Re: Long Rant

[Eugen Leitl <ui22204@xxxxxxxxxxxxxxxxxxxxxxx>]

On Fri, 4 Aug 1995, Mike Losh wrote:

[ what MISC is good for discussion remark ]
 
> Jeff Fox said:
> >Once again, F21 is not designed for the conventional workstation/academic

Conventional, yes. Yet I do not think F21 to be unsuitable for
scientific computation and intend to proove it by using it in
my project.

> >market and SPECmarks don't make any sense in any way as a metric to
> >make any judgements about F21.
> 
> Excuse any ignorance on my part about the SPEC benchmarks, but is a port 
> of the 
> SPECint integer benchmark possible for F21?  Or is a C compiler MANDATORY 
> for 
> the benchmark to be considered a true "SPEC" benchmark?  It seems to me 

No it is not mandatory, though purists might object. The problem is:
we'd had to emulate the FPU the F21 haven't got by longish math routines
we'd had to write and optimize first. Even then, one can't actually
beat dedicated hardware (number cruncher logics) by software. We
can add up each node's contribution arithmetically, of course, though
I don't know how price/performance would relate e.g. to AMD DX4 or
i486 DX2: they are simply too cheap thanks to mass production.

> that one
> could port the source into hand-optimized F21 machine code which performs 

I have to assure you that writing correct maths code at good performance
is not simple. I wouldn't wish to debug transc. numerical code, ack!

> the 
> same or equivalent computations.  If SPECint requires wide (64 bit?)  

Most floats are wider, though some are 64 bits.

> computations, F21 may not be extremely efficient, having to string 20 bit 
> ops 
> together, but so what?  At least then we could say "F21 provides a SPECint 
> of XX
> out of SRAM, YY out of DRAM, at cost $$."  The Great Masses could start 

Those value would be identical, if code is in SRAM. Now code run time is
the bottleck.

> thinking, "Hmm, I can get 1/3 of the SPECInt of an Alpha workstation for 
> 1/10th 
> the cost..." (or whatever the numbers are going to be).  Hopefully, SPECint 
> running on a cluster of 8 (or whatever) of F21s will be really impressive 
> for 
> the price.

I think you are terribly overoptimistic here. This guys will be seeing
bad float values and... Hold it, you just said SPECint? Int values won't
bee too bad, unless: you use 32 or even 64 bit ints. You use * /. Alas,
unless I am mistaken, this benchmark has lots of these. Dedicating
more resources to the ALU and pushing bus width will help a lot. I don't
think one should put a fast mul/div in it, though this would allow using
them as DSPs (one guy grew rapidly disinterested in MISC once he learned
there aren't fast mult/divs in it). I think a fast multiplier will
take much more transistors than the F21, am I right, Jeff? I don't need
* /, though.
 
> SPEC benchmarks may not be interesting to most MISC readers (who envision 
> 20 bit
> apps), but I think most workstation buyers/users expect to see them.  If 

You bet.

> F21 and
> later chips are going to be a commercial success in the scientific and 
> engineering market, we must convince many of these people to take a look.  

The main problem of MISC: it has to create it's own market, and
there isn't money for marketing available. If would be a good embed
controller if it had on-chip memory. It has too much power for most
embed controller applications and additional chips costs do not
compare favourably with other ec you can buy now from major players.
Scientific/engineering market is too small/conservative to be of
any value and fun markets is big bucks/major players and the cake
is already cut (and partly eaten), already. Anything I forgot?

[ snip ]

> more specialized niches like video (although that could be a big market by 
> itself: set top boxes, home web-crawlers, etc.).

It _is_ a big market. If Jeff were not UltraTech but NEC or Nintendo and
had his product out by now, there wouldn't be any problems. If he can 
persuade big players into buying the technology, we will see F21s cropping
up everywhere. Taos is being evaluated as a par set top box OS to process
video streams by some major Nippon company. Alas, TaosVM won't map well to 
F21 stack engine, though it couln't hurt to port it (6 man-months).

> Of course, the best way to convince someone that scientific and engineering 
> work
> is possible with MISC is to show them someone who is already doing such 
> work.  

Even if they would see it happen, it won't sell if they have to learn
a new language/enviroment and switch to parallel programming first.
No way.

> I'm sure that many of the old mainframe diehards expecting to see their 
> benchmarks running on micros where later converted when they saw powerful 
> mainframe-type applications RUNNING on micros.  
> 
> May those who have proposed MISC applications have a great success and 
> pull the 
> others (kicking and screeming) along!

Not bloody likely, though I'd wish I see it happen.

> --
> Michael A. Losh                 Standard disclaimers apply.
>