I have re-organized the files related to my "stream" memory bandwidth
benchmark, and included some new results from the Cray J916 with up to
8 cpus.
The Cray J916 machine is serial number 9003, and so should probably be
considered a production machine. It has full J916 memory subsystem, as
opposed to the previous results with only half of the banks installed.
I am referring to those previous prototype results as J98 (instead of
the J94 that was previously listed in my tables).
In the reorganized tables, I have combined the single-cpu and parallel
results into a single file to make my life easier, and I have
reformatted the tables so that they are much easier to read. I removed
the sorted tables, since they were so hard to read. If you want to
sort the results, feel free to do it yourself....
I have also included some derived metrics, like bandwidth expressed in
bytes/cycle, and "machine balance" which is the ratio of sustainable
memory bandwidth (in words/cycle) to FP performance (in FP ops/cycle).
I considered removing the old results for machines no longer available,
but I left them in for history's sake. It is very interesting to see
that most of the current cache-based machines have memory transfer
rates that are really not much faster than machines of 4-5 years ago.
More results are always welcome --- especially on high-end shared memory
machines. Right now I am most interested in:
Cray 2, Cray 3, Cray 4
NEC SX series
Convex C4 series
Convex Exemplar
Fujitsu supercomputers (any)
Hitachi supercomputers (any)
-- -- John D. McCalpin mccalpin@perelandra.cms.udel.edu Assistant Professor, College of Marine Studies, Univ. of Delaware
This archive was generated by hypermail 2b29 : Tue Apr 18 2000 - 05:23:04 CDT