I listened closely to the mp3 A-B comparison of a C# scale passage suggested by Philippe, and, although it was not clearly specified, by listening again to the Roland video (about halfway at the Vintage Piano title) I found that probably A was the Roland instrument and B pianoteq with the Erard preset.
I was not sure because A may sound a bit like one of the earlier C1 or C2 preset which I find have some less realistic sections. I think Philippe might have chosen this passage because it may show the progress in the naturalness of the presets over the years. The Erard is much more natural-sounding and I don't know if this is due to better analysis techniques of the real instrument or rendering possibilities added with each versions.
Also, A suffers from a very thin sounding upper register, reminding me of the earlier looping sample pianos when memory was limiting the length of high frequency loops. If you look closely at the Roland video, you will find that most of the playing is in the low to mid registers with arpeggios to the high notes with pedal on.
I play on a 10 year old Yamaha P80, which was probably one of the most natural sounding at the time, and I was very surprised when I stumbled upon this forum entry which compared the P80 itself and pianoteq on smal Chopin excerpts:
http://www.pianoworld.com/ubb/ubb/ultim...tml#000003
Now these examples focus on the mid and lower part of the register which explains why both sound very similar to my ear. Maybe Roland has some work to do still...
On the cpu usage, I would like to say that on my Windows XP machine (AMD Athlon 64 X2 4200+ 2.2 GHz dual core with 2 GB memory) there is absolutely no problem, pianoteq using only 10-15% of cpu at idle and 25-30% while playing the included Chopin waltz at 48kHz sampling and with unlimited polyphony.
I can't see why Roland needs 4 cpus (if they are all used for computing the model and not simply controlling the keyboard, pedals etc.) unless their model lacks the algorithmic optimisations that surely went into pianoteq and are unique to it. It is one thing to program equations in a straightforward fashion and another to devise new ways of calculating a mathematical model. I suspect a lot of very clever programming went in the design of pianoteq to obtain such a good product and not so much in a simplification of the model itself. (I may be wrong of course...)
I say this because I did quite a lot of scientific programming myself and I remember accelerating by a factor of 100 a badly written program with simple factoring techniques. As a better example, consider the FFT (Fast Fourier Transform) algorithm which is about 100 times faster than the original DFT (Discrete Fourier Transform). Efficient programming is important and cannot be entirely compensated by faster cpus.
Sorry for the long post! Sometimes there is a lot to say...