← sources

We almost bought a second GPU for an OCR pipeline. Optimizing the existing one gave 15x instead. How often is “buy more hardware” the wrong fix?

source: reddit (reddit:computervision)published 2026-10-25 14:03ingested 2026-10-26 02:36

retained text

a custom multi-language OCR pipeline on a single GPU was choking as load grew (\~4 detections/sec, 3s+ tail latency). The plan was a second GPU. Instead we optimized serving on the same card and got 55–125/s with \~100ms tail latency, and memory use actually dropped.

I can’t share the specifics since it’s internal work, but I’m curious how common this is. Have you seen teams scale hardware when the real issue was utilization? What was the bottleneck in your case?

document fetched 2026-10-26 00:23 · canonical https://reddit.com/r/computervision/comments/1xf36zh/we_almost_bought_a_second_gpu_for_an_ocr_pipeline