IBM hexadecimal floating-point

A single-precision HFP number (called "short" by IBM) is stored in a 32-bit word:

1		7			24		(width in bits)
S	Exp			Fraction
31	30	...	24	23	...	0	(bit index)*
* IBM documentation numbers the bits from left to right, so that the most significant bit is designated as bit number 0.

In this format the initial bit is not suppressed, and the radix (hexadecimal) point is set to the left of the significand (fraction in IBM documentation and the figures).

Since the base is 16, the exponent in this form is about twice as large as the equivalent in IEEE 754, in order to have similar exponent range in binary, 9 exponent bits would be required.

Since the base is 16, there can be up to three leading zero bits in the binary significand. That means when the number is converted into binary, there can be as few as 21 bits of precision. Because of the "wobbling precision" effect, this can cause some calculations to be very inaccurate. This has caused considerable criticism.[6]

A good example of the inaccuracy is representation of decimal value 0.1. It has no exact binary or hexadecimal representation. In hexadecimal format, it is represented as 0.19999999...₁₆ or 0.0001 1001 1001 1001 1001 1001 1001...₂, that is:

S	Exp	Fraction
0	100 0000	0001 1001 1001 1001 1001 1010

This has only 21 bits, whereas the binary version has 24 bits of precision.

Six hexadecimal digits of precision is roughly equivalent to six decimal digits (i.e. (6 − 1) log₁₀(16) ≈ 6.02). A conversion of single precision hexadecimal float to decimal string would require at least 9 significant digits (i.e. 6 log₁₀(16) + 1 ≈ 8.22) in order to convert back to the same hexadecimal float value.

The double-precision HFP format (called "long" by IBM) is the same as the "short" format except that the fraction field is wider and the double-precision number is stored in a double word (8 bytes):

1		7			56		(width in bits)
S	Exp			Fraction
63	62	...	56	55	...	0	(bit index)*
* IBM documentation numbers the bits from left to right, so that the most significant bit is designated as bit number 0.

The exponent for this format covers only about a quarter of the range as the corresponding IEEE binary format.

14 hexadecimal digits of precision is roughly equivalent to 17 decimal digits. A conversion of double precision hexadecimal float to decimal string would require at least 18 significant digits in order to convert back to the same hexadecimal float value.

Called extended-precision by IBM, a quadruple-precision HFP format was added to the System/370 series and was available on some S/360 models (S/360-85, -195, and others by special request or simulated by OS software). The extended-precision fraction field is wider, and the extended-precision number is stored as two double words (16 bytes):

High-order part
1		7			56		(width in bits)
S	Exp			Fraction (high-order 14 digits)
127	126	...	120	119	...	64	(bit index)*
Low-order part
8					56		(width in bits)
Unused				Fraction (low-order 14 digits)
63	...		56	55	...	0	(bit index)*
* IBM documentation numbers the bits from left to right, so that the most significant bit is designated as bit number 0.

28 hexadecimal digits of precision is roughly equivalent to 32 decimal digits. A conversion of extended precision HFP to decimal string would require at least 35 significant digits in order to convert back to the same HFP value. The stored exponent in the low-order part is 14 less than the high-order part, unless this would be less than zero.

Available arithmetic operations are add and subtract, both normalized and unnormalized, and compare. Prenormalization is done based on the exponent difference. Multiply and divide prenormalize unnormalized values, and truncate the result after one guard digit. There is a halve operation to simplify dividing by two. Starting in ESA/390, there is a square root operation. All operations have one hexadecimal guard digit to avoid precision loss. Most arithmetic operations truncate like simple pocket calculators. Therefore, 1 − 16⁻⁸ = 1. In this case, the result is rounded away from zero.[7]

Starting with the S/390 G5 in 1998,[8] IBM mainframes have also included IEEE binary floating-point units which conform to the IEEE 754 Standard for Floating-Point Arithmetic. IEEE decimal floating-point was added to IBM System z9 GA2[9] in 2007 using millicode[10] and in 2008 to the IBM System z10 in hardware.[11]

Modern IBM mainframes support three floating-point radices with 3 hexadecimal (HFP) formats, 3 binary (BFP) formats, and 3 decimal (DFP) formats. There are two floating-point units per core; one supporting HFP and BFP, and one supporting DFP; there is one register file, FPRs, which holds all 3 formats. Starting with the z13 in 2015, processors have added a vector facility that includes 32 vector registers, each 128 bits wide; a vector register can contain two 64-bit or four 32-bit floating-point numbers.[12] The traditional 16 floating-point registers are overlaid on the new vector registers so some data can be manipulated with traditional floating-point instructions or with the newer vector instructions.

The IBM HFP format is used in:

• SAS 5 Transport files (.XPT) as required by the Food and Drug Administration (FDA) for New Drug Application (NDA) study submissions,[13]
• GRIB (GRIdded Binary) data files to exchange the output of weather prediction models (IEEE single-precision floating-point format in current version),
• GDS II (Graphic Database System II) format files (OASIS is the replacement), and
• SEG Y (Society of Exploration Geophysicists Y) format files (IEEE single-precision floating-point was added to the format in 2002).[14]

As IBM is the only remaining provider of hardware using the HFP format, and as the only IBM machines that support that format are their mainframes, few file formats require it. One exception is the SAS 5 Transport file format, which the FDA requires; in that format, "All floating-point numbers in the file are stored using the IBM mainframe representation. [...] Most platforms use the IEEE representation for floating-point numbers. [...] To assist you in reading and/or writing transport files, we are providing routines to convert from IEEE representation (either big endian or little endian) to transport representation and back again."[13] Code for IBM's format is also available under LGPLv2.1.[15]

• IBM System/360 and successors
• RCA Spectra 70
• English Electric System 4
• GEC 4000 series minicomputers
• Interdata 16-bit and 32-bit computers
• SDS Sigma series
• Data General minicomputers
• ICL 2900 Series computers
• Siemens 7.700 and 7.500 series mainframes and successors

The article "Architecture of the IBM System/360" explains the choice as being because "the frequency of pre-shift, overflow, and precision-loss post-shift on floating-point addition are substantially reduced by this choice."[16] This allowed higher performance for the large System/360 models, and reduced cost for the small ones. The authors were aware of the potential for precision loss, but assumed that this would not be significant for 64-bit floating-point variables. Unfortunately, the designers seem not to have been aware of Benford's Law which means that a large proportion of numbers will suffer reduced precision.

The book "Computer Architecture" by two of the System/360 architects quotes Sweeney's study of 1958-65 which showed that using a base greater than 2 greatly reduced the number of shifts required for alignment and normalisation, in particular the number of different shifts needed. They used a larger base to make the implementations run faster, and the choice of base 16 was natural given 8-bit bytes. The intention was that 32-bit floats would only be used for calculations that would not propagate rounding errors, and 64-bit double precision would be used for all scientific and engineering calculations. The initial implementation of double precision lacked a guard digit to allow proper rounding, but this was changed soon after the first customer deliveries.[17]

• IEEE 754 Standard for Floating-Point Arithmetic
• Microsoft Binary Format

• Sweeney, D. W. (1965). "An analysis of floating-point addition". IBM Systems Journal. 4 (1): 31–42. doi:10.1147/sj.41.0031.
• Tomayko, J. (Summer 1995). "System 360 Floating-Point Problems". IEEE Annals of the History of Computing. 17 (2): 62–63. doi:10.1109/MAHC.1995.10006. ISSN 1058-6180.
• Harding, L. J. (1966), "Idiosyncrasies of System/360 Floating-Point", Proceedings of SHARE 27, August 8–12, 1966, Presented at SHARE XXVII, Toronto, Canada
• Harding, L. J. (1966), "Modifications of System/360 Floating Point", SHARE Secretary Distribution, pp. 11–27, SSD 157, C4470
• Anderson, Stanley F.; Earle, John G.; Goldschmidt, Robert Elliott; Powers, Don M. (January 1967). "The IBM System/360 Model 91: Floating-Point Execution Unit". IBM Journal of Research and Development. 11 (1): 34–53. doi:10.1147/rd.111.0034.
• Padegs, A. (1968). "Structural aspects of the System/360 Model 85, III: Extensions to floating-point architecture". IBM Systems Journal. 7 (1): 22–29. doi:10.1147/sj.71.0022.
• Schwarz, E. M.; Sigal, L.; McPherson, T. J. (July 1997). "CMOS floating-point unit for the S/390 Parallel Enterprise Server G4". IBM Journal of Research and Development. 41 (4.5): 475–488. doi:10.1147/rd.414.0475.