This is a document copied from Intel's World Wide Web server
http://www.intel.com/...     on 5. December 1994 - 19.45 CET
____________________________________________________________


Floating Point Flaw in the Pentium(tm) Processor



Overview

Earlier this year, as a part of Intel's ongoing testing and product 
development work, a flaw was discovered in the floating point unit (FPU) 
of the Pentium processor. This flaw affects the accuracy of the floating 
point divide instruction for certain combinations of input operands  (i.e. 
certain combinations of specific numbers). The impact of the flaw is that 
for one in nine billion possible independent divides, the precision of the 
result is reduced. The flaw can occur in all three (single, double, and 
extended) precisions, and is independent of rounding mode. The flaw was 
detected during the course of over seven trillion divisions of testing 
that have been performed on the Pentium processor, and a detailed 
investigation has determined the cause. 

Description

For certain rare divisions (approximately one in nine billion divide 
pairs), the result of a floating point divide is returned with reduced 
precision. The flaw is due to a problem with the lookup table used by the 
iterative algorithm that performs the divide instruction. There is no 
simple characterization of the values for which this flaw occurs. However, 
Intel has run trillions of different random divides to validate that the 
occurrence is less than one in nine billion: the precision of the result 
ranges from the 4th to the 19th significant digit (or the 4th to the 19th 
place after the decimal point in numbers expressed in scientific notation, 
an example of which is 0.12345 x 105). Statistically, for these values, 
the 4th significant digit is the least likely position for an error. 
Importantly, intensive testing has shown that no known physical constant 
causes the reduced precision. 

Types of Applications Analyzed

The impact of the flaw varies by the rate of use of floating point 
instructions, the input data fed to them, the use of the results in 
further computations, and the needed accuracy of the application. Intel 
has assessed the likely impact of the flaw on three types of applications 
that might be run on a Pentium processor-based system. These include: (i) 
commercial PC applications on desktop and mobile platforms running 
MS-DOS*, Windows*, and OS/2*. This class includes basic spreadsheets and 
accounting applications and personal finance; (ii) technical and 
workstation applications, such as engineering and scientific, advanced 
multimedia, educational and financial applications (thus this class 
includes "power users" of spreadsheets such as financial analysts and 
financial engineers); and (ii) server and transaction processing 
applications. 

Commercial PC Applications

These applications were addressed by extensive internal validation and 
surveying external experts. A large majority of these applications do not 
invoke the floating point unit at all, much less floating point divides. 
This includes applications such as word processing, text editing and 
e-mail. In the commercial PC domain, the majority of applications that do 
use floating point do not invoke an appreciable number of different 
divides and hence do not introduce meaningful failures that will pose a 
data integrity problem during the useful life of the part. For 
applications such as corporate financial analysis and forecasting, 
marketing analysis, planning and so forth, the likelihood of encountering 
reduced precision divides is low. This is because typical calculations 
here are dominated by comparisons and additions. The input-output 
operations and the time for human conception of the results consume orders 
of magnitude more time than the processor spends performing arithmetic 
operations. This effect limits the number of divisions that are computed 
per day to well below what is necessary to have any appreciable 
probability of experiencing a meaningful inaccuracy. As an example, 
consider a large budget calculation implemented as a 700x700 cell 
spreadsheet, which is run an average of a few times a day. This will 
produce fewer than 10,000 divisions a day (on average): so few divisions 
that no reduced precision is likely to be seen for thousands of years. A 
more typical spreadsheet user doing 1000 divides per day will encounter 
this reduced precision once in every 27,000 years of use. 

Technical/Workstation Applications

The second class of applications studied were workstation applications, 
broken down further: scientific/engineering and financial. Scientific and 
engineering applications were studied extensively using leading vendors' 
analysis and simulation packages. These applications, such as structural 
mechanics, fluid mechanics, computational biology, chemistry, and 
mathematics, were studied to determine the numerical effects of the 
reduced precision. For structural mechanics -- among the worst problems 
for numerical convergence -- an error in the fourth significant digit 
would appear no more often than once every thousand years. Many other 
applications that iterate to convergence, such as circuit simulation, are 
even less susceptible to reduced precision due to the algorithms they 
employ. The average over a large range of applications was an expected 
meaningful error rate of less than one per thousand years. Particular 
technical applications employing unusually large numbers of divides and 
requiring extra precision need to be evaluated individually. 

Many workstation financial applications perform a significant number of 
divides, but the impact of the reduced precision is likely to be low, 
varying with the likelihood of its propagation. The typical financial user 
will not likely see a meaningful error in the usable life of the machine. 
In fact, the likelihood is less than one error per 270 years of use. 

However, as the most intensive floating point use is in this category and 
these are the applications which could possibly suffer meaningful error, 
users of these applications should contact Intel for more information. 

Server Applications

Database and file server applications typically do essentially no floating 
point computations and thus are unffected by the flaw. 

Conclusions

There is a flaw in the precision of the floating point divide operation of 
the Pentium processor. The impact of the flaw varies by the rate of use of 
floating point instructions, the input data fed to them, the use of the 
results in further computations, and the needed accuracy of the 
application. The flaw is not meaningful for the vast majority of 
commercial PC applications. The flaw is not likely to be meaningful over 
the usable life of the processor for the remainder of these commercial 
applications and most workstation applications. Some workstation 
applications may see occasional reduced precision results from the flaw. 
The flaw has no effect on server applications. 



Intel will work with users of applications involving intensive floating 
point calculations and resolve their concerns in the appropriate fashion, 
including, if necessary, replacing their chips. Please call Intel at 
1-800-628-8686 (international: 916-356-3551) if you have concerns or need 
more information relating to this issue. 

*Other brands and names are the property of their respective owners. 
___________________________________________________________________________

Last modified: 12/1/94

webmaster@www.intel.com 
Tell us what you think! 
Copyright (c) 1994, Intel Corporation. All rights reserved. 


