"Equivalent Power (est.)"

[Lagardo]

OK, so everybody knows that that column doesn't mean much. As far as I can figure it out, the "equivalent power" is over-estimated by a factor of three to five. Now what I'd like to know is whether this is true for everybody, i.e. whether there is anybody that finds that this estimate is too low. For if not, we ought to be able to collect a couple items of crude data of "real power vs. estimated power" and figure out a correction factor that'll make this estimated power a little less glaringly false.

I'm not trying to get precision here, I simply figure we (collectively) ought to be able to do better than to within an order of magnitude. Maybe we could get it to within a factor of two. Or maybe even better.

My P4 is over-estimated around 10-11GHz even though it runs at 2.8 (that's a factor 3.5 or 4 there) and my AthlonXP comes out a little over 5GHz even though it runs at 1.6 which comes to 3.1 -- so judging only from my two boxes, the "equivalent power" should be decreased by a factor of 3.25 or so and would match reality much better.

Anybody got very different numbers?

[Edit: typos, typos...]

[Marco_N]

AFAIK, the "equivalent" is based on the Pentium I 90Mhz level of performance. My current P4 Northwood has an improved & faster architecture, larger caches and offers SSE2 SIMD instructions. I also have PC1066 RAMBUS memory, the P90 had EDO RAM if you were lucky. It seems logical that a P90 system has to be clocked a lot higher than 2.26GHz to achieve the same level of cEMs/s performance.
The point you raise, is the factor 3-4 realistic? I agree it seems high for a "generic applications" speed comparison, but that could be explained by the specifics of the SoB application; it's highly optimised for speed. I don't agree it should be around 1. Maybe a "benchmark" option like the (cough)distributed.net(cough) client has have could shed some light on the effects of SIMD and other P4 / Athlon optimized instructions.

- Marco

[edit]
P.S. some insight can be gained from this table:
http://www.mersenne.org/bench.htm

P.P.S.
The values I get in the client are confusing. Stopping, exitting, relaunching, then restarting the client sometimes has a remarkable effect on the displayed cEM/sec rate. Note: this will cause some loss of work (e.g. you'll restart at 11.8% instead of 12.2%). I don't know whether that is a "real" effect or just a problem in measurements.

[eatmadustch]

that will be one point. The more important one, however, is that the cEM/s climbs the higher n gets. This is because the cEM was made for the older core. The newer one is more efficient with higher numbers, so obviously is seems as if your computer is faster. If you to a test with n=100000 then the cEM/s will sink to a very low value. The GHz is based on cEM/s, so it's also not accurate!