I am a coauthor of the paper.
Indeed, there are many rough estimates in the paper: the theoretical analysis (and indeed, the NFS method itself) relies on some unproven conjectures and asymptotical analysis. Also, the practical cost estimates rely on technological and ecnomical considerations which are rapidly changing. None the less, we believe the assumptions and their conclusions are quite sound when used as order-of-magnitude estimates.
As for the two technical points you mentioned:
> > The bandwidth of the fastest PC memory is 3.2GB/sec
> The GX specs specifically state that they support 4.2 GB per second.
Indeed, but both PC3200 (DDR400) and dual PC800 (RDRAM) have a bandwidth of 3.2GB/sec.
> I checked pricewatch and found at least 6ns for pretty cheap
These "6ns" parts do not have a 6ns random-access latency.
For instance, check these figures.
Slashdot Mirror
> but does it look pretty?
I'd say it does:
a color-coded simulation of the mesh routing algorithm (16MB)
As for the two technical points you mentioned:
> > The bandwidth of the fastest PC memory is 3.2GB/sec
> The GX specs specifically state that they support 4.2 GB per second.
Indeed, but both PC3200 (DDR400) and dual PC800 (RDRAM) have a bandwidth of 3.2GB/sec.
> I checked pricewatch and found at least 6ns for pretty cheap
These "6ns" parts do not have a 6ns random-access latency. For instance, check these figures.