32 vs 64 bit

[Johan Nielsen]

I have a question where I know that the specific answer will always be "it depends." So I'm asking this question in a more generic sense. The question is, if an application suppports 64 bit is there generally a performance gain by implementing in 64 bit vs. 32 bit. Once again, in a generic sense, what if any are the reasons for this gain in performance (ability to address more memory, etc.)
If anyone has experience switching from 32 to 64 bit implementations of any application, please share any performance findings.

Thanks.
Johan

[Steven E. Protter]

When you address memory in bigger blocks, its usually more efficient than doing it in smaller blocks.

A 64 bit system can address more physical memory as well.

We found our Oracle performance(hp-ux) went up substantially when we went to the 64 bit version, though some of the jump was due to the migration to the faster processor and 64 bit HP-UX.

Still the database peformed better.

I've seen situations were going to 64 bit OS slowed things down. That was when lots of little chunks of memory needed to be addressed. In that circumstance, there is a lot of wasted memory. The OS reads a bigger chunk of memory for a piece of data that is actually very small.

Additionaly sometimes the 64 bit versions of programs that applications vendors put out were done sloppily, not tested thoroughly and simply ran like dogs.

SEP

[Stuart Browne]

The only sorts of area's I thought you'd see improovements (ignoring processor design apart form the bus width) would be with accessing large data packets (as SEP said), and for complex maths (using later-than-32bit values).

Wider bus, more data per clock cycle.

As for how this improoves on things, it really is on a application by application basis.

[dirk dierickx]

the main purpose of a 64bit cpu was never 'being faster'. some of those cpu's are indeed faster then a comparable 32bit cpu, but that is mostly due to architectural design changes in the cpu, like # registers, pipelines, etc...

[Ted M Johnson_1]

Access to a 64bit data packet won't speed processing up. Keeping the same number of registers (as well as the same prediction code and cache), only stuffing them with 2x the amount of data doesn't help performance. In fact, the best you could hope for by just increasing the width of the bus would be equal performance (assuming there is no architecture change).

It would still take the same number of clock cycles *or more* to process 64bits instead of 32bits.

Or, in other words: ALL of the performance gains SEP saw were due to switching processors. Going from a 32bit PA chip to a 64bit PA chip increases the clock speed, the cache, pipeline code, TLB size, etc. Just moving to 64bit code without changing anything else will slow your code down.