Matrix multiplication on workstations

(see also: bench_bm_mm.txt)

The first test (Table 1) is simple compilation; the second one (Table 2) is "best effort".

TABLE 1. Program: p25 (simple do-loops, vector loop inside). MFLOPS is defined as 2 N**3 / (CPU time).

Machine	Opt. level	ND	N	CPU (sec)	MFLOPS
HP 710 32 MB	1	400	200	8
(ur)			400	63
	2, 3	400	200	2
			400	17	8
HP 715/75 64 MB	1	400	200	5
(jupiter)			400	39
	2, 3	400	200	1
			400	12	11
SGI Indy R4600/133 SC	1	400	200	4
64 MB (uranus)			400	37
	2	400	200	1
			400	13
	3	400	200	1
			400	13	11
SGI Indy R4600/100 PC	1	400	200
32 MB (atlas)			400	47
	2	400	200
			400	15
	3	400	200	1
			400	15	10
SGI Indy R5000/150 SC	1	400	200
64 MB (new uranus)			400
**	2	400	200
			400	9.5
	3	400	200	1
			400	9	14
Pentium Pro P6/200	1	400	200	1
64 MB (f9pc00)			400	6
	2, 3	400	200	0.6
			400	4.6	28

** Compiled on R4600/133 SC.
Increase of the CPU time if we put the summation loop to be innermost:

SGI Indy R4600/133 SC	times 2
HP 715/75	times 3.5

TABLE 2. Program: p25v (calls DGEMM).

Machine	Opt. level	ND	N	CPU (sec)	MFLOPS
SGI Indy R4600/133 SC	all	400	200	0.6
64 MB (uranus) *			400	4.2	30
		400	200
			401	4.1	30
SGI Indy R4600/100 PC	all	400	200	0.9
32 MB (atlas) *			400	6.7	19
		400	200
			401	6.7	19
SGI Indy R5000/150 SC	all	400	200	0.32	50
64 MB (uranus) * **			400	2.64	48
Pentium Pro P6/200	all	400	200	1.6
64 MB (f9pc00) *1			400	12.6	10

* Non-interleaved memory.
** Compiled on R4600/133 SC.
*1 -lblas.