Throughput benchmarks for all OpenCLIP models

OpenCLIP Throughput Benchmark (image embeddings per second)

This should give an idea of relative throughput of the models. I could not discern what would be fastest from the names alone.

This is just a speed test. Obviously the larger models will perform better on evaluation benchmarks at the tradeoff of speed. Find the models that meet your throughput requirements then benchmark those for performance on the task you are doing.

Tested on an NVIDIA RTX 3090. CPU is an AMD 7950x, though that should not affect the benchmark much.

Code

Images are pre-loaded into memory. The benchmark measures preprocess + encode.

    t0 = time.time()

    preprocd = []
    for image in images:
        preprocd.append(clip_preprocess(image).unsqueeze(0).to(device))
 
    with torch.no_grad():
        embeds = []
        for image in preprocd:
            embeds.append(clip_model.encode_image(image).to("cpu"))

    t1 = time.time()
    print(arch, pretrained, len(images)/(t1-t0))

Average throughput per architecture

Architecture	Images Per Second
ViT-B-32	170.317
nllb-clip-base	166.847
xlm-roberta-base-ViT-B-32	163.626
ViT-B-32-256	161.603
ViT-B-32-quickgelu	161.004
roberta-ViT-B-32	159.02
coca_ViT-B-32	156.964
ViT-B-16-SigLIP	152.068
ViT-B-16	150.31
RN50-quickgelu	147.837
ViT-B-16-quickgelu	145.091
ViT-B-16-SigLIP-256	141.253
ViT-B-16-SigLIP-i18n-256	140.578
RN50	129.033
ViT-B-16-plus-240	126.04
RN50x4	110.681
EVA02-B-16	100.629
RN101-quickgelu	93.8347
convnext_base	93.7734
RN101	92.2537
convnext_base_w	90.138
convnext_base_w_320	89.1247
convnext_large_d	83.5527
ViT-B-16-SigLIP-384	79.4187
RN50x16	65.2571
convnext_large_d_320	60.0695
ViT-L-16-SigLIP-256	59.8122
ViT-L-14-CLIPA	57.4581
ViT-L-14	54.3296
coca_ViT-L-14	53.7369
ViT-L-14-quickgelu	52.6421
ViT-SO400M-14-SigLIP	50.0526
EVA02-L-14	43.3144
ViT-B-16-SigLIP-512	42.4944
ViT-H-14-CLIPA	31.9688
ViT-L-14-CLIPA-336	31.7083
convnext_xxlarge	31.5165
ViT-L-16-SigLIP-384	30.9679
xlm-roberta-large-ViT-H-14	30.9487
nllb-clip-large	30.9176
ViT-H-14	30.8775
ViT-H-14-quickgelu	30.0928
ViT-L-14-336	29.4929
EVA02-L-14-336	26.5657
RN50x64	26.1571
EVA01-g-14	20.4507
EVA01-g-14-plus	20.4077
ViT-g-14	20.0696
ViT-SO400M-14-SigLIP-384	17.6165
ViT-H-14-CLIPA-336	16.354
ViT-bigG-14-CLIPA	12.1527
ViT-bigG-14	12.0347
ViT-bigG-14-CLIPA-336	6.52044
EVA02-E-14-plus	5.52984
EVA02-E-14	5.5032

Raw data for all pretrains

Architecture	Pretrained	Images Per Second
RN50	openai	107.945
RN50	yfcc15m	139.889
RN50	cc12m	139.266
RN50-quickgelu	openai	164.015
RN50-quickgelu	yfcc15m	139.801
RN50-quickgelu	cc12m	139.694
RN101	openai	100.867
RN101	yfcc15m	83.64
RN101-quickgelu	openai	102.801
RN101-quickgelu	yfcc15m	84.868
RN50x4	openai	110.681
RN50x16	openai	65.2571
RN50x64	openai	26.1571
ViT-B-32	openai	163.519
ViT-B-32	laion400m_e31	171.204
ViT-B-32	laion400m_e32	170.896
ViT-B-32	laion2b_e16	169.895
ViT-B-32	laion2b_s34b_b79k	173.671
ViT-B-32	datacomp_xl_s13b_b90k	171.88
ViT-B-32	datacomp_m_s128m_b4k	172.488
ViT-B-32	commonpool_m_clip_s128m_b4k	169.047
ViT-B-32	commonpool_m_laion_s128m_b4k	170.736
ViT-B-32	commonpool_m_image_s128m_b4k	172.236
ViT-B-32	commonpool_m_text_s128m_b4k	173.43
ViT-B-32	commonpool_m_basic_s128m_b4k	168.535
ViT-B-32	commonpool_m_s128m_b4k	170.343
ViT-B-32	datacomp_s_s13m_b4k	170.372
ViT-B-32	commonpool_s_clip_s13m_b4k	170.721
ViT-B-32	commonpool_s_laion_s13m_b4k	170.097
ViT-B-32	commonpool_s_image_s13m_b4k	170.926
ViT-B-32	commonpool_s_text_s13m_b4k	169.276
ViT-B-32	commonpool_s_basic_s13m_b4k	169.176
ViT-B-32	commonpool_s_s13m_b4k	167.898
ViT-B-32-256	datacomp_s34b_b86k	161.603
ViT-B-32-quickgelu	openai	163.239
ViT-B-32-quickgelu	laion400m_e31	159.452
ViT-B-32-quickgelu	laion400m_e32	161.956
ViT-B-32-quickgelu	metaclip_400m	159.77
ViT-B-32-quickgelu	metaclip_fullcc	160.604
ViT-B-16	openai	148.401
ViT-B-16	laion400m_e31	151.745
ViT-B-16	laion400m_e32	151.355
ViT-B-16	laion2b_s34b_b88k	147.765
ViT-B-16	datacomp_xl_s13b_b90k	148.236
ViT-B-16	datacomp_l_s1b_b8k	152.87
ViT-B-16	commonpool_l_clip_s1b_b8k	150.399
ViT-B-16	commonpool_l_laion_s1b_b8k	149.891
ViT-B-16	commonpool_l_image_s1b_b8k	150.263
ViT-B-16	commonpool_l_text_s1b_b8k	151.263
ViT-B-16	commonpool_l_basic_s1b_b8k	151.263
ViT-B-16	commonpool_l_s1b_b8k	150.274
ViT-B-16-quickgelu	metaclip_400m	145.243
ViT-B-16-quickgelu	metaclip_fullcc	144.938
ViT-B-16-plus-240	laion400m_e31	125.802
ViT-B-16-plus-240	laion400m_e32	126.279
ViT-L-14	openai	53.6898
ViT-L-14	laion400m_e31	54.5316
ViT-L-14	laion400m_e32	54.1111
ViT-L-14	laion2b_s32b_b82k	54.339
ViT-L-14	datacomp_xl_s13b_b90k	54.6971
ViT-L-14	commonpool_xl_clip_s13b_b90k	54.8216
ViT-L-14	commonpool_xl_laion_s13b_b90k	54.2918
ViT-L-14	commonpool_xl_s13b_b90k	54.155
ViT-L-14-quickgelu	metaclip_400m	52.2084
ViT-L-14-quickgelu	metaclip_fullcc	53.0757
ViT-L-14-336	openai	29.4929
ViT-H-14	laion2b_s32b_b79k	30.8775
ViT-H-14-quickgelu	metaclip_fullcc	30.0928
ViT-g-14	laion2b_s12b_b42k	20.0461
ViT-g-14	laion2b_s34b_b88k	20.093
ViT-bigG-14	laion2b_s39b_b160k	12.0347
roberta-ViT-B-32	laion2b_s12b_b32k	159.02
xlm-roberta-base-ViT-B-32	laion5b_s13b_b90k	163.626
xlm-roberta-large-ViT-H-14	frozen_laion5b_s13b_b90k	30.9487
convnext_base	laion400m_s13b_b51k	93.7734
convnext_base_w	laion2b_s13b_b82k	88.6564
convnext_base_w	laion2b_s13b_b82k_augreg	90.0901
convnext_base_w	laion_aesthetic_s13b_b82k	91.6674
convnext_base_w_320	laion_aesthetic_s13b_b82k	89.222
convnext_base_w_320	laion_aesthetic_s13b_b82k_augreg	89.0274
convnext_large_d	laion2b_s26b_b102k_augreg	83.5527
convnext_large_d_320	laion2b_s29b_b131k_ft	60.3519
convnext_large_d_320	laion2b_s29b_b131k_ft_soup	59.7872
convnext_xxlarge	laion2b_s34b_b82k_augreg	31.4495
convnext_xxlarge	laion2b_s34b_b82k_augreg_rewind	31.504
convnext_xxlarge	laion2b_s34b_b82k_augreg_soup	31.5961
coca_ViT-B-32	laion2b_s13b_b90k	156.336
coca_ViT-B-32	mscoco_finetuned_laion2b_s13b_b90k	157.592
coca_ViT-L-14	laion2b_s13b_b90k	53.6466
coca_ViT-L-14	mscoco_finetuned_laion2b_s13b_b90k	53.8271
EVA01-g-14	laion400m_s11b_b41k	20.4507
EVA01-g-14-plus	merged2b_s11b_b114k	20.4077
EVA02-B-16	merged2b_s8b_b131k	100.629
EVA02-L-14	merged2b_s4b_b131k	43.3144
EVA02-L-14-336	merged2b_s6b_b61k	26.5657
EVA02-E-14	laion2b_s4b_b115k	5.5032
EVA02-E-14-plus	laion2b_s9b_b144k	5.52984
ViT-B-16-SigLIP	webli	152.068
ViT-B-16-SigLIP-256	webli	141.253
ViT-B-16-SigLIP-i18n-256	webli	140.578
ViT-B-16-SigLIP-384	webli	79.4187
ViT-B-16-SigLIP-512	webli	42.4944
ViT-L-16-SigLIP-256	webli	59.8122
ViT-L-16-SigLIP-384	webli	30.9679
ViT-SO400M-14-SigLIP	webli	50.0526
ViT-SO400M-14-SigLIP-384	webli	17.6165
ViT-L-14-CLIPA	datacomp1b	57.4581
ViT-L-14-CLIPA-336	datacomp1b	31.7083
ViT-H-14-CLIPA	datacomp1b	31.9688
ViT-H-14-CLIPA-336	laion2b	16.3764
ViT-H-14-CLIPA-336	datacomp1b	16.3315
ViT-bigG-14-CLIPA	datacomp1b	12.1527
ViT-bigG-14-CLIPA-336	datacomp1b	6.52044
nllb-clip-base	v1	166.847
nllb-clip-large	v1	30.9176

TACIXAT/open_clip_benchmarks.md

OpenCLIP Throughput Benchmark (image embeddings per second)

Code

Average throughput per architecture

Raw data for all pretrains