Stagnant, the cropped images, as presented here, are 800px high. You want to present them so we can easily compare them by scrolling up and down. Clicking through to Flicker and selecting the original size is not convenient for this purpose.
The linear dimension compression is 800/1294 = 61.8%. The area compression is (800/1294)^2 = 37.5% of the original area, or 290,000px (370x800) / 773,812px (598x1294) = 37.5%
The improvement in SNR in dB with the area reduction is 20log(sq-root(area1/area2)). You can also substitute the ratio of MPs for the ratio of areas. (See:
Detailed computation of DxOMark Sensor normalization - DxOMark)
It's easier to work just with the linear dimension reduction: 20log(height1/height2)
= 20log(1294/800)
= 4.2dB higher SNR .
The full-scene "thumbnails" you presented are 333px high. Since the K-3 frame height is 4,000px, the improvement in SNR is:
20log(4000/333)
= 21.6dB higher SNR.
This improvement is equivalent to shooting at ISO1055 instead of at ISO128000. That's why the small full-scene versions hardly show any noise compared to the 100% crops, and are suitable for web presentation.
So knowing both the original size & the presentation size are critical when assessing noise performance. Using higher ISO means that an image is suitable for a smaller viewing size, or a larger image size, but viewed from a greater distance.
Dan.