I have experimented extensively with Amarra, and wouldn't be without it.
The best solution I found was to use a powerful Mac (more power is a good thing) with 8GB of RAM (more RAM is a good thing), Amarra/iTunes, and outputting through an Empirical Offramp 3 and an excellent USB cable (I like the Offramp as it allows me a world of DAC choices).
I also found that with my 24/96 DAC, I was best to convert all my files to 24/96 and store them at that resolution, uncompressed. Its a bit like the idea of feeding your LCD with a file that matches its natural resolution. This takes load off both the Mac (no unpacking of ALAC files on the fly), and off the DAC (no real-time upsampling to do). Both can then work more accurately. And you really should dedicate a Mac to the task, not use it for anything else.
Peoples' preferences about upsampling versus 'native' blends/confuses two different issues. Upsampling, if done well can be very beneficial, up to a point (which depends on your DAC chip), but doing it in real time taxes your DAC's performance. I use AIFF as it can handle tags etc, unlike WAV, and use SampleManager software for the conversion to 24/96.
Remember that the digital to analog conversion essentially upsamples in the way that it joins the dots in the analog domain between the discrete samples. Different algorithms for joining the dots sound different, even though the imaginary data is supposedly beyond our hearing range. There is a trade-off between how much of the imaginary data is done in the digital domain (upsampling) and how much is done by the DAC process. Doing it offline is a good thing as I have said before, but there is a limit to how high you should upsample - mainly based on whether your DAC works well (or at all) at the higher bit rate.