Jim, I don't think those scores that you posted can be valid scores.
It's really hard for me to believe that super-confidential performance
data from AMD and Intel would appear in a EE 300-level class.
Anyway, for the moment, let's assume that the benchmarks are genuine.
IPC stands for "Instructions Per Clock". The higher the IPC, the more
performance you'll get at the same clock speed. L1 and L2 Miss Ratio
is a measure of how well the caches do their job. Larger caches mean
lower miss ratios, which mean fewer memory accesses have to go out to
a slower memory subsystem.
From your post, it seems that the IPC for the K7 and for Coppermine
vary widely according to benchmark:
IPC for: uncompress expresso fft Alt. Bench.
K7: 1.413 1.104 1.143 1.257
CuMine: 2.066 1.758 1.133 0.806
This shows that for equivalent clock speeds, Coppermine has better
performance on two benchmarks, equivalent performance on one, and
lower performance on the "Alternative Benchmark."
The L1 and L2 miss ratios, however, is no surprise. Coppermine
consistently shows slightly higher miss ratios in both the L1 and L2
caches. This is because the L1 cache on CuMine (32K or 64K?) is
smaller than the oversized 128K L1 cache of K7, and the 256K L2 cache
on CuMine is also bound to be smaller than the minimum L2 cache size
(512K) for K7. The miss ratios, in general, are very small because it
seems that the benchmarks are very small programs themselves and thus
can easily fit itself and most of the data it works on in the L1 and
L2 caches.
So if this document is the real thing, that means Coppermine will beat
the K7 in some benchmarks, and fall behind in others. Of course, the
veracity of this document depends on whether you trust inside
information to come from a Stanford undergrad EE class. ;-)
Tenchusatsu |
