STREAM results for UE6000 with 250 MHz UltraSPARC-2 (standard, no prefetch)

From: Ashok Singhal (Ashok.Singhal@Eng.Sun.COM)
Date: Thu Apr 17 1997 - 15:23:48 CDT


Here are some STREAM numbers for the UE6000
with the 250 MHz UltraSPARC-2. Most of the difference
arises not from the frequency of the processor but from the
fact that it supports multiple outstanding misses.

These are standard, auto-parallelized C using the
C compiler that's been shipping for a while
so no prefetch instructions are generated.

For the 2 CPU case I ran two cases: (a)both CPUs on same
board and (b) CPUs on different board. For the others I let the
scheduler decide.

The cache size of all CPUs on this system was 1MB.
Array sizes were increased from 1600000 (36.6 MB used) to
6400000 (146.5 MB) to make sure caches were not being helpful.
I tried a couple of different offsets for the 1 and 2 cpu cases
but didn't bother for the others (other than increasing the
array size).
This system had only 12 CPUs (but we're pretty close to the bus
limit anyway so the 16 CPU result we had earlier could
still remain).

CPUs COPY SCALE SUM TRIAD
1 281 281 224 261
2 434 429 440 500 (both CPUs on same board)
2 460 457 513 538 (CPUs on different boards)
3 694 684 603 701
4 921 900 796 891
6 1265 1257 1116 1201
8 1517 1528 1371 1501
10 1653 1741 1573 1760
12 1785 1750 1884 1848

Regards,
Ashok



This archive was generated by hypermail 2b29 : Tue Apr 18 2000 - 05:23:06 CDT