Publish CPU, Iris Xe and RTX 5050 OCR benchmarks

kleb commented

2026-11-03 22:48:53 +01:00

Owner

The OCR page had no measured results. Publish all 10 supported local configurations: five CPU models, one Iris Xe model and four RTX 5050 models. No hosted APIs were used; the five unsupported model/device combinations remain explicitly skipped. These are six-page synthetic smoke measurements, not a broad document-quality evaluation.

Paddle CPU now enables the standard KV cache in a distinct cohort after its pinned generation config caused a warmup timeout. CUDA behavior is unchanged. Plans default OCR GPU work to the RTX 5050 workstation, and the authorized 24 GiB host-commit retry is documented. Existing results and failed-run logs are preserved. The page discloses from the evidence that all six Iris Xe pages hit the token cap, and slow GPU throughput retains three decimals.

Proven by 439 Python tests, 60 Bun tests, typecheck, schema/privacy and source-receipt validation, the completed hardware runs, and desktop/mobile OCR filtering, exports and evidence downloads.

The OCR page had no measured results. Publish all 10 supported local configurations: five CPU models, one Iris Xe model and four RTX 5050 models. No hosted APIs were used; the five unsupported model/device combinations remain explicitly skipped. These are six-page synthetic smoke measurements, not a broad document-quality evaluation. Paddle CPU now enables the standard KV cache in a distinct cohort after its pinned generation config caused a warmup timeout. CUDA behavior is unchanged. Plans default OCR GPU work to the RTX 5050 workstation, and the authorized 24 GiB host-commit retry is documented. Existing results and failed-run logs are preserved. The page discloses from the evidence that all six Iris Xe pages hit the token cap, and slow GPU throughput retains three decimals. ![ocr-cpu-reviewed.png](https://git.kleb.sh/attachments/2fc52930-0340-4d1a-8134-eb20f2b564b8) ![ocr-gpu-reviewed.png](https://git.kleb.sh/attachments/1b78a600-a3ef-46df-80c3-89e52416383b) Proven by 439 Python tests, 60 Bun tests, typecheck, schema/privacy and source-receipt validation, the completed hardware runs, and desktop/mobile OCR filtering, exports and evidence downloads.

ocr-cpu-desktop.png

65 KiB

ocr-gpu-mobile.png

36 KiB

ocr-cpu-reviewed.png

76 KiB

ocr-gpu-reviewed.png

46 KiB

kleb added 1 commit

2026-11-03 22:48:54 +01:00

Publish CPU, Iris Xe and RTX 5050 OCR benchmarks 9f85a65a27

kAgent commented

2026-11-03 22:48:54 +01:00

First-time contributor

kReview review

Verdict: no findings

No supported defects were found in the reviewed changes.

kReview checked only the PR diff, metadata, and supplied repository context at 0b7bc63881. It reports only issues supported by this material.

Estimated cost: ~$0.72 total (1.0M in / 2.1k out); this run ~$0.10 (201.8k in / 236 out) / gpt-6.1-terra.

Run time: 1:29 total; this run 0:13.

## kReview review **Verdict:** no findings No supported defects were found in the reviewed changes. _kReview checked only the PR diff, metadata, and supplied repository context at `0b7bc63881`. It reports only issues supported by this material._ _Estimated cost: ~$0.72 total (1.0M in / 2.1k out); this run ~$0.10 (201.8k in / 236 out) / gpt-6.1-terra._ _Run time: 1:29 total; this run 0:13._

kleb added 1 commit

2026-11-03 22:49:49 +01:00

Keep slow OCR throughput visible 154c01f651

kleb added 1 commit

2026-11-03 22:52:04 +01:00

Disclose truncated OCR output without changing measurements 02bc8a3101

kleb added 1 commit

2026-11-03 22:52:53 +01:00

Keep the OCR limit count together 0b7bc63881