I have an EVM for the battery fuel gauge so I can connect that to my laptop and log the battery draining from start to finish through the USB port so it doesn't matter if the Pandora board shuts down on its own when the battery is empty. I still haven't done a full charge to empty cycle all at once but the fuel gauge learns the battery profile and was predicting the usual 8.5-9 hours. Since I have run the battery down over a couple of days with my testing, the prediction is accurate as the battery capacity is at or above the rating given by the manufacturer.
The load doesn't change much whether you idle at the desktop or run an emulator or app. I think the change is around a 50mA range based on processing and how quickly the pixels are changing so doing pretty much anything with just the CPU core is going to perform about the same. Average power with the LCD at full brightness is around 1.6W. Adding the SGX core into the mix (like with Quake) added another 30-60mA of drain I think. It fluctuates with what's going on. It's been a couple of months since I have run these tests but the hardware hasn't changed to affect new results with production boards so this is all still valid.
What really changes the power usage is when the LCD goes off, cranking up the internal speaker volume or plugging something into the USB port (a simple mouse or keyboard will decrease battery life by 22-25%. That's how little the whole system uses!). If you slow the clock for things that don't need it (I run Picodrive at 200MHz in Angstrom for example), you get the 10 hour life and if I turned down the LCD brightness it would be a bit more.
When the LCD goes off and I'm playing music, the battery life was in the 12 to 15 (if I remember to slow down clock) hour range. We still don't have power management in the code so everything is lit up, the core is running at the voltage for 500MHz and the clock was usually at 500MHz so there is all kinds of room for improvement. If an N900 or Palm Pre can run all day on an 1100-1500mAh battery (maybe?), we should be able to do way better with 4000mAh. The hardware has the capability to shut down anything not in use so it's purely a software effort needed. We only have one person volunteering to work on the power management code right now so software seems to develop at a slow serial rate with the extra stuff like this.
EDIT: I also don't recommend going around boasting a 100 hour battery life playing music. As some people have pointed out, getting anywhere near that will require some serious talent to code. That might be a community project some day but no one should count on that as a launch feature to shame your iPod buddies.
