Like many questions about what setting works best: It depends.
The native ISO for almost all Canon DLSRs over the last few years has been ISO 100. 'Full stop' intervals, such as ISO 200, ISO 400, ISO 800, etc. increase the analog amplification of the signal readout of the sensor. The 1/3 stops in between those full stops use software adjustments during in-camera processing of the data coming off the sensor. Here's what happens if you select, for example, ISO 160 when you take a shot. The sensor is set to ISO 200. The camera overexposes the shot by 1/3 stop by increasing Exposure Compensation (E.C.) 1/3 more stops than the user selected value. When the data from the sensor is read into the processor, 1/3 stop of pull is applied to the data. The effect this has is that a shot taken at ISO 160 has slightly less noise in the shadows at the expense of slightly less headroom in the highlights for an overall slight reduction in dynamic range. Settings 1/3 stop above the 'full stop' ISO settings work in the reverse: the camera exposes by -1/3 stop and then pushes exposure by 1/3 stop when the sensor readout data is processed.
So what does this mean when selecting what ISO to use for a specific shot?
If you are shooting video or allowing the in-camera settings to be applied to the RAW data and then saving the files in the JPEG format:
- If you are in dim light where shadow noise is the biggest concern, select the -1/3 stop ISO setting such as 160, 320, 640, 1250, etc. that is nearest to the aperture and shutter speed settings you desire. Effectively you are telling the camera to automatically expose to the right by 1/3 stop and then apply -1/3 stop when converting the analog information from the sensor to digital.
- If you are in a setting where there aren't many shadows and not blowing out the highlights is the most significant concern, select the full stop ISO settings such as 100, 200, 400, 800, 1600, etc.
- You should probably avoid the +1/3 stop ISO settings (ISO 125, 250, 500, 1000, etc.) altogether. With the +1/3 stop setting you give up the dynamic range of the 'full stop' settings. But since the signal off the sensor is increased by 1/3 stop via software, the noise in the image is also increased by 1/3 stop.
If you are saving the files as RAW data it becomes a little murkier. You should be able to get just as good results in terms of shadow noise by using +1/3 stop more E.C. to increase the Tv/Av combination and selecting 'full stop' ISO values as you would by reducing the ISO setting -1/3 stop and leaving the E.C. setting alone. But if that pushes some of the highlights over the edge into full saturation on any of the three color channels, then you effectively give up the same dynamic range that using a -1/3 stop ISO value would have given up.
In the case of RAW files the Signal to Noise Ratio (SNR) is largely determined by the amount of light allowed to enter the camera by the Av/Tv combination selected compared to the sensor's fairly constant read noise. When shooting in automatic exposure modes (P, Tv, Av), by telling the camera's metering system you are shooting at ISO 320, it will select an Av/Tv value that allows 1/3 stop more light into the camera than if you tell it you are shooting at ISO 400.
Here is a link to test shots ordered from the lowest to highest amount of shadow noise from a Canon 60D. In order of lowest to highest measured noise at each ISO setting the sequence is 160, 320, 640, 100, 200, 400, 800, 1250, 125, 250, 500, 1000, 1600, 2500, 2000, 3200, 4000, 5000, 6400. ISO 1250 has roughly the same amount of noise as ISO 125! Here's a test with similar results using the canon 5D Mark II, and video shot with a 7D. The graph included in this one is fairly precise and shows the expected performance of the Canon 5DII. My own personal experience with the Canon 5DII is that there is little performance difference up to and including ISO 1250. ISO 2000 is marginally noisier than ISO 2500 and ISO 1600. ISO 5000 is the last setting I can use before the noise performance falls off the cliff.
Based on this study, Canon started adopting this method sometime between the 1D Mark IIN and the 1D Mark III and original 5D.
The superior high ISO/low noise performance of a Full Frame sensor compared to an APS-C sensor (of the same generation of technology) is due to the physical size of the sensor and thus the total amount of light falling on the sensor. In the case of Canon cameras, the current APS-C sensors all have a pixel pitch of just over 4µm. The pixel pitches of current Canon FF sensors range from 6.25-6.9µm. When the linear width is converted to surface area, the FF sensors have pixels that cover over twice the area of their APS-C counterparts and thus collect twice as much light per pixel under the same lighting conditions and Tv/Av settings.