Though I like your price/speed formula, I must object to the statement "Because FLOPs/s is a better judge of speed that Mhz, in my experience", not because FLOPs is worse than Mhz (it isn't), but more because they are both USELESS. MHz is pointless (remeber that comparison of an Alpha which beat the pants of an Athlon at twice the clock speed). FLOPs are pointless too. Suppose we take a segment of code which is one million floating poitn operations, all independent. And we ran the code on two processor.
Both processors are 4 wide issue, 4 wide fetch, 4 wide retire (or 4-way superscalar of the Intel junkies). Also assume that it always has enough functial units to always execute the code (in other words, the code has been chosen specifically to run on the CPU). Now, one of these processors is in-order, and the other is out-of-order. If you ran the code on both processors at the same speed, it would take the same amount of time to finish. Therefore, the number of FLOPs you could use to describe each processor is the same.
Now suppose you ran some real world code on both chips. Almost exhaustively floating point, but it actually DID something (i.e. had dependencies and branches and stuff). The out-of-order processor would probably finish 60% faster than the in-order one.
FLOPs aren't a good measure of speed. I personally trust the SPEC benchmarks a bit more, since they have rules about how much your compiler is allowed to cheat in its score.
As a side note, can anyone find a version of this article (or one of similar subject matter) in English? The German translation just doesn't cut it for me.
I had a similar problem. In my case, when I played mp3's, my SCSI CD-R drive went crazy. It turns out there was a memory conflict between my Plug 'n' Pray SB64 and my ethernet card (don't ask me how my CD-R drive got envolved, I don't really have any idea). Anyway, I too was absolutely certain that there were no io/irq conflicts, until I logged in at console and tried using mpg123 (I think I actually logged in as root), and I got this bizarre console message indicating that there was a problem. Anyway, you might want to double and triple check for io conflicts. Note I had the same problem under both '95 and NT at one point or another.
Slashdot Mirror
Though I like your price/speed formula, I must object to the statement "Because FLOPs/s is a better judge of speed that Mhz, in my experience", not because FLOPs is worse than Mhz (it isn't), but more because they are both USELESS. MHz is pointless (remeber that comparison of an Alpha which beat the pants of an Athlon at twice the clock speed). FLOPs are pointless too. Suppose we take a segment of code which is one million floating poitn operations, all independent. And we ran the code on two processor.
Both processors are 4 wide issue, 4 wide fetch, 4 wide retire (or 4-way superscalar of the Intel junkies). Also assume that it always has enough functial units to always execute the code (in other words, the code has been chosen specifically to run on the CPU). Now, one of these processors is in-order, and the other is out-of-order. If you ran the code on both processors at the same speed, it would take the same amount of time to finish. Therefore, the number of FLOPs you could use to describe each processor is the same.
Now suppose you ran some real world code on both chips. Almost exhaustively floating point, but it actually DID something (i.e. had dependencies and branches and stuff). The out-of-order processor would probably finish 60% faster than the in-order one.
FLOPs aren't a good measure of speed. I personally trust the SPEC benchmarks a bit more, since they have rules about how much your compiler is allowed to cheat in its score.
As a side note, can anyone find a version of this article (or one of similar subject matter) in English? The German translation just doesn't cut it for me.
I had a similar problem. In my case, when I played mp3's, my SCSI CD-R drive went crazy. It turns out there was a memory conflict between my Plug 'n' Pray SB64 and my ethernet card (don't ask me how my CD-R drive got envolved, I don't really have any idea). Anyway, I too was absolutely certain that there were no io/irq conflicts, until I logged in at console and tried using mpg123 (I think I actually logged in as root), and I got this bizarre console message indicating that there was a problem. Anyway, you might want to double and triple check for io conflicts. Note I had the same problem under both '95 and NT at one point or another.