high resistive losses, but the thicker the copper, the
more difficult it is to make it flex for millions of cycles
without stress fatiguing in the tight bend radius re quired by the small package. The cable materials and adhesives all affect the fatigue life of the copper. After many iterations and extensive testing, a Kapton- copper-Kapton sandwich laminated with special adhesive was selected. Because the cable connectors are copper and the printhead interconnects are thin-film aluminum, conventional solder bonding was not possible. Ultra sonic-wire or stitch bonding provided the answer. Ac k n o wl e d g me n t s While there are numerous people who contributed to the development of this project we would be remiss in not giving special recognition to: Chuck Dodge for his superb industrial design and Craig Sanford who developed the mechanical package; Paul Stoft of HPL who encouraged the initial printer effort and those in the Loveland, San Diego, Avondale, and Colorado Springs Labs who provided helpful early advice; Frank Ura, Per Gloersen, Glenn Weberg and Bill Ebert of HPL and Bill Sperry of San Diego Divi sion who developed the thin film print head process, and Robert Waits, Fred Scheu, and Ken Scholz who set up the process at APD; Ken Salisbury, Manny Kohli, and Dick Barth who designed the printer mechanism; Al Kovalick who designed the printer drive electronics using in part a motor speed control invented by Bill Egbert; Jim Steinmetz for the flex cable and head bonding technique; Larry Jackson and Tom Braun for printer and print head testing; John Uebbing, Perry Jeung, and John Fischer of OED who developed our attractive display and to She-Chung Sun for the cathode driver and display driver cir cuitry; Dick Osgood and Bill Mitchell who designed the keyboard; Ken Hodor who designed the PICK circuit and later helped put the product into produc tion; Dennis Harms and Paul Williams who wrote the software; Carl Landsness who designed the power supply and recharger; Dick Leininger, Hal Andrews, and especially Bob Meadows who served as product manufacturing engineers on the project; Roger Evans and Jim Fleming at APD and Bernard Tsai in Singa pore who provided the production engineering sup port; Bill Bull and his crew on the production line; Don Grant and the QA group; the manufacturing divi sion tool designers and builders whose state-of-the- art injection molding methods help make the product possible; Bob Dudley and Chung Tung for their managerial support. E The New Accuracy: Making 2=8 by Denni s W. Har ms One of t he engi neer i ng ef f or t s t hat made t he HP- 35 possi bl e was t he devel opment of t he al gor i t hms f or t he mat hemat i cal f unct i ons. Thi s wor k was done by Davi d S. Cochr an. 1 These al gori t hms have served t he HP hand-hel d cal cul at or f ami l y wel l , but have now been r epl aced by mor e accur at e ones on t he HP- 91 . As one of t he mor e obvi ous exampl es of t he i mpr ove ment , 23 = 8. 000000003 on t he HP-35 but 23 = 8000000000 on t he HP-91 . Ot her exampl es of t he i mprovement are shown i n t he t abl e bel ow: O P E R A T I O N H P - 3 5 H P - 9 1 1 0 5 2 9 9 9 9 9 9 9 9 6 4 x 1 0 5 1 1 0 0 0 0 0 0 0 0 0 x 1 0 5 2 2 5 5 9 7 6 5 6 2 5 0 5 5 9 7 6 5 6 2 5 0 0 0 9 5 5 9 0 4 9 0 0 0 2 0 5 9 0 4 9 0 0 0 0 0 1 5 5 7 5 9 3 7 5 0 0 3 4 7 5 9 3 7 5 0 0 0 0 < - 2 ) 3 e r r o r - 8 0 0 0 0 0 0 0 0 0 I n ( 1 . 0 0 0 0 1 ) 9 9 9 9 9 0 0 0 0 0 x 1 0 ~ 6 I n ( 99995) t an 89 999999s tan 90 000001 sin 000 OOOOOOr sin 359 9999999 tan 720 t an ( 720 x i 90) ' 9 9 9 9 9 5 0 0 0 0 x 1 0 - 5 0 0 0 1 2 0 0 0 0 x 1 0 ~ 5 - 5 0 0 0 1 2 5 0 0 4 x 1 0 ~ 5 5 8 3 6 3 5 0 0 . 1 6 5 7 2 9 5 7 7 9 . 5 1 - 5 8 8 2 3 5 2 9 4 1 - 5 7 2 9 5 7 7 9 5 1 1 7 x 1 0 " 9 1 . 7 4 5 3 2 9 2 5 2 X 1 0 " 9 - 2 1 8 4 4 1 3 8 7 9 - 1 7 4 5 3 2 9 2 5 2 x 1 0 ~ 9 4 0 0 0 0 0 0 0 0 0 x 1 0 ~ 9 0 0 0 0 0 0 0 0 0 0 2 0 2 9 7 5 2 0 9 0 0 0 0 0 0 0 0 0 0 0 The over r i di ng consi der at i on i n t he HP- 35 was savi ng ROM ( r ead- onl y memor y) . Onl y t hr ee ROMs of mi cr ocode wer e used t o i mpl ement al l t he f unct i ons of t he cal cul at or. Bet t er accuracy, comparabl e t o t hat of t he HP-91 , woul d have consumed t wo addi t i onal ROMs and woul d have i ncreased bot h t he si ze and cost of t he HP-35. Si nce t he sl i de rul e was t he st andard of t he day, t he ' Th i s d i s c u s s i o n i s e s s e n t i a l l y t h e s a me a s a p a p e r p r e s e n t e d b y t h e a u t h o r a t El e c t r o 7 6 i n Bo s t o n . Ma s s a c h u s e t t s . U. S. A. . Ma y 1 1 - 1 4 . 1 9 7 6 . i mpr ovement i n accur acy bought by t hese addi t i onal ROMs di d not seem wor t h t he expense. Two t hi ngs have happened t o make t he addi t i onal ROM wort h t he cost . Fi r st , ROM has become cheaper and smal l er by a f act or of f our . Second, cal cul at or s ar e now bei ng appl i ed mor e wi del y t han ori gi nal l y ant i ci pat ed, and users now have a hi gher l evel of sophi st i cat i on and expect t he same f r om t hei r cal cu l at or s. Pr ogr ammabl e cal cul at or s, especi al l y, r equi r e gr eat er accur acy. The user of a nonpr ogr ammabl e cal cul at or moni t or s hi s cal cul at i on st ep by st ep, so unr easonabl e i nt er medi at e resul t s caused by al gori t hm i nadequaci es are recogni zed when t hey appear . However , when a cal cul at or i s pr ogr ammabl e, t he user i s of t en awar e of onl y t he i nput and t he f i nal answer . Avoi dabl e and Unavoi dabl e Er r or s Di spl ayed er r or s can be di vi ded i nt o t wo t ypes, avoi dabl e and unavoi dabl e. The f i r st i s caused by i nadequaci es i n t he al gori t hms. Exampl es of t hi s t ype of error are shown i n t he t abl e above and ar e t ypi f i ed by 23 = 8. 000000003. Thi s t ype of er r or has been al most compl et el y el i mi nat ed f r om t he HP- 91 . The second t ype of er r or i s t he unavoi dabl e er r or caused by usi ng f i ni t e comput er s t o appr oxi mat e nonf i ni t e pr ocesses. For i nstance, the deci mal representati on of 1 1I i s the repeati ng non- f i ni t e deci mal : 1/ 7 = 0. 142857 142857 142857 . . . whi ch cannot be r epr esent ed exact l y wi t h t en f i gur es. The cl osest appr oxi mat i ons ar e: 0. 142857 1429 ( cl osest ) and 0. 142857 1428 ( next cl osest ) . 16 Copr. 1949-1998 Hewlett-Packard Co. When we wr i t e 1/ 7 = 0. 142857 1429 we ar e l yi ng onl y by a l i t t l e bi t . Unf or t unat el y, 1/ ( 0. 142857 1429) = 6. 99999 99979 00000 00062 99. . . whi ch r ounds t o 6. 99999 9998 i nst ead of 7. Ther ef or e, t he cal cul at or cannot sat i sf y t he i dent i t y 1/(1/7) = 7. Thi s i s not a mi st ake. I t i s an unavoi dabl e consequence of roundi ng. There i s no t en-f i gure approxi mat i on t o 1/ 7 whose reci procal rounds t o 7. To see t hi s, l et us l ook at t he next cl osest approxi mat i on t o 1/ 7. 1/ ( 0. 142857 1428) = 7. 00000 00028 00000 00112 . . . whi ch rounds to 7.00000 0003 i nstead of 7. Thi s i s farther from 7 t han t he r eci pr ocal of t he ot her appr oxi mat i on t o 1/ 7. Per haps t he best way t o avoi d t he conf usi ng st at ement " 1/ ( 1/ 7) = 6. 99999 9998" i s to wri te [ 1/ ( 1/ 7) 10] 10 = 6. 999999998 wher e t he subscr i pt 10 i ndi cat es a r oundi ng t o 10 pl aces. How I mpr ovement s Wer e Achi eved Accur acy i n t he HP- 91 has been i mpr oved pr i nci pal l y by r e wri t i ng t he al gori t hms' subrout i ne st ruct ure. An exampl e of t hi s i s wi t h t he yx r out i ne. Al l HP cal cul at or s cal cul at e yx by t he f ormul a y X _ e x I n y Thi s i s eval uat ed by subr out i ne cal l s t o I n, mul t i pl y, and ex ponent i al On t he HP-35 t hese subrout i nes were i dent i cal t o t he keyboar d f unct i ons of t he same name. The f unct i on yx di d t he same t hi ng as t he key sequence x^y, In, x, ex I t i s not surpri si ng t hen t hat 23 = 8. 000000003, f or t hi s answer has been subj ect t o t hree i nt ermedi at e roundi ngs. The i mprove ment was made by r ewr i t i ng t he basi c subr out i nes t o al l ow f or car r yi ng i nt er medi at e r esul t s t o ext r a di gi t s. Enough ext r a di gi t s ar e c ar r i ed and c ar ef ul l y l ook ed af t er t o as s ur e t he desi r ed accur acy. Speci al subr out i nes of t hi s t ype ar e al so used i n t he t r i gonomet r i c f unct i ons, i n r ect angul ar - t o- pol ar conver si ons, and i n pol ar - t o- r ect angul ar conver si ons. The second met hod of i mpr ovi ng accur acy i s t o t r ap cr i t i cal ar gument s and cal cul at e t he f unct i ons at t hese ar gument s i n a speci al way. These cr i t i cal ar gument s i ncl ude number s near 1 when cal cul at i ng I n or l og, number s near 0 when cal cul at i ng s i n " 1 , c o s " 1 , o r t a n " 1 , a n d n u mb e r s n e a r z e r o o r mu l t i pl es of 7T/ 2 when cal cul at i ng si n, cos, or t an. Anot her i mpr ovement was made by pr escal i ng t he t r i gono met r i c f unct i ons i n t he user ' s uni t s. The HP- 35 conver t ed de grees t o radi ans and t hen subt ract ed mul t i pl es of 2i r t o convert t he i nput ar gument t o t he uni t ci r cl e. Thi s made t an 720 = 4 x 10~9. On t he HP- 91, i f t he c al c ul at or i s i n degr ees or grads, mul t i pl es of 360 or 400 respect i vel y are subt ract ed unt i l t he resul t i s wi t hi n t he uni t ci rcl e. Thi s makes t an 720 = 0 and i mpr oves symmet r y pr oper t i es of t he t r i gonomet r i c f unct i ons. Possi bl e I mprovement s Ther e i s a poi nt wher e t he new al gor i t hms appear at f i r st gl ance t o be l ess accurat e. I f we cal cul at e t an I T i n radi ans, t he answer i s 4. 10 x 10~10ont he HP- 91 wher eas t he HP- 55 yi el ds 0. Thi s i s mi sl eadi ng, f or i t on t he keyboar d shoul d r eal l y be l abel ed JTI O, showi ng t hat i t i s t en di gi t s of -n. Thus 7r 10= 3. 141592654 To ten digits, tan w10 = 4. 1 2067 61 50 x 1 0~ 10. One can see that t he HP- 91 gi ves t he bet t er answer t o t hi s pr obl em. The r eason t hat t he HP- 91 does not gi ve t he cor r ect answer i s t hat pr escal i ng i s done wi t h a f i ni t e- l engt h I T. To gi ve t he correct answer to thi s probl em, a i t of l ength 20 woul d have to be used. I t can be shown t hat t o produce t he correct answer t o al l t ri gonomet ri c probl ems i n radi an measure, i t t akes a w of l engt h 120. The HP- 91 act ual l y uses 13 di gi t s of i r . Anot her pl ace wher e t he accur acy can be i mpr oved i s i n t he st at i st i cal f unct i ons I f t he f ol l owi ng dat a i s put i nt o t he HP- 55 us i ng t he 1+ k ey , t he s t andar d dev i at i on i s not c al c ul at ed exact l y. 6666666123 6666666246 6666666369 The st andar d devi at i on cal cul at ed i s 0. 000000000 i nst ead of t he correct val ue of 1 23. 0000000. Thi s probl em coul d be sol ved by st or i ng di f f er ent sums i n t he st at i st i cs r egi st er s. However , t he sums t hat ar e now st or ed ar e usef ul f or vect or ar i t hmet i c. The user can usual l y avoi d t hi s pr obl em by l eavi ng of f t he r edundant l eadi ng si xes. I f t he dat a i s keyed i n as f ol l ows, t he cor r ect st andar d devi at i on i s cal cul at ed. 123 246 369 Acknowl edgment s I woul d l i ke t o express my appreci at i on t o Dave Cochran and Pr of essor Wi l l i am Kahn, who made t hemsel ves avai l abl e f or consul t at i on, and Bi l l Egber t f or hi s wor k i n document i ng and f l ow char t i ng t he or i gi nal al gor i t hms. Ref erences 1 D S Cochr an, " Al gor i t hms and Accur acy i n t he HP- 35. Hewl et t - Packar d Jour nal , June 1972. 2 R B Nef f and L. Ti l l man. " Thr ee New Pocket Cal cul at or s Smal l er . Less Cost l y, Mor e Power f ul . Hewl et t - Packar d Jour nal , November 1975 3 P H St er benk, " Fl oat i ng Poi nt Comput at i on , Pr ent i ce Hal l . 1974 Denni s W. Har ms Denni s Harms was born, grew up, and at t ended col l ege i n Waver l y, I owa, gr aduat i ng f r om War t bur g Col l ege i n 1969 wi t h a BS degr ee i n mat hemat i cs. He went on t o r ecei ve hi s MS and PhD degr ees f r om I owa St at e Uni ver si t y i n 1970 and 1974, t eachi ng under gr aduat es cal cul us and f i ni t e ^ V . x m a t h e m a t i c s w h i l e p u r s u i n g h i s own studi es He j oi ned HP i n 1 973, wr ot e many appl i cat i ons pr o gr ams f or t he HP- 65 and HP- 55, di d mi cr opr ogr ammi ng f or t he k H P - 9 1 , a n d i m p r o v e d t h e a l gor i t hms used i n HP hand- hel d cal cul at or s. He' s a member of t he Amer i can Mat hemat i cal Soci et y. Denni s r el axes by pl ayi ng dupl i cat e br i dge, basket bal l , and t enni s. He' s mar r i ed t o a l i brari an and l i ves i n Sant a Cl ara. Cal i f orni a. 17 Copr. 1949-1998 Hewlett-Packard Co.