home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.sys.next.misc
- Path: sparky!uunet!mcsun!Germany.EU.net!incom!orfeo!qb!vhs
- From: vhs@rhein-main.de (Volker Herminghaus-Shirai)
- Subject: Re: 100 Mips Intel NeXT.
- Message-ID: <1992Dec11.174053.943@qb.rhein-main.de>
- Sender: vhs@qb.rhein-main.de (Volker Herminghaus-Shirai)
- Reply-To: vhs@rhein-main.de
- References: <1992Dec10.090411.12142@ichips.intel.com>
- Date: Fri, 11 Dec 92 17:40:53 GMT
- Lines: 151
-
- In article <1992Dec10.090411.12142@ichips.intel.com> lschultz@pdx255 (Leonard Schultz)
- writes:
- >
- > In article <1992Dec8.205422.746@qb.rhein-main.de> vhs@rhein-main.de writes:
- >
- > >The 80x86 line is forever chained to an instruction set that blocks
- > >most possibilities to make use of modern technologies other than
- > >manufacturing processes. Going superscalar is a pain with CISCs, the
- > >same holds for superpipelined. Modern compilers work best with a lot of
- > >general purpose registers - the 80x86es have three or so, the rest has
- > >special functions.
- >
- > Why is it a pain to go superscalar or superpipelined with CISCs, and
- > why is it easier with RISCs?
-
- The decoding takes too long with CISCs. This leads to late recognition of
- interdependencies between instructions in the pipeline. In order to
- prevent resulting trouble, you must either put much more silicon into the
- pipeline (a pain) or accept a general pipeline slowdown (also a pain).
- Also, especially with 80x86 code you have a whole bunch of self-modifying
- code running around mugging instructions that may already be in the
- pipeline. Therefore, additional silicon is needed again to make sure the
- code runs as expected (another pain).
-
- > And what modern technologies are x86
- > machines barred from using? I can't think of any. From the diagram
- > in EE Times, the Pentium(tm) Processor looks superscalar to me.
-
- Reasonable register allocation technologies used in compilers, for instance.
- The Pentium is almost superscalar, since it does have separate instruction
- pipelines, but unlike superscalar processors it does not have separate ALUs.
- That means at one point in the execution each instruction uses the same
- piece of hardware, while in other SS processors the instruction streams run
- fully in parallel. And I bet going superscalar *was* a pain.
-
- > >If you compare benchmarks using the same compiler (gcc) on e.g. a
- > >SPARC and an 80x86 you'll see that the SPARC code gets about twice
- > >as fast when you turn optimization on, while the 80x86 only gets a ~5%
- > >performance fart (Tested on Dhrystone 2.1).
- >
- > What the bleep does this prove? That unoptimized x86 compilers are
- > more efficient than unoptimized sparc compilers? Who cares? Compare
- > SPECmarks on optimized code if you want to make a comparison for gcc.
-
- It the bleep proves that it's damn hard to optimize 80x86 code. You
- won't get much of a performance gain. As I said we used identical
- compilers on both machines. Thus the unoptimized outputs stem from
- identical parse trees.
- If you compare based on standard benchmarks, don't compare even
- SPECmarks. At least let the manufacturer hand you SPECint92 and
- SPECf92p values separately (SPECmarks was 1989) and compare each
- single benchmark out of the complete suite to find out where the
- system performs well.
-
- > >Intel has always been lagging around a year behind its competitors
- > >in performance. However, they have learned to announce very early
- > >and deliver bad versions of the chips early so they sometimes appear
- > >up-to-date. But if you compare the real-world availability dates of
- > >intel 80x86 chips with their RISC competitors, they're almost always
- > >a generation behind.
- > >
- >
- > What product was announced early, in your opinion?
-
- 386, each stage of the 486 (25, 33, 50, 66) except SX, and
- Pentium, and this is not only my opinion.
- In addition the early 386s were buggy in 32-bit multiply
- (M$ even rewrote their C library to enable it to be used
- on these faulty processors).
- The early 486s were buggy in the FP unit and somewhere else I don't
- remember. The 486/50 took an additional 6 months after they were
- finally delivered to get rid of a heat problem.
-
- > >I would prefer NeXT to go with PA-RISC or, better yet, alpha (expensive
- > >though). Best compromise might be SPARC: High availability, multiple
- > >vendors, low price (much lower than 80x86es!) higher performance,
- > >pretty much iron out there alread, with much more to come.
- >
- > Ah, from a sexy/horsepower point of view, we would all like to see this.
- > Wouldn't we? But let's look at the realities of this, which NeXT must
- > face. PA-RISC is a proprietary processor. So if NeXT designs a system,
- > where are they going to buy the chips from?
-
- Where is NeXT going to buy the Pentium if they design a Pentium-based one?
-
- > And what about Alpha?
- > Unproven, expensive, not a mass-production levels, DEC unproven as a
- > CPU supplier.
-
- They have been delivering 21064s (alphas) for months. At 150 MHz.
- What are you afraid of DEC as a chip supplier (other than to cut into
- intel's semi-monopoly).
-
- > With all of NeXT's experiences with bad decisions (C-cube,
- > DSP, optical), why would they bet the farm on another?
-
- Whew! So you think the DSP was a bad decision? Maybe because it's from
- Motorola? What's wrong with having a DSP?
-
- > SPARC? I guess it will have alot to do with what markets NeXT is
- > aiming for, and how many different architectures they want to support.
- > Remember, SPARC _systems_ are typically more expensive than x86 systems.
-
- They are also a lot better equipped (i.e. the equivalent of a:
- - local bus 32-bit SCSI
- - local bus accelerated 32-bit graphics card
- - local bus 32-bit ethernet board
- - dual local bus 32-bit serial interfaces
- - local bus 32-bit disk controller, sound input and
- most importantly, a
- - local bus 32-bit wide keyboard interface!!! ;-)
- - a decent boot prom that lets you fix hard disks etc. even
- if you can't boot the O/S any more.
-
- > And there are NO significant performace differences.
-
- Whoooaaarr! That was a good one. Ever had a look at an SS-10? Or, even
- better from price-performance, an LX? And those are original Sun
- products, expect much lower prices from the clone makers soon.
-
- > From NeXT's perspective, the best route for the near future is Intel.
- > If TTM (Time To Market) is important to NeXT (and I'm sure most of you
- > want your next NeXT workstation soon).
-
- The Pentium is again delayed, this time until end 1993. Talk about time
- to market
-
- > They already have the OS
- > running on x86 machines. Why port to a different architecture even if
- > significant performance gains are there?
-
- Because of these performance gains (among a *lot* of other reasons).
-
- > As for the long term, NeXT has NDAs with many vendors. They will make
-
- Thus endeth your posting...?
-
- > --
- >
- > Len Schultz, lschultz@ichips.intel.com
- > Intel Corp., M/S JF1-19, 5200 NE Elam Young Pkwy,
- > Hillsboro, Oregon 97124-6497
-
-
- Volker
-
- --
- Volker Herminghaus-Shirai (vhs@qb.rhein-main.de)
-
- Looks good on the outside, but -
- intel inside
-