---

Single versus Double precision
(60,64 or even 48 bits - but not 32 bits - versus 120 and 128 bits).

• CDC Cyber 170/835 :

  compiler	FTN	FTN	FTN	FTN5	FTN5	FTN5
  opt	opt=0	opt=1	opt=2	opt=0	opt=1	opt=2
  single	977.56	783.10	658.84	1477.74	840.72	718.28
  double	3756.22	3697.49	3404.08	4124.88	3568.91	3436.58

• CDC Cyber 170/875, FTN5, opt=2: double/single = 382.667/128.349 = 298%
• ETA 10P, SystemV 1.0b, FTN77 1.5c: d/s = 835.9/xxx.x
• ETA 10Q, SystemV 1.1, FTN77 1.xx: d/s = 548.2/150.2
• Cray X/MP-216 (SN:402), UNICOS 4.0, cft77 2.0.19: 641.54/39.99; cft 1.15: 648.77/55.67
• Cray X/MP-28 (SN:xxx), UNICOS 4.0, cft77 2.0.19: 558.61/32.38; cft 1.15: 564.99/45.47
• Cray 2 (SN:2027), UNICOS 5.0.8, cft77 3.1: 781.41/41.49; cft 3.1b: 899.30/56.06
• Dec/Vax machines, Fortran v4, /G precision versus /D precision :
  • 11/780: 189414/2097 = 9032%;
  • 11/785: 6812/1296;
  • muVax2: 2315/2247;
  • 8530: 594/474.
  Only on the Vax 11/780 is the G precision emulated.
• 64/48-bits precision: ModComp 11/45 = 3559/3009, 11/75 = 2449/1542, 32/85 = 1806/2924.

---

[Return]