trans. (T. Onoma) wrote:
> On Saturday 30 October 2004 12:00 am, Hal Fulton wrote:
> | It's not a matter of computer science, but of economics.
> |
> | Implementing those algorithms in silicon is expensive, esp. if we don't
> | have practice at it. Call it hundreds of millions, counting the research
> | that precedes it.
> |
> | Convincing people to buy an expensive chip that their software doesn't
> | support is a huge PR task, not quite on a par with convincing al Qaeda
> | to become Methodists.
> |
> | Rewriting the compilers and the runtime environments to support these
> | chips is another gargantuan task.
> |
> | And all of this to solve a problem that is very minor and is chiefly
> | bothersome to those who have had fewer than four semesters of computer
> | science.
> |
> | Given a choice between improved floating point and a mission to Mars,
> | I'd take the latter.
> 
> Ha ha! When the spacecraft crashes and burns b/c of an overlooked rounding 
> error, I'll accept your apology ;) 

I work in a place that actually sends missions to Mars, and I can assure 
you that floating-point arithmetic is a well-understood phenomenon here.

> Honestly, you're way overstating this and undervaluing the problem. Many 
> persons have wasted hour upon hour compensating for this. This whole thread 
> kept going b/c of an overlooked rounding error, that could have easily 
> propagated into live code and been a real headache for someone --what value 
> do you put on all those efforts?

There is nothing you or anyone else can do to make a finite number of 
bits behave like the real number line. Whatever you do to fix things up 
for the utterly unimportant class of rationals whose binary expansion 
repeats in less than 64 bits will come at the expense of some other 
characteristics that genuine experts have agreed is more important. 
Unless you're a world-class numerical analyst, your odds of improving on 
754 are smaller than roundoff error. Give it a rest.

Steve