Two CPU Comparisons are given:
1)  Electronic-Structure SCF run
2) VECTORIZED Electronic-Structure Convolution Integral.
 

• KKR-CPA Electronic Structure Method

Details:
I have ran my OLD KKR-CPA code on several platforms to compare performance.
I have done one interation of FCC NM Fe90S (requiring lots of CPA)
Code is a mixed cluster-real-space + BZ integration method for speed.
Cluster method used 6 real-space shells in FCC
BZ method is a ray-prism method (similar to special directions).
(Note: cluster is more scalar, but a fast method BZ method was written
highly vectorized for CRAY. However, much longer is spent in BZ method,
but more E's are done cluster method, so there is a large averaging out.)

Machine	Time (secs)	Compiler Options
SUN Hyper Sparc	2580	-dalign -native -O
J-CRAY	1045	-a static -dp
SUN Ultra 170 Mhz	936	-dalign -native -O3
HP              120 Mhz	870	+E7 -K  -O
CRAY YMP	553	-a static -dp
IBM      590	565	-O3 -qdpc -qstrict -qtune=pwr2                                                   -qarch=pwr2    -lesslp2
DEC ALPHA 267 MHZ WS	354	-O -tune host -fast
DEC ALPHA 400 MHZ WS	243	-O -tune host -fast
DEC ALPHA 500MHZ  P-AU	184	-O -tune host -fast
DEC ALPHA 600 MHZ

 

• A CONVOLUTION INTEGRAL

A First-Principles, Concentration-Wave Calculation is a KKR-CPA-based approach for a complex
convolution integral  [i.e., integrate over k: T(k+q; E)T(k; E)].   The code is highly vectorized, so the CRAY
C90 should be by far the fastest machine. (FOR F77 Compiler and 1 q and 1 Energy point.)

---

Machine	Time (mins)	Compiler Options
SUN  Sparc  10/50	8:03	-dalign -native -O3
SUN Hyper Sparc	4:06	-dalign -native -O3
SUN Ultra 170 Mhz	2:00	-dalign -native -O3
HP    J2-10 120 Mhz	2:27	-O
SGI RS-8000 pwr challenge	1:37	-O3 -static  one part w/ -O1
IBM      590	1:24	-O3 -qdpc -qstrict
DEC ALPHA 267 MHZ WS	1:24	-O -tune host -fast
DEC ALPHA 400 MHZ WS	1:02	-O -tune host -fast
DEC ALPHA 500MHZ  P-AU	0:46	-O -tune host -fast
DEC ALPHA 600 MHZ
CRAY YMP-C90	0:23	-dp -static