home *** CD-ROM | disk | FTP | other *** search
- Path: comma.rhein.de!serpens!not-for-mail
- From: mlelstv@serpens.rhein.de (Michael van Elst)
- Newsgroups: comp.sys.amiga.hardware
- Subject: Re: 80 MIPS with 1260 Explaination
- Date: 19 Feb 1996 12:28:29 +0100
- Organization: dis-
- Message-ID: <4g9mst$79l@serpens.rhein.de>
- References: <kluebke.14.824124623@mi.uni-koeln.de> <734.6617T840T985@Th0r.foo.bar> <4g9hjk$ah6@percy.cs.bham.ac.uk>
- NNTP-Posting-Host: serpens.rhein.de
-
- R.D.Martin-CSSE94@cs.bham.ac.uk (Rich Martin) writes:
-
- >If any one of these instructions relies on the result of an instruction further
- >on in the pipeline, then the whole system has to be stalled while the
-
- Just the pipeline that depends on results to be computed has to be stalled.
-
- >The Susperscalar architecture also compounds this problem. By having multiple
- >copies of the same CPU in the same chip
-
- It doesn't have multiple copies of the same CPU in the same chip. It does
- have replicated processing units though.
-
- >pipeline. If the 68060 had four copies (I don`t know exactly how many copies
- >it actually has) of the CPU
-
- It has two integer and one floating point pipeline.
-
- >another instruction in the pipeline, causing ALL copies of the CPU to be
- >stalled while the results are obtained from instructions further on.
-
- I wonder where you got this nonsense from.
-
- >Loops also cause a problem as the processing of the conditions of these loops
- >happens about half way through the pipeline, therefore when the CPU finds that
- >it has to branch to a different address, it has already loaded and started
- >execution of another 8 instructions, therefore effectively wasting these
- >instructions which have to be dumped.
-
- Dumping these instructions would be the naiv approach. Some CPUs with deep
- pipelines do this. But more recent CPUs (including the 060) can decode and execute
- branches early. The 68060 has a branch prediction cache that enables it to
- "guess" the program flow correctly most of the time. The effect that you describe
- is therefore pretty rare.
-
- >explains why somtimes turning the superscalar off can increase performance
- >(less pipelines stalls due to there being less instructions being executed
- >simultaniously)
-
- This is again nonsense. Pipeline stalls happen with one or two pipelines. Going
- to a single pipeline may reduce the number of stalls but that just says that
- two pipelines _may_ be as slow as a single one. However, most often they are
- faster. Two pipelines are never slower than a single pipeline (unless the
- CPU designer goofed).
-
- >It is perfectly possible to get the 68060 to execute 82.5 MIPS. The secret is
- >to compile a load of rubbish.
-
- Or average programs. If the 68060 could execute two instructions per cycle all
- the time you would see 100 MIPS (at 50MHz).
-
- > for single thread sequential instruction execution. New versions of
- > compilers are needed which reorganise instructions to allow these new
- > breed of CPU`s to run at full throttle.
-
- This is of course true but you are exaggerating the effects completely.
-
- --
- Michael van Elst
-
- Internet: mlelstv@serpens.rhein.de
- "A potential Snark may lurk in every tree."
-