Compaq Deskpro 386/33L with two floating-point units : multi-coprocessing ?

• The Compaq Deskpro box is "386/33L" with an 386 daughter board (having 64 K cache and 2 separate sockets for an x87 and a Weitek 3167 or the Cx EMC87) or "486/33L" with an 486 daughter board .

• (16 bits Fortran) MSFortran, Lahey, Prospero, RMFarland ... ref.time = DAT35; (extract from) :

  compiler	Prospero	MicroSoft	IBM IBM RM	Lahey	Digital
  compiler	f66 f77		Pro. F/2 Farland		Research
  version	2.144 1.241	3.31 4.0 5.0	1.0 1.0 2.11	2.04 2.22 3.00	4.1
  let t1 be time with CxEMC87; t2 be time with x87 and t3 be time with x87 and CxEMC87
  Cx486DRx2 + CxEMC87	370 151	224 202 223	xxx 184 121	136 134 131	172
  for x87 = 387DX or CT38700 : t3 is smaller than min (t1, t2) and the gain may reach 20%
  Cx486DRx2 + 387DX	396 169	239 215 230	xxx 210 147	159 154 150	199
  Cx486DRx2 + 387DX + EMC87	290 134	183 163 180	xxx xxx 112	123 120 118	154
  Cx486DRx2 + 38700	371 153	224 215 230	xxx 186 125	137 136 133	171
  Cx486DRx2 + 38700 + EMC87	292 xxx	182 185 179	xxx xxx 112	xxx 127 117	155
  for x87 = 3C87 or Cx387+ : t2 is equal to t3 : the 2nd fpu has no influence
  Cx486DRx2 + 3C87	385 165	238 214 234	xxx 434 132	151 146 141	191
  Cx486DRx2 + 3C87 + EMC87	385 166	xxx 214 235	xxx xxx 134	152 148 144	191
  Cx486DRx2 + Cx387+	292 138	184 166 185	xxx xxx 115	127 123 122	158
  Cx486DRx2 + Cx387+ + EMC87	296 140	186 168 187	xxx xxx 117	129 125 123	161
  for x87 = 4C87DLC : t3 is greater than max (t1, t2) and the loss may be severe
  Cx486DRx2 + 4C87DLC	308 151	197 177 196	xxx xxx 127	140 134 133	180
  Cx486DRx2 + 4C87DLC + EMC87	387 169	xxx xxx 240	xxx xxx 136	152 150 145	196

• (32 bits Fortran) SVS and MicroWay, ref.time = DAT35; (extract from) :

  (Dos32) compiler	SVS 2.4	mf77 3.0
  Cx486DRx2 + CxEMC87	462	469 468
  for x87 = 387DX or CT38700 : t3 is smaller than min (t1, t2) and the gain may reach 20%
  Cx486DRx2 + 387DX	524	540 539
  Cx486DRx2 + 387DX + EMC87	437	453 452
  Cx486DRx2 + 38700	471	488 487
  Cx486DRx2 + 38700 + EMC87	437	452 451

• (32 bits Fortran) PowerStationFortran 1.0, ref.time = DAT50; (extract from) :

  compiler switches	g3	g3 oP	g3 oX	g4	g4 oP	g4 oX
  Cx486DRx2 + CxEMC87	536	549	457	537	550	459
  for x87 = 3C87 or Cx387+ : t2 is equal to t3 : the 2nd fpu has no influence
  Cx486DRx2 + Cx387+	527	537	470	527	538	472
  Cx486DRx2 + Cx387+ + EMC87	529	539	472	528	540	475
  for x87 = 4C87DLC : t3 is greater than max (t1, t2) and the loss may be severe
  Cx486DRx2 + 4C87DLC	581	594	521	589	594	522
  Cx486DRx2 + 4C87DLC + EMC87	621	634	533	620	635	534

• (32 bits Fortran) WatCom 9.01, ref.time = DAT50; (extract from) :

  (Dos32) Watcom version	9.01				9.5	10.5
  source	original								FOREST
  compiler switches	4		3 oT		4 oT		4 fp3 oX	5 fp5 oX
  floating point instruction	no	yes	no	yes
  ref.time = DAT35
  let t1 be time with CxEMC87; t2 be time with x87 and t3 be time with x87 and CxEMC87
  Cx486DRx2 + CxEMC87	1414	436	1419	abend	xxxx	xxxx	xxxx	xxxx	xxxx
  for x87 = 387DX or CT38700 : t3 is smaller than min (t1, t2) and the gain may reach 20%
  Cx486DRx2 + 387DX	1495	552		abend
  Cx486DRx2 + 387DX + EMC87	1117	418	1132	abend
  Cx486DRx2 + 38700	1390	453	1396	446
  Cx486DRx2 + 38700 + EMC87	1244	418	1266	411
  for x87 = 3C87 or 4C87DLC or Cx387+ : the 2nd fpu has no influence or there is some performance loss
  Cx486DRx2 + 3C87	1234	500	1248	abend
  Cx486DRx2 + 3C87 + EMC87	1516		1521	513
  Cx486DRx2 + 4C87DLC	1321	625	1135	abend
  Cx486DRx2 + 4C87DLC + EMC87	1611	643	1616	abend
  Cx486DRx2 + Cx387+	1132	445	1147	abend
  Cx486DRx2 + Cx387+ + EMC87	1155	650	1171	abend

• (16 bits C) MicroSoft C 5.1, 6.0 and 7.0, ref.time = DAT35; (extract from) :

  Microsoft C version	3.0		4.0		5.1	6.00A		7.0
  switch	fpa	fpc87	fpa	fpi87	fpc87	fpa	fpc87	fpa	fpc87
  switch	g2				g0		g0		g0 g2
  let t1 be time with CxEMC87; t2 be time with x87 and t3 be time with x87 and CxEMC87
  Cx486DRx2 + CxEMC87					228		220		106 101
  for x87 = 387DX or CT38700 : t3 is smaller than min (t1, t2) and the gain may reach 20%
  Cx486DRx2 + 387DX					235		226		127 122
  Cx486DRx2 + 387DX + EMC87					187		181		97 94
  Cx486DRx2 + 38700					231	917	221	926	109 104
  Cx486DRx2 + 38700 + EMC87					187		179		97 93
  for x87 = 3C87 or Cx387+ : t2 is equal to t3 : the 2nd fpu has no influence
  Cx486DRx2 + 3C87							231		111 114
  Cx486DRx2 + 3C87 + EMC87					239		232		120 116
  Cx486DRx2 + Cx387+					192		184		101 98
  Cx486DRx2 + Cx387+ + EMC87					194		186		102 99
  for x87 = 4C87DLC : t3 is greater than max (t1, t2) and the loss may be significant
  Cx486DRx2 + 4C87DLC					203	696	196	694	112 108
  Cx486DRx2 + 4C87DLC + EMC87					246		234		121 115

• (16 bits C) MicroSoft C 8.0, ref.time = DAT35; (extract from) :

  compiler switch	fpa			fpi87			fpi87 o2
  compiler switches	g0	g2	g3	g2 o2	g3 o2	g3	g0	g1	g2	g3
  let t1 be time with CxEMC87; t2 be time with x87 and t3 be time with x87 and CxEMC87
  Cx486DRx2 + CxEMC87				99	94	104
  for x87 = 387DX or CT38700 : t3 is smaller than min (t1, t2) and the gain may reach 10%
  Cx486DRx2 + 387DX				119	115	121
  Cx486DRx2 + 387DX + EMC87				91	88	94
  Cx486DRx2 + 38700	926		914
  Cx486DRx2 + 38700 + EMC87				91	89	94
  for x87 = 3C87 or Cx387+ : t2 is equal to t3 : the 2nd fpu has no influence
  Cx486DRx2 + 3C87				114	109	114
  Cx486DRx2 + 3C87 + EMC87				114	109	115
  Cx486DRx2 + Cx387+				94	92	97
  Cx486DRx2 + Cx387+ + EMC87				96	93	99

• The influence of the second floating-point unit is smaller or even nil if the CPU is among [ 386, CT38600, Ti486DLC ] than with the Cx486DRx2 .
• Please give your comments !