This begs the question, was it really faster to store that stuff (e.g. skinned meshes, shadow volumes) back in 2004 when Doom3 came out? Has the hardware really changed that much in the 8 or so years between the original and the BFG releases? Or was this another case of premature optimization and doing things the way they were always done before?
In addition to memory bandwidth vs. cpu throughput, the number of cores in your average consumer cpu has gone up from 1-2 to 4 (or "8 cores" with hyperthreading). Reading between the lines, this was a motivator for the changes. Sharing memory between the cores is expensive and clearly some of the changes were made in order to reduce sharing between the CPUs.
So if any of the Doom 3 coders are reading this, can you give some insight on these changes? :)
A low-end desktop Sandy-Bridge i3 CPU running a single-threaded task on one core is almost 5 times faster than a Venice Athlon 64 singlecore that was released in 2005 if memory serves right. When you factor in the second core and HyperThreading (if it helps the workload) it wouldn't be unreasonable to claim that performance has increased by an order of magnitude or more.
But I looked, and holy crap: released 2004. I don't know why, but that makes me feel particularly old.