With offline processing writing to a (big) file any modern system would be powerful enough. If you want to make 1 change a day and leave it to re-calc over night.

But for twiddling (even controlled parameter twiddling with a method-of-experiments plan) something a lot faster is needed.

So where those demo's being done on the fly?

Otherwise just what the Interception Modernisation Programme needs. Now all they need is some of that magic find-all-suspects-with-a-matching-comms-pattern software GCHQ seem so keen on.

How hard could that be?

Isn't it funny when the first comment is done by someone clearly not knowing what they're talking about? Makes the whole article seem cleverer than it is!

<looks for the "Obligatory will-it-run-Crysis" icon.....>

I like the PCI-e way of connecting these GPU-CPU combos as it's a common standard and so will be easy to speed adoption across vendors . That means we might even be able to mix-and-match bits from different vendors to get the most suitable bundle for a given requirement. Whilst I'm sure there are more "elegant" solutions coming, they'll probably be proprietary and mean vendor lock-in.

"Yeah, hi, can I get a quattro formaggio, a SuperServer 6016GT-TF-TM2 and a side order of garlic bread, thanks..."

I also like the idea of using standard PCI-E interconnects for adding on GPU daughterboards. Odd though, I used to think of them as "graphics cards"... Hmmm...

Of course, having anything that would conform allows for upgrading along the curve, as well as substituting "lesser" boards for on-the-cheap(?!) release versions for the masses (hopefully?).