To clarify my last post, I suggest we don't use Atmel's CAN drivers and instead work them from scratch. But if they are really good, then no need to mess with them. In reality, it will probably be a task of looking what they did and seeing if we can improve upon their driver implementation. If it's anything like the CAN processing on the PIC24/PIC32, you can probably offload much of the CAN receiving messages to a DMA channel, which will be key for efficiency. This could be done with TX too, but may not benefit too many applications. I should probably start reading the Atmel uC manual to see how they implement this stuff.