048x2048 JPEG

1

1

54.3

54.3

55.0

55.0

18.4

FP8

image passage

2048x2048 JPEG

1

2

98.7

96.0

96.3

96.4

20.7

FP8

image passage

2048x2048 JPEG

1

3

143.8

145.1

145.4

145.4

21.1

FP8

image passage

2048x2048 JPEG

1

5

238.8

243.3

243.7

243.8

21.7

FP8

image passage

2048x2048 JPEG

1

8

379.3

390.9

391.3

391.7

23.0

FP8

image passage

2048x2048 JPEG

4

1

52.0

51.8

52.2

52.3

19.2

FP8

image passage

2048x2048 JPEG

4

2

97.0

97.2

97.4

97.9

20.7

FP8

image passage

2048x2048 JPEG

4

3

146.6

146.5

148.1

148.4

20.8

FP8

image passage

2048x2048 JPEG

4

5

243.4

244.7

249.7

260.2

21.4

FP8

image passage

2048x2048 JPEG

4

8

379.3

392.1

394.5

399.7

23.0

FP8

image passage

2048x2048 JPEG

8

1

51.7

51.5

51.8

51.9

19.4

FP8

image passage

2048x2048 JPEG

8

2

97.0

97.3

97.5

99.1

20.7

FP8

image passage

2048x2048 JPEG

8

3

145.8

146.5

147.6

149.1

20.9

FP8

image passage

2048x2048 JPEG

8

5

240.9

244.7

247.6

248.2

21.6

FP8

image passage

2048x2048 JPEG

8

8

380.6

392.3

396.6

409.9

23.0

NVIDIA H100 80GB HBM3#

The following table contains the performance data for Llama Nemotron Embed VL 1B v2 on NVIDIA H100 80GB HBM3.

Precision

Input Type

Image Size

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP8

text

20

1

1

3.5

3.5

3.6

3.7

193.8

FP8

text

20

1

3

5.1

5.3

6.1

6.4

436.0

FP8

text

20

1

5

5.2

5.3

6.2

6.3

710.2

FP8

text

20

1

7

5.6

5.4

6.1

11.2

933.0

FP8

text

20

1

9

7.1

5.6

14.7

16.0

999.3

FP8

text

20

1

11

6.8

6.0

10.5

12.6

1218.5

FP8

text

20

1

13

9.8

8.0

16.8

19.4

1099.3

FP8

text

20

1

15

6.3

6.1

9.0

9.4

1809.5

FP8

text

300

64

1

72.0

71.8

73.1

73.5

805.6

FP8

text

300

64

3

193.1

194.7

196.5

196.8

947.6

FP8

text

300

64

5

323.1

329.2

331.0

331.5

948.1

FP8

text

512

64

1

122.6

122.5

123.8

124.1

491.6

FP8

text

512

64

3

344.8

347.7

350.9

351.6

539.1

FP8

text

512

64

5

571.6

582.8

585.8

586.3

541.1

FP8

image passage

336x336 JPEG

1

1

17.9

17.6

18.7

19.3

56.0

FP8

image passage

336x336 JPEG

1

2

27.4

27.0

28.9

29.9

73.3

FP8

image passage

336x336 JPEG

1

3

41.6

41.2

44.1

46.2

72.9

FP8

image passage

336x336 JPEG

1

5

69.7

70.1

73.5

74.2

73.9

FP8

image passage

336x336 JPEG

1

8

111.0

113.5

118.9

123.5

77.7

FP8

image passage

336x336 JPEG

4

1

15.6

15.3

16.5

17.0

64.1

FP8

image passage

336x336 JPEG

4

2

28.6

28.1

30.6

30.7

70.4

FP8

image passage

336x336 JPEG

4

3

42.8

42.6

44.9

45.9

71.0

FP8

image passage

336x336 JPEG

4

5

70.2

71.2

73.2

73.6

74.0

FP8

image passage

336x336 JPEG

4

8

111.9

115.1

119.0

122.3

78.2

FP8

image passage

336x336 JPEG

8

1

15.1

14.9

15.8

16.5

66.3

FP8

image passage

336x336 JPEG

8

2

28.7

28.6

30.2

31.0

70.1

FP8

image passage

336x336 JPEG

8

3

42.9

43.0

44.3

44.6

70.9

FP8

image passage

336x336 JPEG

8

5

70.7

72.0

73.5

73.5

73.7

FP8

image passage

336x336 JPEG

8

8

113.3

116.2

120.0

127.6

77.4

FP8

image passage

512x512 JPEG

1

1

18.2

17.8

19.2

20.5

55.1

FP8

image passage

512x512 JPEG

1

2

27.6

27.1

29.6

30.9

72.8

FP8

image passage

512x512 JPEG

1

3

42.3

41.5

46.1

49.3

71.7

FP8

image passage

512x512 JPEG

1

5

69.6

70.1

72.6

76.8

73.6

FP8

image passage

512x512 JPEG

1

8

113.5

115.2

124.8

125.2

76.3

FP8

image passage

512x512 JPEG

4

1

15.9

15.5

16.5

16.8

63.3

FP8

image passage

512x512 JPEG

4

2

28.7

28.4

29.9

30.4

70.2

FP8

image passage

512x512 JPEG

4

3

42.5

42.7

44.2

44.4

71.4

FP8

image passage

512x512 JPEG

4

5

70.9

72.2

74.1

74.3

72.7

FP8

image passage

512x512 JPEG

4

8

111.8

115.3

117.6

118.2

77.4

FP8

image passage

512x512 JPEG

8

1

15.3

15.1

15.9

16.1

65.6

FP8

image passage

512x512 JPEG

8

2

28.8

28.7

29.8

30.0

69.7

FP8

image passage

512x512 JPEG

8

3

43.1

43.5

44.7

45.5

70.5

FP8

image passage

512x512 JPEG

8

5

71.5

72.3

74.2

75.1

72.7

FP8

image passage

512x512 JPEG

8

8

112.6

116.5

118.4

118.7

77.4

FP8

image passage

1024x1024 JPEG

1

1

64.7

64.0

66.8

67.7

15.5

FP8

image passage

1024x1024 JPEG

1

2

118.6

117.9

122.3

124.6

16.9

FP8

image passage

1024x1024 JPEG

1

3

178.5

179.7

184.9

185.7

17.0

FP8

image passage

1024x1024 JPEG

1

5

296.6

300.1

308.9

313.2

17.4

FP8

image passage

1024x1024 JPEG

1

8

470.5

482.9

495.8

516.7

18.4

FP8

image passage

1024x1024 JPEG

4

1

62.2

61.6

62.8

63.0

16.1

FP8

image passage

1024x1024 JPEG

4

2

120.6

120.3

121.7

126.5

16.7

FP8

image passage

1024x1024 JPEG

4

3

179.4

181.0

182.6

183.5

17.0

FP8

image passage

1024x1024 JPEG

4

5

296.9

301.8

304.3

306.1

17.6

FP8

image passage

1024x1024 JPEG

4

8

469.0

484.2

497.5

498.9

18.7

FP8

image passage

1024x1024 JPEG

8

1

62.0

62.0

62.5

62.8

16.1

FP8

image passage

1024x1024 JPEG

8

2

120.0

120.5

121.7

122.3

16.7

FP8

image passage

1024x1024 JPEG

8

3

179.8

181.2

182.1

187.6

16.9

FP8

image passage

1024x1024 JPEG

8

5

296.4

302.2

303.6

304.9

17.6

FP8

image passage

1024x1024 JPEG

8

8

468.1

484.3

486.2

493.5

18.7

FP8

image passage

2048x2048 JPEG

1

1

66.3

66.1

68.3

68.8

15.1

FP8

image passage

2048x2048 JPEG

1

2

118.4

118.2

121.1

124.6

17.0

FP8

image passage

2048x2048 JPEG

1

3

178.7

179.4

183.0

186.0

17.0

FP8

image passage

2048x2048 JPEG

1

5

296.4

301.7

306.3

306.9

17.5

FP8

image passage

2048x2048 JPEG

1

8

469.7

486.0

491.8

492.2

18.6

FP8

image passage

2048x2048 JPEG

4

1

63.7

63.5

64.7

65.2

15.7

FP8

image passage

2048x2048 JPEG

4

2

120.5

120.7

122.4

124.4

16.7

FP8

image passage

2048x2048 JPEG

4

3

180.2

181.8

183.8

184.9

16.9

FP8

image passage

2048x2048 JPEG

4

5

298.1

303.7

306.0

306.4

17.4

FP8

image passage

2048x2048 JPEG

4

8

471.1

486.3

488.7

489.3

18.2

FP8

image passage

2048x2048 JPEG

8

1

63.4

63.3

63.9

64.2

15.8

FP8

image passage

2048x2048 JPEG

8

2

119.6

120.0

120.7

121.5

16.8

FP8

image passage

2048x2048 JPEG

8

3

180.1

181.1

183.4

184.1

16.9

FP8

image passage

2048x2048 JPEG

8

5

297.7

302.6

309.0

309.9

17.5

FP8

image passage

2048x2048 JPEG

8

8

469.8

485.5

487.6

498.2

18.6

NVIDIA A100 SXM4 80GB#

The following table contains the performance data for Llama Nemotron Embed VL 1B v2 on NVIDIA A100 SXM4 80GB.

Precision

Input Type

Image Size

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP8

text

20

1

1

3.5

3.4

3.8

3.8

229.4

FP8

text

20

1

3

5.7

5.8

6.2

6.2

439.6

FP8

text

20

1

5

6.7

6.7

7.1

7.4

624.6

FP8

text

20

1

7

6.7

6.6

8.2

8.4

872.2

FP8

text

20

1

9

9.8

10.0

10.5

10.6

773.6

FP8

text

20

1

11

10.1

10.1

11.1

12.2

923.5

FP8

text

20

1

13

9.0

9.0

10.9

11.3

1170.8

FP8

text

20

1

15

11.5

11.7

12.6

13.8

1073.4

FP8

text

300

64

1

382.2

382.1

384.0

384.7

163.4

FP8

text

300

64

3

1100.1

1110.5

1114.5

1115.5

171.2

FP8

text

300

64

5

1822.5

1859.4

1863.0

1863.9

171.2

FP8

text

512

64

1

840.2

839.6

848.1

849.3

75.2

FP8

text

512

64

3

2459.8

2484.6

2496.9

2498.0

77.0

FP8

text

512

64

5

4064.9

4146.5

4163.6

4167.8

77.0

FP8

image passage

336x336 JPEG

1

1

34.5

34.3

34.7

35.5

29.0

FP8

image passage

336x336 JPEG

1

2

64.9

64.8

66.7

68.2

30.9

FP8

image passage

336x336 JPEG

1

3

97.3

97.6

100.3

102.8

31.2

FP8

image passage

336x336 JPEG

1

5

161.7

163.2

169.7

171.1

32.2

FP8

image passage

336x336 JPEG

1

8

254.7

261.6

268.2

269.5

33.9

FP8

image passage

336x336 JPEG

4

1

33.8

33.5

35.2

35.7

29.6

FP8

image passage

336x336 JPEG

4

2

65.6

65.4

67.8

68.7

30.7

FP8

image passage

336x336 JPEG

4

3

97.8

98.1

100.9

101.5

31.1

FP8

image passage

336x336 JPEG

4

5

162.1

164.3

168.2

169.5

32.1

FP8

image passage

336x336 JPEG

4

8

255.1

264.1

266.1

267.5

34.4

FP8

image passage

336x336 JPEG

8

1

33.6

33.3

34.5

35.1

29.8

FP8

image passage

336x336 JPEG

8

2

65.6

65.4

67.2

67.9

30.6

FP8

image passage

336x336 JPEG

8

3

98.2

98.7

101.0

101.4

31.0

FP8

image passage

336x336 JPEG

8

5

162.4

165.6

167.4

167.7

32.1

FP8

image passage

336x336 JPEG

8

8

255.5

264.8

265.5

265.8

34.2

FP8

image passage

512x512 JPEG

1

1

34.7

34.5

35.6

36.2

28.8

FP8

image passage

512x512 JPEG

1

2

65.1

64.7

67.3

68.9

30.9

FP8

image passage

512x512 JPEG

1

3

97.9

97.5

102.7

109.0

31.1

FP8

image passage

512x512 JPEG

1

5

161.9

163.1

169.5

171.3

31.9

FP8

image passage

512x512 JPEG

1

8

255.1

261.4

274.1

276.0

34.2

FP8

image passage

512x512 JPEG

4

1

33.9

33.7

34.8

35.2

29.5

FP8

image passage

512x512 JPEG

4

2

65.3

65.2

66.9

67.4

30.8

FP8

image passage

512x512 JPEG

4

3

97.8

98.2

100.9

101.6

31.1

FP8

image passage

512x512 JPEG

4

5

162.0

164.5

167.7

168.0

32.2

FP8

image passage

512x512 JPEG

4

8

256.9

265.9

269.7

270.0

34.1

FP8

image passage

512x512 JPEG

8

1

33.7

33.5

34.6

35.0

29.7

FP8

image passage

512x512 JPEG

8

2

65.5

65.4

67.2

67.3

30.7

FP8

image passage

512x512 JPEG

8

3

98.0

98.5

100.3

100.8

31.0

FP8

image passage

512x512 JPEG

8

5

162.4

165.6

167.4

167.6

32.1

FP8

image passage

512x512 JPEG

8

8

255.4

264.1

265.5

265.6

34.1

FP8

image passage

1024x1024 JPEG

1

1

181.2

180.5

182.2

186.8

5.5

FP8

image passage

1024x1024 JPEG

1

2

359.8

360.3

368.4

369.0

5.6

FP8

image passage

1024x1024 JPEG

1

3

538.6

541.1

552.1

552.9

5.7

FP8

image passage

1024x1024 JPEG

1

5

888.7

904.5

914.1

914.6

5.9

FP8

image passage

1024x1024 JPEG

1

8

1400.0

1451.2

1458.1

1458.5

6.3

FP8

image passage

1024x1024 JPEG

4

1

181.5

181.1

183.0

184.1

5.5

FP8

image passage

1024x1024 JPEG

4

2

361.0

363.3

364.1

364.4

5.6

FP8

image passage

1024x1024 JPEG

4

3

539.2

544.4

546.7

547.0

5.6

FP8

image passage

1024x1024 JPEG

4

5

890.1

908.6

910.1

910.4

5.9

FP8

image passage

1024x1024 JPEG

4

8

1402.7

1453.7

1454.6

1454.9

6.3

FP8

image passage

1024x1024 JPEG

8

1

182.0

181.6

183.1

183.6

5.5

FP8

image passage

1024x1024 JPEG

8

2

361.2

362.8

364.1

364.7

5.6

FP8

image passage

1024x1024 JPEG

8

3

539.3

544.6

545.6

545.9

5.6

FP8

image passage

1024x1024 JPEG

8

5

890.1

908.2

909.1

909.4

5.9

FP8

image passage

1024x1024 JPEG

8

8

1403.2

1454.2

1454.7

1454.9

6.3

FP8

image passage

2048x2048 JPEG

1

1

181.5

180.5

183.6

186.9

5.5

FP8

image passage

2048x2048 JPEG

1

2

359.2

359.5

366.5

369.9

5.6

FP8

image passage

2048x2048 JPEG

1

3

536.7

540.4

548.6

550.4

5.7

FP8

image passage

2048x2048 JPEG

1

5

1006.9

910.8

933.4

1871.2

5.6

FP8

image passage

2048x2048 JPEG

1

8

1400.5

1450.3

1456.7

1457.6

6.3

FP8

image passage

2048x2048 JPEG

4

1

183.7

183.3

185.3

186.1

5.4

FP8

image passage

2048x2048 JPEG

4

2

360.4

362.2

363.3

363.9

5.6

FP8

image passage

2048x2048 JPEG

4

3

538.4

543.7

545.1

545.8

5.7

FP8

image passage

2048x2048 JPEG

4

5

889.5

907.7

908.7

908.9

5.9

FP8

image passage

2048x2048 JPEG

4

8

1401.9

1452.9

1454.0

1454.5

6.3

FP8

image passage

2048x2048 JPEG

8

1

184.0

183.7

185.4

185.9

5.4

FP8

image passage

2048x2048 JPEG

8

2

360.4

362.1

363.1

363.2

5.6

FP8

image passage

2048x2048 JPEG

8

3

538.6

543.8

545.0

545.3

5.7

FP8

image passage

2048x2048 JPEG

8

5

889.5

907.5

908.5

909.2

5.9

FP8

image passage

2048x2048 JPEG

8

8

1402.2

1453.3

1454.1

1454.3

6.3

NVIDIA L40S#

The following table contains the performance data for Llama Nemotron Embed VL 1B v2 on NVIDIA L40S.

Precision

Input Type

Image Size

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP8

text

20

1

1

4.8

4.8

4.9

4.9

157.6

FP8

text

20

1

3

6.6

6.7

6.9

7.0

361.0

FP8

text

20

1

5

8.6

8.7

9.0

9.0

480.0

FP8

text

20

1

7

6.8

6.9

7.0

7.0

796.8

FP8

text

20

1

9

8.9

9.0

9.4

9.6

808.3

FP8

text

20

1

11

7.9

8.0

8.3

9.2

1124.1

FP8

text

20

1

13

8.2

8.5

9.5

9.8

1256.2

FP8

text

20

1

15

8.8

8.8

9.8

9.9

1354.3

FP8

text

300

64

1

233.0

232.9

233.7

233.9

267.0

FP8

text

300

64

3

675.3

681.5

685.2

687.5

278.7

FP8

text

300

64

5

1117.7

1139.9

1145.5

1147.4

278.9

FP8

text

512

64

1

394.7

394.3

399.1

399.5

159.3

FP8

text

512

64

3

1164.2

1175.8

1181.4

1183.7

162.3

FP8

text

512

64

5

1927.5

1967.5

1974.9

1976.5

162.1

FP8

image passage

336x336 JPEG

1

1

26.6

26.6

26.7

26.7

37.6

FP8

image passage

336x336 JPEG

1

2

45.2

45.3

45.5

45.7

44.4

FP8

image passage

336x336 JPEG

1

3

68.2

68.8

68.9

68.9

44.4

FP8

image passage

336x336 JPEG

1

5

113.7

115.8

115.9

116.0

45.5

FP8

image passage

336x336 JPEG

1

8

180.1

186.3

186.4

186.5

48.1

FP8

image passage

336x336 JPEG

4

1

24.4

24.4

24.4

24.5

41.0

FP8

image passage

336x336 JPEG

4

2

46.3

46.5

46.6

46.7

43.3

FP8

image passage

336x336 JPEG

4

3

69.4

70.1

70.1

70.1

43.8

FP8

image passage

336x336 JPEG

4

5

114.9

117.1

117.2

117.2

45.3

FP8

image passage

336x336 JPEG

4

8

181.4

187.8

188.2

189.3

48.2

FP8

image passage

336x336 JPEG

8

1

24.1

24.1

24.1

24.1

41.5

FP8

image passage

336x336 JPEG

8

2

46.6

46.8

46.8

46.9

43.1

FP8

image passage

336x336 JPEG

8

3

69.7

70.4

70.4

70.5

43.6

FP8

image passage

336x336 JPEG

8

5

115.2

117.5

117.5

117.5

45.2

FP8

image passage

336x336 JPEG

8

8

181.6

188.1

188.2

188.2

48.2

FP8

image passage

512x512 JPEG

1

1

26.7

26.7

26.8

26.9

37.4

FP8

image passage

512x512 JPEG

1

2

45.2

45.3

45.5

45.5

44.4

FP8

image passage

512x512 JPEG

1

3

68.2

68.8

69.0

69.0

44.4

FP8

image passage

512x512 JPEG

1

5

113.8

115.9

116.0

116.1

45.5

FP8

image passage

512x512 JPEG

1

8

180.2

186.4

186.5

186.6

48.1

FP8

image passage

512x512 JPEG

4

1

24.5

24.5

24.6

24.6

40.8

FP8

image passage

512x512 JPEG

4

2

46.3

46.5

46.6

46.6

43.3

FP8

image passage

512x512 JPEG

4

3

69.4

70.1

70.1

70.2

43.8

FP8

image passage

512x512 JPEG

4

5

115.1

117.1

117.4

118.7

45.2

FP8

image passage

512x512 JPEG

4

8

181.3

187.8

187.9

187.9

48.2

FP8

image passage

512x512 JPEG

8

1

24.2

24.2

24.3

24.3

41.3

FP8

image passage

512x512 JPEG

8

2

46.6

46.8

46.8

46.8

43.1

FP8

image passage

512x512 JPEG

8

3

69.7

70.3

70.4

70.4

43.6

FP8

image passage

512x512 JPEG

8

5

115.3

117.4

118.0

118.3

45.2

FP8

image passage

512x512 JPEG

8

8

182.0

188.1

188.7

192.2

48.1

FP8

image passage

1024x1024 JPEG

1

1

123.1

123.1

123.7

123.8

8.1

FP8

image passage

1024x1024 JPEG

1

2

236.1

237.2

237.9

238.1

8.5

FP8

image passage

1024x1024 JPEG

1

3

353.7

357.2

358.0

358.2

8.6

FP8

image passage

1024x1024 JPEG

1

5

584.8

596.5

597.9

598.1

8.9

FP8

image passage

1024x1024 JPEG

1

8

922.0

954.9

956.9

957.1

9.5

FP8

image passage

1024x1024 JPEG

4

1

121.0

121.0

121.2

121.3

8.3

FP8

image passage

1024x1024 JPEG

4

2

237.8

239.1

239.3

239.4

8.4

FP8

image passage

1024x1024 JPEG

4

3

355.6

359.1

359.3

360.3

8.6

FP8

image passage

1024x1024 JPEG

4

5

586.9

598.8

599.1

599.5

8.9

FP8

image passage

1024x1024 JPEG

4

8

924.9

958.5

958.9

960.6

9.5

FP8

image passage

1024x1024 JPEG

8

1

120.8

120.8

121.0

121.0

8.3

FP8

image passage

1024x1024 JPEG

8

2

238.2

239.4

239.5

239.5

8.4

FP8

image passage

1024x1024 JPEG

8

3

355.6

359.3

359.4

359.5

8.6

FP8

image passage

1024x1024 JPEG

8

5

587.6

599.3

601.1

601.9

8.9

FP8

image passage

1024x1024 JPEG

8

8

925.4

959.1

959.3

959.7

9.5

FP8

image passage

2048x2048 JPEG

1

1

125.3

125.1

125.8

125.9

8.0

FP8

image passage

2048x2048 JPEG

1

2

236.4

237.6

238.0

238.0

8.5

FP8

image passage

2048x2048 JPEG

1

3

354.0

357.5

358.1

358.2

8.6

FP8

image passage

2048x2048 JPEG

1

5

585.4

597.3

597.9

597.9

8.9

FP8

image passage

2048x2048 JPEG

1

8

923.9

957.0

958.0

958.2

9.5

FP8

image passage

2048x2048 JPEG

4

1

122.4

122.5

122.7

122.9

8.2

FP8

image passage

2048x2048 JPEG

4

2

237.7

238.9

239.1

239.1

8.5

FP8

image passage

2048x2048 JPEG

4

3

355.3

358.9

359.1

359.1

8.6

FP8

image passage

2048x2048 JPEG

4

5

587.2

598.8

599.1

601.8

8.9

FP8

image passage

2048x2048 JPEG

4

8

925.5

958.6

959.7

964.1

9.5

FP8

image passage

2048x2048 JPEG

8

1

121.9

121.8

122.4

122.5

8.2

FP8

image passage

2048x2048 JPEG

8

2

238.0

239.0

239.2

239.3

8.4

FP8

image passage

2048x2048 JPEG

8

3

355.5

359.0

359.2

359.3

8.6

FP8

image passage

2048x2048 JPEG

8

5

587.0

598.8

599.0

599.0

8.9

FP8

image passage

2048x2048 JPEG

8

8

925.3

958.7

958.8

958.9

9.5

NVIDIA L4#

The following table contains the performance data for Llama Nemotron Embed VL 1B v2 on NVIDIA L4.

Precision

Input Type

Image Size

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP8

text

20

1

1

10.2

10.1

10.2

10.3

88.8

FP8

text

20

1

3

17.6

18.1

18.5

18.6

157.7

FP8

text

20

1

5

19.5

19.2

19.9

21.6

236.2

FP8

text

20

1

7

20.1

20.2

21.2

21.5

319.3

FP8

text

20

1

9

21.1

21.3

22.5

22.7

380.9

FP8

text

20

1

11

22.4

22.7

23.9

24.4

436.0

FP8

text

20

1

13

31.4

32.2

41.5

42.1

378.3

FP8

text

20

1

15

37.3

36.8

46.2

46.9

366.9

FP8

text

300

64

1

1096.9

1098.2

1106.0

1108.7

57.9

FP8

text

300

64

3

3239.8

3271.2

3291.7

3293.9

58.5

FP8

text

300

64

5

5352.3

5464.2

5478.1

5481.6

58.5

FP8

text

512

64

1

2159.2

2157.8

2182.8

2191.2

29.5

FP8

text

512

64

3

6405.2

6471.2

6501.6

6507.9

29.6

FP8

text

512

64

5

10591.4

10810.3

10853.6

10862.4

29.6

FP8

image passage

336x336 JPEG

1

1

85.3

85.0

87.0

87.1

11.7

FP8

image passage

336x336 JPEG

1

2

167.1

168.0

169.5

169.9

12.0

FP8

image passage

336x336 JPEG

1

3

250.5

252.5

254.2

254.6

12.2

FP8

image passage

336x336 JPEG

1

5

412.9

421.1

423.1

423.9

12.6

FP8

image passage

336x336 JPEG

1

8

655.7

679.3

681.0

681.7

13.4

FP8

image passage

336x336 JPEG

4

1

85.2

85.2

85.5

85.7

11.7

FP8

image passage

336x336 JPEG

4

2

168.6

169.4

169.7

169.9

11.9

FP8

image passage

336x336 JPEG

4

3

251.8

254.3

254.6

254.7

12.1

FP8

image passage

336x336 JPEG

4

5

416.1

424.4

425.0

425.8

12.5

FP8

image passage

336x336 JPEG

4

8

655.8

679.5

680.1

680.3

13.4

FP8

image passage

336x336 JPEG

8

1

85.3

85.3

85.5

85.7

11.7

FP8

image passage

336x336 JPEG

8

2

169.6

170.5

170.8

171.0

11.9

FP8

image passage

336x336 JPEG

8

3

252.4

254.9

255.4

255.5

12.1

FP8

image passage

336x336 JPEG

8

5

416.7

425.2

425.7

425.9

12.5

FP8

image passage

336x336 JPEG

8

8

658.8

682.8

683.6

683.8

13.3

FP8

image passage

512x512 JPEG

1

1

86.6

86.5

88.0

88.6

11.6

FP8

image passage

512x512 JPEG

1

2

167.4

168.3

170.0

170.3

12.0

FP8

image passage

512x512 JPEG

1

3

250.2

252.6

254.3

254.5

12.2

FP8

image passage

512x512 JPEG

1

5

414.1

422.4

424.1

424.4

12.6

FP8

image passage

512x512 JPEG

1

8

653.6

677.0

678.4

678.8

13.4

FP8

image passage

512x512 JPEG

4

1

85.4

85.4

85.8

85.9

11.7

FP8

image passage

512x512 JPEG

4

2

168.4

169.3

169.5

169.6

11.9

FP8

image passage

512x512 JPEG

4

3

251.7

254.3

254.6

254.8

12.1

FP8

image passage

512x512 JPEG

4

5

415.7

424.1

424.7

425.0

12.5

FP8

image passage

512x512 JPEG

4

8

655.3

679.0

679.5

679.6

13.4

FP8

image passage

512x512 JPEG

8

1

86.1

86.1

86.5

86.6

11.6

FP8

image passage

512x512 JPEG

8

2

168.8

169.6

169.9

170.0

11.9

FP8

image passage

512x512 JPEG

8

3

252.2

254.7

255.1

255.2

12.1

FP8

image passage

512x512 JPEG

8

5

417.7

426.4

426.8

427.1

12.5

FP8

image passage

512x512 JPEG

8

8

656.9

680.7

682.4

682.6

13.4

FP8

image passage

1024x1024 JPEG

1

1

457.4

457.6

458.9

459.3

2.2

FP8

image passage

1024x1024 JPEG

1

2

902.1

906.8

909.6

909.8

2.2

FP8

image passage

1024x1024 JPEG

1

3

1348.7

1362.4

1365.4

1365.8

2.3

FP8

image passage

1024x1024 JPEG

1

5

2223.0

2268.0

2272.1

2272.9

2.3

FP8

image passage

1024x1024 JPEG

1

8

3508.4

3634.6

3641.6

3643.4

2.5

FP8

image passage

1024x1024 JPEG

4

1

456.6

456.5

457.6

457.7

2.2

FP8

image passage

1024x1024 JPEG

4

2

903.8

908.1

909.6

910.0

2.2

FP8

image passage

1024x1024 JPEG

4

3

1348.2

1361.7

1363.1

1363.5

2.3

FP8

image passage

1024x1024 JPEG

4

5

2225.1

2270.6

2272.7

2273.1

2.3

FP8

image passage

1024x1024 JPEG

4

8

3506.3

3633.0

3636.6

3637.4

2.5

FP8

image passage

1024x1024 JPEG

8

1

456.0

456.0

456.6

456.7

2.2

FP8

image passage

1024x1024 JPEG

8

2

903.3

907.9

908.7

908.9

2.2

FP8

image passage

1024x1024 JPEG

8

3

1349.3

1362.8

1364.2

1364.4

2.3

FP8

image passage

1024x1024 JPEG

8

5

2226.2

2271.5

2273.4

2273.7

2.3

FP8

image passage

1024x1024 JPEG

8

8

3507.6

3634.5

3637.3

3637.8

2.5

FP8

image passage

2048x2048 JPEG

1

1

459.5

459.2

461.4

461.8

2.2

FP8

image passage

2048x2048 JPEG

1

2

899.8

904.0

907.0

907.5

2.2

FP8

image passage

2048x2048 JPEG

1

3

1345.5

1358.8

1361.8

1362.2

2.3

FP8

image passage

2048x2048 JPEG

1

5

2219.2

2264.1

2268.1

2269.1

2.3

FP8

image passage

2048x2048 JPEG

1

8

3500.3

3626.5

3632.9

3633.8

2.5

FP8

image passage

2048x2048 JPEG

4

1

457.0

457.0

457.9

458.2

2.2

FP8

image passage

2048x2048 JPEG

4

2

902.4

906.8

908.2

908.8

2.2

FP8

image passage

2048x2048 JPEG

4

3

1346.7

1360.3

1362.2

1362.3

2.3

FP8

image passage

2048x2048 JPEG

4

5

2224.7

2269.9

2272.8

2273.3

2.3

FP8

image passage

2048x2048 JPEG

4

8

3505.9

3632.6

3636.4

3636.7

2.5

FP8

image passage

2048x2048 JPEG

8

1

456.6

456.6

457.5

457.7

2.2

FP8

image passage

2048x2048 JPEG

8

2

903.0

907.6

908.7

908.8

2.2

FP8

image passage

2048x2048 JPEG

8

3

1347.9

1361.6

1362.8

1363.0

2.3

FP8

image passage

2048x2048 JPEG

8

5

2225.5

2270.8

2272.4

2272.7

2.3

FP8

image passage

2048x2048 JPEG

8

8

3503.6

3630.4

3633.7

3634.6

2.5

NVIDIA A10G#

The following table contains the performance data for Llama Nemotron Embed VL 1B v2 on NVIDIA A10G.

Precision

Input Type

Image Size

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP8

text

20

1

1

6.9

7.0

7.2

7.3

124.9

FP8

text

20

1

3

10.9

10.9

11.2

11.3

249.3

FP8

text

20

1

5

11.8

11.7

12.3

12.6

380.2

FP8

text

20

1

7

17.6

17.6

19.5

19.7

357.7

FP8

text

20

1

9

14.9

15.1

15.6

16.9

532.9

FP8

text

20

1

11

16.5

16.4

17.8

19.1

586.1

FP8

text

20

1

13

21.7

23.0

24.0

24.2

534.9

FP8

text

20

1

15

21.4

22.1

23.5

23.6

616.3

FP8

text

300

64

1

951.5

951.6

954.7

955.1

66.6

FP8

text

300

64

3

2787.7

2814.7

2822.9

2824.8

67.9

FP8

text

300

64

5

4606.7

4700.2

4706.6

4708.7

67.9

FP8

text

512

64

1

1806.0

1804.9

1819.4

1821.3

35.2

FP8

text

512

64

3

5322.8

5377.1

5394.6

5397.8

35.6

FP8

text

512

64

5

8788.5

8965.8

8996.8

9001.3

35.6

FP8

image passage

336x336 JPEG

1

1

72.7

72.7

72.9

72.9

13.8

FP8

image passage

336x336 JPEG

1

2

138.0

138.6

138.8

139.0

14.6

FP8

image passage

336x336 JPEG

1

3

206.6

208.6

208.9

209.0

14.7

FP8

image passage

336x336 JPEG

1

5

342.0

348.4

349.1

352.5

15.2

FP8

image passage

336x336 JPEG

1

8

539.0

558.3

558.6

558.7

16.2

FP8

image passage

336x336 JPEG

4

1

70.9

70.9

71.0

71.0

14.1

FP8

image passage

336x336 JPEG

4

2

138.9

139.4

139.6

139.8

14.5

FP8

image passage

336x336 JPEG

4

3

207.3

209.3

209.5

209.6

14.7

FP8

image passage

336x336 JPEG

4

5

342.6

349.2

349.5

351.1

15.2

FP8

image passage

336x336 JPEG

4

8

539.9

559.1

559.5

562.8

16.2

FP8

image passage

336x336 JPEG

8

1

70.6

70.5

70.6

70.7

14.2

FP8

image passage

336x336 JPEG

8

2

138.9

139.5

139.6

139.7

14.5

FP8

image passage

336x336 JPEG

8

3

207.6

209.5

209.7

210.9

14.7

FP8

image passage

336x336 JPEG

8

5

342.6

349.4

349.5

350.5

15.2

FP8

image passage

336x336 JPEG

8

8

540.0

559.3

559.4

561.9

16.2

FP8

image passage

512x512 JPEG

1

1

73.2

72.9

73.6

74.0

13.7

FP8

image passage

512x512 JPEG

1

2

137.7

138.4

138.6

138.8

14.6

FP8

image passage

512x512 JPEG

1

3

206.9

208.3

208.9

209.3

14.7

FP8

image passage

512x512 JPEG

1

5

341.6

348.3

349.0

349.7

15.2

FP8

image passage

512x512 JPEG

1

8

538.8

558.0

558.4

558.5

16.2

FP8

image passage

512x512 JPEG

4

1

71.1

71.1

71.3

71.4

14.1

FP8

image passage

512x512 JPEG

4

2

138.8

139.3

139.5

139.5

14.5

FP8

image passage

512x512 JPEG

4

3

207.3

209.3

209.5

209.6

14.7

FP8

image passage

512x512 JPEG

4

5

342.3

349.2

349.4

349.5

15.2

FP8

image passage

512x512 JPEG

4

8

539.6

559.0

559.3

559.3

16.2

FP8

image passage

512x512 JPEG

8

1

70.7

70.7

70.8

70.8

14.1

FP8

image passage

512x512 JPEG

8

2

138.8

139.5

139.6

139.6

14.5

FP8

image passage

512x512 JPEG

8

3

207.4

209.5

209.6

209.8

14.7

FP8

image passage

512x512 JPEG

8

5

342.4

349.3

349.5

350.3

15.2

FP8

image passage

512x512 JPEG

8

8

539.9

559.2

559.4

561.0

16.2

FP8

image passage

1024x1024 JPEG

1

1

371.7

371.6

372.1

372.5

2.7

FP8

image passage

1024x1024 JPEG

1

2

729.0

732.5

733.5

733.9

2.8

FP8

image passage

1024x1024 JPEG

1

3

1089.1

1099.5

1101.5

1103.1

2.8

FP8

image passage

1024x1024 JPEG

1

5

1799.1

1835.2

1837.1

1838.6

2.9

FP8

image passage

1024x1024 JPEG

1

8

2834.2

2935.6

2940.2

2942.3

3.1

FP8

image passage

1024x1024 JPEG

4

1

369.2

369.1

369.5

369.8

2.7

FP8

image passage

1024x1024 JPEG

4

2

729.7

733.1

733.9

734.4

2.8

FP8

image passage

1024x1024 JPEG

4

3

1089.9

1100.5

1102.4

1103.0

2.8

FP8

image passage

1024x1024 JPEG

4

5

1799.1

1835.3

1836.5

1840.2

2.9

FP8

image passage

1024x1024 JPEG

4

8

2834.4

2937.1

2937.9

2938.1

3.1

FP8

image passage

1024x1024 JPEG

8

1

368.8

368.6

369.2

370.0

2.7

FP8

image passage

1024x1024 JPEG

8

2

730.0

733.4

734.4

734.8

2.8

FP8

image passage

1024x1024 JPEG

8

3

1090.3

1101.0

1102.7

1103.2

2.8

FP8

image passage

1024x1024 JPEG

8

5

1798.8

1835.3

1836.2

1836.4

2.9

FP8

image passage

1024x1024 JPEG

8

8

2834.7

2937.2

2938.3

2938.7

3.1

FP8

image passage

2048x2048 JPEG

1

1

374.0

373.7

375.3

376.5

2.7

FP8

image passage

2048x2048 JPEG

1

2

729.3

732.7

733.7

734.4

2.8

FP8

image passage

2048x2048 JPEG

1

3

1089.6

1100.0

1102.5

1104.6

2.8

FP8

image passage

2048x2048 JPEG

1

5

1801.1

1837.5

1839.2

1839.5

2.9

FP8

image passage

2048x2048 JPEG

1

8

2837.4

2939.5

2941.3

2945.5

3.1

FP8

image passage

2048x2048 JPEG

4

1

371.0

370.9

371.6

372.2

2.7

FP8

image passage

2048x2048 JPEG

4

2

733.2

736.8

740.3

740.8

2.7

FP8

image passage

2048x2048 JPEG

4

3

1090.2

1101.1

1102.0

1102.2

2.8

FP8

image passage

2048x2048 JPEG

4

5

1800.8

1837.0

1839.1

1840.4

2.9

FP8

image passage

2048x2048 JPEG

4

8

2838.2

2940.6

2942.0

2942.3

3.1

FP8

image passage

2048x2048 JPEG

8

1

370.3

370.2

370.9

371.1

2.7

FP8

image passage

2048x2048 JPEG

8

2

730.6

734.2

734.6

735.4

2.8

FP8

image passage

2048x2048 JPEG

8

3

1090.9

1101.6

1103.2

1104.6

2.8

FP8

image passage

2048x2048 JPEG

8

5

1800.8

1837.2

1838.4

1838.9

2.9

FP8

image passage

2048x2048 JPEG

8

8

2838.2

2940.7

2942.0

2942.3

3.1

Llama Nemotron Embed 1B v2#

Memory Footprint#

The following table contains the memory footprint data for Llama Nemotron Embed 1B v2.

Compute Capability

Precision

Approximate GPU Memory Size (GiB)

12.0

fp8

2.98

12.0

fp16

6.53

10.0

fp8

2.98

10.0

fp16

6.09

9.0

fp8

2.98

9.0

fp16

4.91

8.9

fp8

2.98

8.9

fp16

4.91

8.6

fp16

5.22

8.0

fp16

5.09

NVIDIA H100 80GB HBM3#

The following table contains the performance data for Llama Nemotron Embed 1B v2 on NVIDIA H100 80GB HBM3.

Precision

Input Type

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP8

passage

300

64

1

159

159

162

163

401.8

FP8

passage

300

64

3

267

269

275

277

709.4

FP8

passage

300

64

5

392

390

401

411

814.5

FP8

passage

512

64

1

222

218

233

235

286.3

FP8

passage

512

64

3

431

430

450

455

440.1

FP8

passage

512

64

5

615

617

694

701

504.3

FP8

query

20

1

1

7

7

8

8

140.7

FP8

query

20

1

3

10

10

10

10

291.7

FP8

query

20

1

5

16

17

17

17

307.0

FP8

query

20

1

7

22

23

24

24

310.9

FP8

query

20

1

9

25

26

31

31

329.7

FP8

query

20

1

11

30

32

38

38

332.8

FP8

query

20

1

13

36

37

43

44

340.0

FP8

query

20

1

15

42

43

48

50

327.5

FP16

passage

300

64

1

186

186

190

192

342.7

FP16

passage

300

64

3

345

346

354

356

550.7

FP16

passage

300

64

5

525

524

535

544

608.3

FP16

passage

512

64

1

269

265

280

284

237.2

FP16

passage

512

64

3

563

564

581

584

337.3

FP16

passage

512

64

5

846

861

946

956

366.0

FP16

query

20

1

1

8

8

8

9

125.6

FP16

query

20

1

3

12

12

13

13

246.1

FP16

query

20

1

5

20

21

22

22

247.9

FP16

query

20

1

7

25

26

30

30

267.5

FP16

query

20

1

9

34

34

38

39

251.5

FP16

query

20

1

11

43

46

47

48

237.7

FP16

query

20

1

13

46

49

53

54

260.9

FP16

query

20

1

15

55

58

65

65

248.4

NVIDIA A100 SXM4 80GB#

The following table contains the performance data for Llama Nemotron Embed 1B v2 on NVIDIA A100 SXM4 80GB.

Precision

Input Type

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP16

passage

300

64

1

325

324

331

333

196.9

FP16

passage

300

64

3

699

699

717

723

272.1

FP16

passage

300

64

5

1095

1093

1109

1113

291.8

FP16

passage

512

64

1

496

497

502

518

128.8

FP16

passage

512

64

3

1165

1169

1199

1209

163.2

FP16

passage

512

64

5

1872

1870

1904

1916

170.5

FP16

query

20

1

1

8

8

9

10

122.3

FP16

query

20

1

3

11

12

14

14

252.9

FP16

query

20

1

5

20

19

23

24

250.9

FP16

query

20

1

7

27

27

31

32

251.6

FP16

query

20

1

9

30

30

37

42

280.1

FP16

query

20

1

11

41

42

47

51

250.9

FP16

query

20

1

13

46

49

53

57

262.0

FP16

query

20

1

15

50

51

60

63

269.7

NVIDIA A100 PCIe 80GB#

The following table contains the performance data for Llama Nemotron Embed 1B v2 on NVIDIA A100 PCIe 80GB.

Precision

Input Type

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP16

passage

300

64

1

362

362

367

369

176.8

FP16

passage

300

64

3

786

789

803

812

241.9

FP16

passage

300

64

5

1240

1239

1258

1262

257.6

FP16

passage

512

64

1

550

548

565

570

116.2

FP16

passage

512

64

3

1327

1335

1354

1359

143.4

FP16

passage

512

64

5

2145

2145

2180

2187

149.0

FP16

query

20

1

1

10

9

10

10

104.7

FP16

query

20

1

3

12

12

12

12

246.7

FP16

query

20

1

5

18

19

20

20

276.7

FP16

query

20

1

7

25

24

28

28

268.9

FP16

query

20

1

9

30

31

36

36

291.4

FP16

query

20

1

11

38

39

44

44

270.3

FP16

query

20

1

13

43

44

52

52

272.4

FP16

query

20

1

15

49

52

56

57

276.4

NVIDIA L40S#

The following table contains the performance data for Llama Nemotron Embed 1B v2 on NVIDIA L40S.

Precision

Input Type

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP8

passage

300

64

1

277

278

280

281

230.3

FP8

passage

300

64

3

615

617

629

632

309.4

FP8

passage

300

64

5

976

975

983

987

327.4

FP8

passage

512

64

1

443

441

448

454

144.3

FP8

passage

512

64

3

1071

1077

1089

1090

177.6

FP8

passage

512

64

5

1736

1735

1752

1758

184.1

FP8

query

20

1

1

10

10

10

11

98.0

FP8

query

20

1

3

15

17

17

17

192.5

FP8

query

20

1

5

24

23

28

28

200.8

FP8

query

20

1

7

34

34

39

40

197.0

FP8

query

20

1

9

44

45

51

51

197.3

FP8

query

20

1

11

53

56

62

62

196.8

FP8

query

20

1

13

63

67

73

73

190.6

FP8

query

20

1

15

73

79

84

84

188.3

FP16

passage

300

64

1

350

350

353

354

182.8

FP16

passage

300

64

3

821

823

831

842

231.7

FP16

passage

300

64

5

1320

1317

1330

1346

242.1

FP16

passage

512

64

1

567

566

573

575

112.7

FP16

passage

512

64

3

1440

1448

1464

1475

132.2

FP16

passage

512

64

5

2370

2371

2391

2396

134.8

FP16

query

20

1

1

12

12

12

12

85.7

FP16

query

20

1

3

19

21

22

22

150.3

FP16

query

20

1

5

32

29

36

36

152.8

FP16

query

20

1

7

46

50

50

50

148.0

FP16

query

20

1

9

55

57

65

65

156.3

FP16

query

20

1

11

68

72

79

79

154.5

FP16

query

20

1

13

79

86

93

93

154.4

FP16

query

20

1

15

84

86

95

101

165.0

NVIDIA L4#

The following table contains the performance data for Llama Nemotron Embed 1B v2 on NVIDIA L4.

Precision

Input Type

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP8

passage

300

64

1

724

725

730

733

88.3

FP8

passage

300

64

3

1865

1876

1891

1892

102.1

FP8

passage

300

64

5

3117

3127

3147

3150

102.1

FP8

passage

512

64

1

1300

1300

1314

1318

49.2

FP8

passage

512

64

3

3551

3577

3606

3618

53.6

FP8

passage

512

64

5

5940

5968

6000

6007

53.6

FP8

query

20

1

1

13

13

14

15

74.2

FP8

query

20

1

3

21

23

23

23

138.0

FP8

query

20

1

5

32

31

39

39

152.2

FP8

query

20

1

7

46

46

53

54

145.8

FP8

query

20

1

9

56

61

69

69

150.3

FP8

query

20

1

11

67

69

76

77

154.0

FP8

query

20

1

13

86

91

98

99

141.2

FP8

query

20

1

15

98

105

113

114

141.0

FP16

passage

300

64

1

1011

1011

1019

1019

63.3

FP16

passage

300

64

3

2702

2716

2734

2754

70.6

FP16

passage

300

64

5

4527

4527

4550

4571

70.6

FP16

passage

512

64

1

1705

1710

1733

1740

37.5

FP16

passage

512

64

3

4770

4799

4860

4872

39.9

FP16

passage

512

64

5

8001

8025

8079

8085

39.9

FP16

query

20

1

1

18

18

18

19

55.8

FP16

query

20

1

3

33

37

38

38

87.7

FP16

query

20

1

5

55

62

63

63

87.4

FP16

query

20

1

7

71

75

88

88

94.6

FP16

query

20

1

9

96

101

114

114

90.3

FP16

query

20

1

11

115

126

139

139

89.8

FP16

query

20

1

13

131

138

164

164

91.4

FP16

query

20

1

15

154

157

189

189

88.4

NVIDIA A10G#

The following table contains the performance data for Llama Nemotron Embed 1B v2 on NVIDIA A10G.

Precision

Input Type

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP16

passage

300

64

1

888

888

899

899

72.1

FP16

passage

300

64

3

2272

2280

2329

2341

83.9

FP16

passage

300

64

5

3795

3801

3828

3840

84.2

FP16

passage

512

64

1

1451

1451

1471

1473

44.1

FP16

passage

512

64

3

3926

3947

3988

4009

48.6

FP16

passage

512

64

5

6577

6571

6632

6657

48.6

FP16

query

20

1

1

21

21

21

21

48.6

FP16

query

20

1

3

39

44

45

45

75.8

FP16

query

20

1

5

61

60

75

75

80.2

FP16

query

20

1

7

85

89

104

105

79.2

FP16

query

20

1

9

109

119

134

134

79.0

FP16

query

20

1

11

144

149

164

164

73.2

FP16

query

20

1

13

159

164

194

194

76.5

FP16

query

20

1

15

175

179

209

209

79.4

Llama Nemotron Embed 300m v2#

Memory Footprint#

The following table contains the memory footprint data for Llama Nemotron Embed 300m v2.

Compute Capability

Precision

Approximate GPU Memory Size (GiB)

12.0

fp8

2.04

12.0

fp16

3.53

10.0

fp8

2.04

10.0

fp16

3.19

9.0

fp8

2.04

9.0

fp16

3.17

8.9

fp8

2.04

8.9

fp16

2.67

8.6

fp16

2.86

8.0

fp16

2.86

NVIDIA H100 80GB HBM3#

The following table contains the performance data for Llama Nemotron Embed 300m v2 on NVIDIA H100 80GB HBM3.

Precision

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP8

20

1

1

7.78

8

8

8.33

126

FP8

20

1

3

8.29

8

9

8.77

348

FP8

20

1

5

8.75

8

9

10.35

547

FP8

20

1

7

10.85

11

12

11.86

585

FP8

20

1

9

13.18

13

15

15.14

643

FP8

20

1

11

15.53

16

18

18.35

652

FP8

20

1

13

17.54

17

21

21.46

673

FP8

20

1

15

20.59

21

24

24.02

662

FP8

300

64

1

69.49

69

71

72.27

896

FP8

300

64

3

103.50

104

107

108.34

1779

FP8

300

64

5

172.27

173

178

179.91

1799

FP8

512

64

1

103.64

101

116

120.12

606

FP8

512

64

3

180.56

182

192

193.55

1040

FP8

512

64

5

298.65

300

318

325.06

1041

FP16

20

1

1

8.02

8

9

8.93

122

FP16

20

1

3

8.58

8

9

9.17

340

FP16

20

1

5

8.79

9

9

9.22

548

FP16

20

1

7

10.91

11

12

12.59

575

FP16

20

1

9

13.33

14

15

15.16

638

FP16

20

1

11

15.36

16

17

17.92

668

FP16

20

1

13

17.60

17

21

21.27

688

FP16

20

1

15

21.16

22

24

24.85

645

FP16

300

64

1

70.88

71

73

74.33

880

FP16

300

64

3

112.79

113

117

118.36

1652

FP16

300

64

5

187.74

189

194

194.95

1653

FP16

512

64

1

109.49

108

128

128.95

574

FP16

512

64

3

202.49

203

216

219.39

919

FP16

512

64

5

336.37

338

352

356.86

926

NVIDIA H100 NVL#

The following table contains the performance data for Llama Nemotron Embed 300m v2 on NVIDIA H100 NVL.

Precision

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP8

20

1

1

8.18

8

9

8.63

120

FP8

20

1

3

8.47

8

9

9.15

345

FP8

20

1

5

9.41

9

10

10.85

508

FP8

20

1

7

11.40

11

13

13.37

576

FP8

20

1

9

14.38

15

17

16.93

584

FP8

20

1

11

16.63

16

20

20.25

602

FP8

20

1

13

20.25

21

24

24.18

590

FP8

20

1

15

22.61

23

26

26.69

600

FP8

300

64

1

68.78

68

72

73.08

907

FP8

300

64

3

119.34

120

125

126.38

1552

FP8

300

64

5

199.83

202

208

210.33

1555

FP8

512

64

1

112.73

112

125

131.93

559

FP8

512

64

3

229.52

232

242

246.59

814

FP8

512

64

5

383.51

391

401

406.96

813

FP16

20

1

1

8.16

8

8

8.70

120

FP16

20

1

3

8.31

8

9

8.83

339

FP16

20

1

5

8.73

9

9

9.28

535

FP16

20

1

7

10.72

11

12

11.83

597

FP16

20

1

9

13.20

13

15

15.20

642

FP16

20

1

11

15.52

16

18

18.31

654

FP16

20

1

13

18.43

19

21

21.44

650

FP16

20

1

15

20.82

21

24

24.46

668

FP16

300

64

1

78.69

79

82

82.79

795

FP16

300

64

3

149.74

151

156

157.84

1250

FP16

300

64

5

249.85

254

260

261.95

1246

FP16

512

64

1

129.64

127

142

150.17

487

FP16

512

64

3

283.14

286

299

302.77

660

FP16

512

64

5

474.03

482

499

502.43

658

NVIDIA A100 SXM4 80GB#

The following table contains the performance data for Llama Nemotron Embed 300m v2 on NVIDIA A100 SXM4 80GB.

Precision

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP16

20

1

1

5.53

5

7

6.98

180

FP16

20

1

3

7.01

7

7

7.13

421

FP16

20

1

5

10.47

11

12

11.85

455

FP16

20

1

7

12.63

12

16

16.39

529

FP16

20

1

9

18.47

20

21

21.29

466

FP16

20

1

11

19.63

19

23

25.16

524

FP16

20

1

13

24.63

25

28

30.14

481

FP16

20

1

15

26.42

28

30

30.31

509

FP16

300

64

1

115.68

115

120

120.39

544

FP16

300

64

3

229.73

231

236

239.21

817

FP16

300

64

5

381.63

385

393

397.63

817

FP16

512

64

1

199.57

195

237

242.43

317

FP16

512

64

3

431.60

433

447

448.86

438

FP16

512

64

5

715.00

724

740

746.65

437

NVIDIA L40S#

The following table contains the performance data for Llama Nemotron Embed 300m v2 on NVIDIA L40S.

Precision

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP8

20

1

1

8.31

8

8

8.66

118

FP8

20

1

3

8.96

9

9

9.50

320

FP8

20

1

5

9.95

10

11

11.45

476

FP8

20

1

7

11.51

12

14

13.74

573

FP8

20

1

9

14.40

14

16

16.49

583

FP8

20

1

11

17.76

18

21

21.69

518

FP8

20

1

13

20.21

20

24

23.74

582

FP8

20

1

15

20.48

21

24

24.73

664

FP8

300

64

1

97.32

97

99

100.20

646

FP8

300

64

3

199.96

202

204

204.93

940

FP8

300

64

5

332.69

336

340

340.51

938

FP8

512

64

1

199.93

189

261

268.05

317

FP8

512

64

3

415.89

419

427

428.23

455

FP8

512

64

5

690.07

701

711

712.80

453

FP16

20

1

1

8.58

9

9

9.05

114

FP16

20

1

3

9.29

9

10

10.16

312

FP16

20

1

5

10.49

10

11

11.16

459

FP16

20

1

7

14.37

14

15

21.76

453

FP16

20

1

9

17.65

18

20

20.21

476

FP16

20

1

11

20.24

21

25

24.91

499

FP16

20

1

13

24.52

24

27

28.41

480

FP16

20

1

15

28.73

30

33

33.26

471

FP16

300

64

1

124.10

124

126

127.18

508

FP16

300

64

3

273.10

275

278

279.24

692

FP16

300

64

5

451.25

458

463

463.71

692

FP16

512

64

1

239.63

230

297

300.99

265

FP16

512

64

3

529.88

534

541

542.72

357

FP16

512

64

5

876.35

892

900

901.77

357

NVIDIA L4#

The following table contains the performance data for Llama Nemotron Embed 300m v2 on NVIDIA L4.

Precision

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP16

20

1

1

8.77

8

10

10.01

113

FP16

20

1

3

11.12

11

12

11.77

259

FP16

20

1

5

16.35

16

19

18.90

295

FP16

20

1

7

23.27

22

26

26.77

290

FP16

20

1

9

29.48

30

33

33.85

289

FP16

20

1

11

34.20

34

40

40.63

302

FP16

20

1

13

41.65

44

47

48.09

288

FP16

20

1

15

45.74

48

55

55.23

300

FP16

300

64

1

339.59

339

343

345.46

187

FP16

300

64

3

918.26

927

933

934.58

207

FP16

300

64

5

1517.32

1547

1554

1555.52

206

FP16

512

64

1

642.27

643

657

661.90

99

FP16

512

64

3

1795.35

1818

1835

1841.23

105

FP16

512

64

5

2976.58

3034

3066

3075.53

105

NVIDIA A10G#

The following table contains the performance data for Llama Nemotron Embed 300m v2 on NVIDIA A10G.

Precision

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP16

20

1

1

9.60

9

10

10.98

103

FP16

20

1

3

12.57

13

14

14.24

232

FP16

20

1

5

20.57

22

23

23.06

237

FP16

20

1

7

28.32

31

32

31.69

234

FP16

20

1

9

35.08

36

41

41.84

244

FP16

20

1

11

43.33

45

50

50.38

236

FP16

20

1

13

52.65

54

59

58.80

234

FP16

20

1

15

57.97

60

68

68.68

234

FP16

300

64

1

304.86

305

309

310.32

209

FP16

300

64

3

780.64

791

799

800.03

241

FP16

300

64

5

1296.85

1320

1330

1331.81

241

FP16

512

64

1

520.74

519

533

538.60

122

FP16

512

64

3

1385.52

1404

1424

1428.40

137

FP16

512

64

5

2294.26

2341

2362

2368.31

137

NVIDIA QA E5 Embedding#

Memory Footprint#

The following table contains the memory footprint data for NVIDIA QA E5 Embedding.

Compute Capability

Precision

Approximate GPU Memory Size (GiB)

12.0

fp16

0.87

10.0

fp16

0.87

9.0

fp16

0.87

8.9

fp16

0.87

8.6

fp16

0.88

8.0

fp16

0.87

NVIDIA H100 80GB HBM3#

The following table contains the performance data for NVIDIA QA E5 Embedding on NVIDIA H100 80GB HBM3.

Precision

Input Type

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP16

passage

300

64

1

99.82703581

100.904438

107.924179

108.6330097

639.0343548

FP16

passage

300

64

3

143.8097451

143.2762855

156.6011532

159.0137593

1330.002429

FP16

passage

300

64

5

239.6826504

239.732372

259.140126

264.9664443

1330.975643

FP16

passage

512

64

1

114.5962003

114.388121

115.9079986

117.0048043

556.5479079

FP16

passage

512

64

3

170.1910781

169.862901

171.2404072

171.8325308

1124.195819

FP16

passage

512

64

5

284.5675255

284.516985

285.61744

286.059655

1121.352035

FP16

query

20

1

1

5.080810388

5.077455

5.35184

5.449206

196.2536704

FP16

query

20

1

3

6.005669286

5.5439535

7.449483

7.6433133

498.5059144

FP16

query

20

1

5

11.90364799

12.3012425

12.7891541

12.93557995

418.3266895

FP16

query

20

1

7

16.50119069

17.2286925

17.9718198

18.13005535

422.0442266

FP16

query

20

1

9

21.43597982

22.277524

23.323499

23.553763

418.3116944

FP16

query

20

1

11

26.03332591

25.9899645

28.3617327

28.5704181

421.3414301

FP16

query

20

1

13

30.71904992

30.885644

33.145684

33.59465

422.1799737

FP16

query

20

1

15

37.27936682

37.933293

39.0920083

39.3421219

401.4463907

NVIDIA A100 SXM4 80GB#

The following table contains the performance data for NVIDIA QA E5 Embedding on NVIDIA A100 SXM4 80GB.

Precision

Input Type

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP16

passage

300

64

1

176.5200289

177.123774

188.566657

190.397038

362.1932991

FP16

passage

300

64

3

336.1264122

336.965212

359.1192051

365.7920741

570.1898092

FP16

passage

300

64

5

560.1892229

562.919036

592.343194

634.589014

569.8085691

FP16

passage

512

64

1

205.2655254

204.74978

208.1572084

210.8121596

311.4421668

FP16

passage

512

64

3

410.9122943

411.082123

412.4502242

412.745361

466.4179481

FP16

passage

512

64

5

681.5252158

681.9785185

683.5981631

684.081592

468.6722081

FP16

query

20

1

1

5.349868331

5.286015

5.6207312

5.717331

186.3291474

FP16

query

20

1

3

7.411916441

7.4081655

7.5194672

7.6627225

403.8435598

FP16

query

20

1

5

11.90756041

12.35885

12.5979345

12.761145

419.1684076

FP16

query

20

1

7

16.58563318

17.30968

17.5332659

17.618394

421.4802771

FP16

query

20

1

9

21.2083815

22.1409605

22.4645555

22.57402875

423.9030531

FP16

query

20

1

11

26.13345736

27.168601

27.7088854

27.8043151

420.5148337

FP16

query

20

1

13

30.75539894

31.155936

32.590465

32.686162

422.3358847

FP16

query

20

1

15

36.41082533

37.3229475

37.8944988

37.9994324

411.6554532

NVIDIA A100 PCIe 80GB#

The following table contains the performance data for NVIDIA QA E5 Embedding on NVIDIA A100 PCIe 80GB.

Precision

Input Type

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP16

passage

300

64

1

188.3624994

191.678584

197.67535

198.7671405

338.7085017

FP16

passage

300

64

3

371.7164853

372.696551

393.8187446

470.9664648

515.255286

FP16

passage

300

64

5

619.4848889

621.652637

648.4527444

728.8011656

515.0944739

FP16

passage

512

64

1

222.7213055

222.302109

226.0124854

227.4071962

286.5090374

FP16

passage

512

64

3

446.9578206

446.956036

448.6925762

449.3574626

428.4463769

FP16

passage

512

64

5

742.316976

742.838223

744.9883932

745.4736348

430.0964269

FP16

query

20

1

1

6.61186138

6.582944

6.977546

7.1138479

149.3209374

FP16

query

20

1

3

7.36042852

7.345678

7.5866995

7.65647025

404.7664237

FP16

query

20

1

5

11.7962158

12.222473

12.503024

12.59291

421.5411133

FP16

query

20

1

7

16.35367295

17.0629145

17.4296151

17.541723

426.3990987

FP16

query

20

1

9

20.86020568

21.905815

22.2696387

22.37563665

429.9399649

FP16

query

20

1

11

25.69697501

26.830452

27.410482

27.724138

427.1729204

FP16

query

20

1

13

30.3572733

31.467969

32.107687

32.228261

427.3957507

FP16

query

20

1

15

35.59845166

36.40144

37.911859

38.023954

420.5844606

NVIDIA L40S#

The following table contains the performance data for NVIDIA QA E5 Embedding on NVIDIA L40S.

Precision

Input Type

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP16

passage

300

64

1

2554.319287

2563.941207

2678.142464

2698.264006

25.04941137

FP16

passage

300

64

3

7349.208273

7502.055819

7889.257348

7968.054606

25.46501769

FP16

passage

300

64

5

11913.18547

12461.92393

12893.43705

12969.40433

25.62482416

FP16

passage

512

64

1

3701.910032

3701.626575

3703.128269

3703.37542

17.28565985

FP16

passage

512

64

3

10730.17861

10985.16205

10987.04189

11029.24857

17.46973511

FP16

passage

512

64

5

17355.37756

14691.33609

22035.36884

22035.66547

17.42093242

FP16

query

20

1

1

32.39684445

32.390068

32.697962

32.812839

30.71068928

FP16

query

20

1

3

82.5131236

85.641417

85.943178

86.0222786

36.28989135

FP16

query

20

1

5

135.5158144

142.938266

143.2624068

143.3466052

36.83096757

FP16

query

20

1

7

191.652608

200.174195

200.5275478

200.61249

36.46886827

FP16

query

20

1

9

246.8746939

257.437655

257.8359189

257.9366438

36.40752495

FP16

query

20

1

11

301.7433573

314.645817

315.1281346

315.2351826

36.38786195

FP16

query

20

1

13

356.6215151

371.558847

372.237291

372.367342

36.38148395

FP16

query

20

1

15

409.5395827

401.376161

429.7671046

429.915561

36.54702515

NVIDIA L4#

The following table contains the performance data for NVIDIA QA E5 Embedding on NVIDIA L4.

Precision

Input Type

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP16

passage

300

64

1

483.6478251

483.56992

505.150513

509.4594582

132.2515378

FP16

passage

300

64

3

1327.952502

1334.361511

1367.915559

1379.568671

143.9367794

FP16

passage

300

64

5

2181.789118

2203.719238

2241.932492

2250.362263

145.4517364

FP16

passage

512

64

1

633.7960827

633.8213355

638.6268792

639.4460989

100.9307126

FP16

passage

512

64

3

1744.892764

1755.272823

1761.547025

1763.103427

109.3979277

FP16

passage

512

64

5

2892.159813

2923.949474

2934.835412

2936.818463

109.4116347

FP16

query

20

1

1

8.00527288

7.97713

8.255327

8.3389483

124.1064504

FP16

query

20

1

3

11.2179374

12.2274185

12.643952

12.7737285

266.1168591

FP16

query

20

1

5

19.89864006

20.623372

21.068909

21.218374

250.3263155

FP16

query

20

1

7

27.61174325

28.90997

29.4092535

29.5545445

252.95772

FP16

query

20

1

9

35.05091948

36.7263

37.341008

37.513897

256.1157064

FP16

query

20

1

11

42.73745639

44.63778

45.480588

45.658214

256.9000856

FP16

query

20

1

13

50.74195156

50.275669

54.0248513

54.245848

255.897636

FP16

query

20

1

15

57.40762279

57.9185695

62.191041

62.46531895

260.999059

NVIDIA A10G#

The following table contains the performance data for NVIDIA QA E5 Embedding on NVIDIA A10G.

Precision

Input Type

Input Tokens

Batch Size

Concurrency

Avg Latency

P50 Latency

P90 Latency

P95 Latency

Throughput (inputs/s)

FP16

passage

300

64

1

377.9075444

376.386471

396.5482632

404.4754582

169.1417219

FP16

passage

300

64

3

929.9058353

932.259007

972.005973

979.1739983

205.8495868

FP16

passage

300

64

5

1545.528521

1555.163799

1597.331709

1610.341433

205.8607711

FP16

passage

512

64

1

469.4904227

468.610089

473.875774

475.110078

136.232079

FP16

passage

512

64

3

1178.924205

1182.383929

1183.031068

1183.212201

162.2102538

FP16

passage

512

64

5

1958.134001

1970.934197

1971.589458

1971.823967

162.2009027

FP16

query

20

1

1

11.05668374

11.059908

11.4824455

11.6029265

89.77224972

FP16

query

20

1

3

19.26746096

20.2599865

20.804533

20.961988

154.9430338

FP16

query

20

1

5

32.06569956

33.962209

34.599702

34.76241

155.5098235

FP16

query

20

1

7

44.76415285

47.4440885

48.0757631

48.2137225

156.015683

FP16

query

20

1

9

57.69580393

60.895778

61.8394424

62.03314885

155.7518518

FP16

query

20

1

11

70.60605841

74.0469635

75.482615

75.69873895

155.546507

FP16

query

20

1

13

83.83823323

82.7783115

89.2197166

89.615594

154.8850883

FP16

query

20

1

15

97.47471991

96.5648555

103.1160317

103.4140632

153.7165527