The hard numbers are 2200 for an 8 core (4 x 2) Opteron system vs 4000 for a 16 core (4 x 4) Barcelona, or about 1.82 scaling. On the surface, this looks pretty bad, but there are some bright spots.
First the bad, POV is a benchmark that scales really well, almost perfectly with core count. It is largely cache resident, and mainly needs raw number crunching power. POV should get much closer to 2x scaling merely going from 8 to 16 cores. That also discounts any improvement that Barcelona has over Opteron, so theoretically it should scale more than 2x. It didn't.
Back to the good. The CPUs were HE or 65W parts said to be running at the same clock speeds. Now 65W isn't a big trick with the Rev F Opterons, but 65W is a trick with a quad core Barcelona. Intel has a massively downclocked 50W Clovertown as well, so it can be done, but this is not a trivial thing.
Both systems were AMD reference platforms, big, loud and functional, but hardly speed demons. They are there to get things working solidly, not to win benchmarks. That should account for a little speed loss overall, especially as things got faster.
Where is the good? The point AMD was trying to make was not to win benches or show high end capability, but simply to show that with the current HE parts, Barcelona is a drop in upgrade with almost 100% performance increases. No more power, no more anything, just more speed.
If you consider that they both use HT 1.x, the same memory, the same power, and the same everything, 1.8x is not a bad upgrade. Going from an Intel 2C to 4C system can also get you similar scaling increases, it all depends on workload.
One other note about the benchmark itself, POV as downloaded uses a lot of x87 math. The x87 improvements from Opteron to Barcelona are somewhere between nothing and minimal. This is about the worst case scenario to show off a Barcelona, AMD has to have been mad to use it as a benchmark. Well either that or they are sandbagging.
In any case, since there are now numbers that can be worked back to a known starting point, the 8-way Opteron, I asked the good folk at the Inq Suburban Paris Research Labs (ISPRL ? pronounced like 'nonogenarian' if you were wondering) for a hard clock speed.
Simply working the numbers back gets you 1.6-1.7Ghz for the Opterons on a more performance tuned platform. All the tested numbers came up as 1.8Ghz, but we have reason to believe the demo boxes might have been a clock bin up from that.
In any case, you could do this with a store bought 1.8Ghz Opteron box so that is the numbers we will be going with. Any loss due to the platform would be equally reflected in both sets of numbers, so it seems like a fair comparison.
So now you know, 1.8GHz is doable now in a 65W AMD quad core. When HT3 parts come with Socket G, speeds will certainly improve, but you can look forward to at least a 1.8x increase with drop in parts
