From: Doug Shapter <dps@kafka.atinc.com>
Date: Tue, 22 Nov 94 13:57:06 PST
To: Chris Wedgwood <cwedgwood@cybernet.co.nz>
Subject: Re: Pentium bug and CRYPTO
In-Reply-To: <m0r9uNz-000SgBC@mserve>
Message-ID: <Pine.3.89.9411221604.B13316-0100000@kafka.atinc.com>
MIME-Version: 1.0
Content-Type: text/plain




On Wed, 23 Nov 1994, Chris Wedgwood wrote:

> chris.claborne@sandiegoca.ncr.com writes:
> [Will the following error (Re
> [Pentium Floating Point Bug ... cause problems with PGP key generation or]
> [any other normal operations with PGP or other crypto.]
> 
> It shouldn't effect PGP in the slightest. Its a bug that effects only
> certain mantissa (23 are known so far) when doing a floating point divide
> (double precision).
> 
> PGP doesn't use floating point for its big-numbers and it has no need for
> double precision. Since most version of PGP compiled for Intel platforms
> will be or the MS-DOS or Windows variety then it is very unlikely that even
> floating point instructions will be used - emulations libraries will be used
> instead for floating point. NT is a slightly different matter - but as I
> said PGP doesn't use floating point for the key generation or ANY of the
> RSA/IDEA code....
> 
> An interesting point about this rather obscure bug though. It won't effect
> over 99% of all Pentium machines in use. It won't effect word, windows or
> any of the other numerous programs that hold a large market share and
> high-usage stats.... but people making a big deal out of this and demanding
> fixes (I have heard new Pentiums don't do this and am going to test this
> next week) could cost Intel millions potentially..... and I doubt whether it
> would effect PovRay or whatever things people might actually use floating
> point for anyways....
> 
> Serious scientific work could suffer severely, and since Intel boxes are good
> power for dollar there are quite a few used in various places for intensive
> calculations.... (e.g. seismic ray-tracing - but that done of 486-DX2-66
> machine because here is NZ they are about half the price of a Pentium so are
> even better value for money).
> 
> Chris
> 

Not much crypto relevance, but ...

We've been concerned about this bug for precisely that reason-- serious 
scientific work. We are contemplating purchasing a Pentium and running 
FreeBSD on it to do scientific computation and while Intel has "fixed" 
the fp problem, I wonder if there are others that have yet to be discovered.

(As to why the bug slipped out from under Intel's quality control, another 
programmer here pointed out that default fp precision is 6 for a 
printf call and that the error occurs in the 7th decimal place. 
Coincidence? Chance? Grist for the conspiracy theory mill?)

Granted the bug won't affect PGP much, but you have to wonder about the 
integrity of a company that lets this kind of hardware slip out the door. 

Doug Shapter                
dps@kafka.atinc.com         
finger dps@kryten.atinc.com for PGP public key