From their description it is clear their denoising algorithm refers to kind of non local means denoising advancements. These methods works best for moderate noise then their efficiency drops very quickly when standard deviation of noise is increasing.
But for me it works reasonably well up to ISO12800 on my NEX-5n (in fact they are better than ISO6400 pics processed with Topaz Denoise which was my choice before).
Example (ISO3200 shot pulled up 2 stops).
Since I'm going to buy Sony A7r in the very near future, I was interested what will I get with its sensor which is quite close to D800E's one. IMO results are nothing short but spectacular:
D800E ISO3200 full size D800E ISO6400 shot downsampled to 16Mp D800E ISO12800 shot downsampled to 16Mp
ISO25600 was not good though, but looking at DxOmark charts I see noticeably higher SNR at ISO25600 for A7r compared to D800E, so there's a good chance ISO25600 will be usable with that Sony.
Some remarks about features of NR algorithms of its kind:
1) They love higher resolution to some extent (see #3), because for every certain part of an image they uses other parts of it to reconstruct taken one. Higher resolution sensor provides more information, so the reconstruction will likely be more «successful»
2) They work better at sharp details. You see it at my ISO3200*2² shot, where fine details are left, but bokeh is getting dirty.
3) So, they use patches in a neighborhood to reconstruct a certain pixel. However, those patches are not clean themselves and when noise getting more and more intense those patches are getting noisier too. And amount of these unacceptably noisy patches leads to errors in reconstruction.