Benchmarks

The following section provides some performance figures for CleverDocs’s CPU and GPU pipelines/containers on various AWS instance types

PDF de-identification

For evaluation, we are used 3 PDF documents with total number of pages is 410 which contain scanned documents.

Please look more detail about performance of CleverDoc in related blog post: Automated PDF de-identification: Performance Analysis

CPU

Instance Type	vCPUs	Memory (GiB)	Price ($/hour)	Throughput (pages/sec)	Latency for 410 pages request (secs)
m6g.8xlarge	32	128	$1.232	4.4	98

GPU

Instance Type	vCPUs	GPUs	Memory (GiB)	GPU Memory (GiB)	Price ($/hour)	Throughput (pages/sec)	Latency for 410 pages request (secs)
g4dn.8xlarge	32	1	128	16	$2.176	7.4	55

Image de-identification

Dataset:

• 100 PNG images of scanned documents with text

• Resolution: 300 dpi

• Width: 2469 px

• Height: 3500 px

CPU

Instance Type	vCPUs	Memory (GiB)	Price ($/hour)	Throughput (images/sec)	Latency for 100 images request (secs)
m6g.8xlarge	32	128	$1.232	6.8	14.5

GPU

Instance Type	vCPUs	GPUs	Memory (GiB)	GPU Memory (GIB)	Price ($/hour)	Throughput (images/sec)	Latency for 100 images request (secs)
g4dn.8xlarge	32	1	128	16	$2.176	9	11