Messages in this thread | | | Date | Thu, 15 Apr 1999 23:39:50 +0200 | From | Kurt Garloff <> | Subject | Re: SMP vs. Specialized chipsets WAS: Re: I have information/driver ... |
| |
On Thu, Apr 15, 1999 at 02:04:11PM -0700, bwoodard@cisco.com wrote: > > The ix86 insn set is hard to support with a simple design. Actually, what > > K6, PII and PIII do is to translate those insns in internal RISC like insns > > which can be nicely handled by the execution units. > > > > If you want to get rid of that much silicone, you have to dwitch > > architecture and use PPC, ARM, Alpha or whatever ... > > > > On the other hand, the x86 code is very dense, resulting in relatively short > > executables. > > > > One more question then. Since ix86 instruction set is so easy to > support what takes up all the die space?
Sorry? Did you read what I wrote? I said: The ix86 insns set is hard to support. You got a lot of instructions with a lot of modes and different lengths with few registers. That's an old CISC design.
> Also I still don't understand why chip designers don't design simpler > chips that behave like multiple processors rather than going to all > the work of having deep pipelines and all the complexity associated > with that e.g. speculative execution, out of order execution... > i.e. Is there some reason they don't make a chip which is essentially > four 486's, cache and all the glue circuitry needed to make them > capable of running SMP.
Well. Most OSes out there don't support SMP. Also, there are apps, which don't profit much from multiple processors. Also note, that if 10% of your app is not parallelizable, you will at most get 10 times the performance of a single CPU by using multiple. Not counted communication overhead.
Maybe processor designers find it more exciting to have those nighty features like specul. execution etc. One goal is to build a single very performing CPU. Consider the AXP-21264, which I consider the most performant affordable chip out there. The AXp is a RISC design, no question. But with the 21264, a lot of extra logic, such as spec. execution, more than one insns per cycle, etc. was added. So, the 21264-500 outperforms the 21164-633 by a factor of 1.5 to 2.0 (!)
If you need plenty of performance, you will get several of those.
Regards,
-- Dipl.Phys. Kurt Garloff <kurt@garloff.de> [Wuppertal, FRG] Plasma physics, high perf. computing [Linux-ix86,-axp, DUX] PGP key: see mailheader / key servers [Linux SCSI driver: DC390] [unhandled content-type:application/pgp-signature] | |