Performance

Training Accuracy Results

Training Accuracy: NVIDIA DGX SuperPOD (8 x 8 x A100 80GB for CLIP B/32 Model)

We followed the training recipe from Open CLIP blog to verify our training pipeline. Our results are displayed in the table below:

Framework	Dataset	Model Name	Batch Size	Samples Seen	ImageNet Top-1
OpenCLIP	LAION 400M	B/32	32k	12B	62.90%
NeMo	Our Multimodal Blend*	B/32	32k	12B	60.13%

Note

Our multimodal dataset is originated from Common Crawl with custom filtering and contains 670M image-caption pairs.

We believe the final accuracy difference is due to the dataset, as LAION 400M is filtered with CLIP scores. To ensure our implementation is consistent with OpenCLIP, we trained OpenCLIP with our dataset and found out that the loss curve and validation accuracy were nearly identical to NeMo’s CLIP.

Training Performance Results

We measured the throughput of training CLIP models on different numbers of DGX A100 nodes and DGX H100 nodes, and we achieved near-linear scaling on both platforms.

We are comparing the out-of-box performance on DGX H100 machines with the same configuration from DGX A100 machines. This comparison is an apple-to-apple assessment, ensuring that we evaluate the relative performance of the two machine types under equivalent conditions and configurations.

The tables and charts below show the performance results.

NVIDIA DGX SuperPODs (16 x 8 x A100 80GB for CLIP g/14 model)

		1	2	4	8	16
	Samples per Second	559	1115	2190	4407	8633
CLIP g/14	Perfect Linear Scaling (Samples)	559	1119	2237	4475	8950
	Speedup	1x	1.99x	3.92x	7.88x	15.43x

NVIDIA DGX SuperPODs (16 x 8 x H100 80GB for CLIP g/14 model)

		1	2	4	8	16
	Samples per Second	935	1795	3502	6771	13829
CLIP g/14	Perfect Linear Scaling (Samples)	935	1869	3739	7478	14955
	Speedup	1x	1.92x	3.75x	7.24x	14.8x

DGX A100 vs. DGX H100: A Comparative Analysis of CLIP Training

Model	Nodes	Global Batch Size	Micro Batch Size	Precision	Global Batch / Sec (A100)	Global Batch / Sec (H100)	Speedup (x)
CLIP B/32	4	16000	500	bf16 (O2)	2.12	5.26	2.5
CLIP H/14	4	3584	112	bf16 (O2)	0.88	1.92	2.2
CLIP g/14	4	2560	80	bf16 (O2)	0.86	2.25	2.6

Inference Performance Results

Latency times are taken as starting with an image on CPU and text input (of length 64) and stopped on output. For framework we use the Torch Automated Mixed Precision (AMP) for FP16 computation. For TRT, we export the various models with the FP16 acceleration. We use the optimized TRT engine setup present in the deployment directory to get the numbers in the same environment as the framework.

GPU: NVIDIA DGX A100 (1x A100 80 GB) Batch Size: Number of Images in a Batch

Model	Batch Size	TRT FP16 Latency (s)	FW FP 16 (AMP) Latency (s)	TRT vs FW Speedup (x)
	1	0.014	0.032	2.3
	2	0.014	0.033	2.4
CLIP B/32	4	0.014	0.028	2.0
	8	0.015	0.028	1.9