Manga OCR Speed Tests Across GPUs
My Manga Translator team and I developed MangaTranslate Manga OCR in-house. It is not the similarly named open-source OCR project. I process a large number of OCR requests every day, so I focus on images processed per second when comparing GPUs.
On my current workload, I measured 3 images per second on a single A10M. In my earlier GPU comparison, the A10M ran at roughly half the speed of an RTX 4080 SUPER, which puts the 4080 SUPER at approximately 6 images per second. On the same workload, our Manga OCR delivered about twice the throughput of PaddleOCR.
GPU Speed Comparison
I have identified how each figure was obtained. The A10M result is directly measured, and the 4080 SUPER figure is calculated from the observed speed ratio. The remaining figures are projections based on theoretical FP16 compute performance. They are not Manga OCR benchmark results.
| GPU | Manga OCR Speed (images/second) | Basis |
|---|---|---|
| RTX 4080 SUPER | Approx. 6.0 | Calculated from the observed A10M-to-4080 SUPER speed ratio |
| RTX 4070 Ti SUPER | Approx. 5.1 | Theoretical compute projection; benchmark pending |
| RTX 5070 Ti | Approx. 5.0 | Theoretical compute projection; benchmark pending |
| RTX 4070 Ti | Approx. 4.6 | Theoretical compute projection; benchmark pending |
| RTX 3090 Ti, standard model | Approx. 4.6 | Theoretical compute projection; benchmark pending |
| RTX 3090 | Approx. 4.1 | Theoretical compute projection; benchmark pending |
| A10G | Approx. 3.6 | Theoretical compute projection; benchmark pending |
| RTX 3080 20GB | Approx. 3.4 | Projected using standard RTX 3080 compute specifications; exact card configuration needs verification |
| RTX 2080 Ti | Approx. 3.3 | Theoretical compute projection; benchmark pending |
| V100 | Approx. 3.2–3.6 | Specifications vary between PCIe and SXM models; benchmark pending |
| A10M | 3.0 | Measured result |
| RTX 4060 Ti | Approx. 2.5 | Theoretical compute projection; benchmark pending |
| RTX 4060 | Approx. 1.7 | Theoretical compute projection; benchmark pending |
| RTX 3090 Ti advertised as 32GB; RTX 3090 advertised as 20GB | Not listed | Non-standard configurations; exact hardware must be verified first |
This table covers the GPUs I am evaluating, but only figures explicitly marked as measured should be treated as Manga OCR test results. For the projections, I used approximately 6 images per second on the 4080 SUPER as a reference and scaled it by each GPU's theoretical FP16 compute performance. This calculation does not account for image preprocessing, the precision used by my model, memory access, or request scheduling. Hardware specifications are available from NVIDIA's GeForce comparison, A10 specifications, and V100 specifications.
What the Results Mean for My Deployment
The A10M's 3 images per second is my clearest single-GPU measurement so far. With many independent requests, I can distribute work across two A10M GPUs. Based on the single-GPU result, their combined throughput could approach 6 images per second. I still need to measure that dual-GPU setup directly, and adding a second GPU will not automatically make an individual request finish twice as fast.
I use throughput on my current workload when comparing Manga OCR with PaddleOCR. The roughly twofold difference is a result from that workload; model versions, image content, and inference settings can change it. As I complete comparable tests on the other GPUs, I will replace the projections in the table with measured images-per-second results.