Re: MISC-d Digest V97 #28
On Thu, 24 Jul 1997, Eugene Leitl wrote:
> On Thu, 24 Jul 1997, Penio Penev wrote:
>
> > On Thu, 24 Jul 1997, Eugene Leitl wrote:
> > [...]
> > If we agrre, that www.TheChipMerchant.com sells ar rock bottom prices, the
> > Cyrix 686 150MHz (200+) is $83, which is is quite a few $10s.
>
> I just bought a 166 K5, which was quite a number of $10. 83$ is perfecly
> allright considering what an F21 is going to cost ($1? $10$ $40? I think
> the latter).
I think somewhere between the former two. More to the left side (in
QUANTITIES)
> > we are looking at: K6-166-200-233 @ $181/233/400 ; 686MX-166/200 @
> > $177/249 ;
>
> Let's see what M2/K6/Pentium are going to cost in 3 months.
Whatever the 686 was a few months ago -- $150 for a non-discontinued
product.
> Please notice
> that a) I can't still buy an F21 (I've been waiting for them for almost 2
> years now) b) the machine can run a lot of things the F21 can't do.
a) true. And Jeff is hoping to change that soon :-)
b) not a Beaowulf pack. You won't buy 100 workstations to run Excel.
> > > last
> > > time I looked MuP21 sold for roughly $40.
> >
> > This is to cover the development costs out of the sales of development
> > systems. Do inquire about the price if you want to buy, say 10K MuP21.
>
> But I don't want to buy 10 k. I'd like to buy let's say 64 F21's, _with_
> a board, and SRAM. Besides of the fact that I can't buy them, I don't
> think the prices will be that spectacular.
If you buy 64 Beaowulf nodes, you might as well buy 10K F21s (or at least
in that ball park).
> > > Motherboards cost next to
> > > nothing,
> >
> > Say, $150 form an MMX mb. Source -- necxdirect.necx.com:8002
>
> Who does need an MMX? Of what possible use is MMX in sci calculations,
> unless one (oops) intends to do lattice algorithms?
The one that bought an MMX processor (remember, there is no vanilla Pentium
any more). Anyway, the non-MMX ones are $110.
> > > and the rest of the components (RAM, HD, etc) are also needed for
> > > F21 nodes.
> >
> > I haven't heard of a Pentim running off 640K DRAM. I've heard of Pentiums
> > running off 1M SRAM, though. (I've heard of R10Ks running of 4M SRAM :-)
>
> I need the RAM.
You need a _total_ amount of RAM N MB. Not a per-node constant, because
that depends on the number of nodes. So, per node, the F21 will me much
cheaper, because it will be configured with less RAM.
> Wide RAM (64 bits).
10 F21s have a RAM bandwidth of 200 bits. And can saturate it, I might
add.
> Fast RAM, preferably PB SRAM.
What is the primary cache miss penalty for the Pentium? What about the
secondary? If you have 10--20 clocks latencies due to cache misses, what
good does it make to have memory with 1 versus 2 clocks latency?
> RAM
> costs the same for F21 (I hope) as for a PC, and RAM usage is roughly the
> same (code density plays no role in my application).
The total RAM for a given problem will be roughly the same for both. For
a fraction of the price of the RAM, you'll have 1 F21 per 640K RAM.
> > > L3/L4 microkernel, a message-passing microkernal with Linux
> > > personality fits in 12 kBytes (i.e. in K6's cache).
> >
> > What abut the application and the FP code? FP data?
>
> Lots of lookups, vast amounts of data (16-64 MBytes), tiny code. I will
> probably even use machine code, hideously ugly as Intel code is.
How about putting the code on-chip for F21? No PCI chip set can beat F21
for lookups.
> > May be, may be not. But certainly 10xF21 will be better than 1xAMD. And
> > about 10 time cheaper, I'd add.
>
> Let's see, I can get 100 F21's for one AMD (K5? K6? Whatever). Sold.
> Where can I get them, preferably without having to develop a host system
> from scratch (and also, not having to pay $50 for a meek piece of epoxy
> sans other semiconductors).
>
> > > Don't get me wrong, I still go for the F21. But it just doesn't compete
> > > with the Beowulf, a different type of a beast entirely.
> >
> > The _only_ advantage of the Beowulf is that you can say 'make' and produce
> > something running on them. The only.
>
> I don't see it quite like that.
You just contradicted that a few lines above -- "preferably ..."
> > In that regard the SuperDSP box is like the F21 -- porting is required.
>
> Yeah, but you can get a QNX and a C compiler -- so porting cannot be
> equaled to writing from scratch. (Fortunately, I don't need an OS for my
> app).
You can get gnu c and compile to an intermediate code. Then write a C
virtual machine for F21 and an F21/C front end for gcc -- that's trivial.
Your internal loops will be hand-coded anyway, if don't want to waste
money on hardware.
> > > However, beware of the Merced, which will be a C6x clone, and
> > > run at 0.5-1 GHz).
> >
> > Vaporware.
>
> Merced, yes. 'C6x, no. ADSP-2106x, no. F21, yes.
Merced is not coming out of the foundry in August, as far as I've heard.
The 'C6x is definitely a contender, but you'll need some pretty elaborate
tricks to squeeze the juice off it, and it is not good for table lookups.
And it costs $100 @ 10K
> Ask Analog for ADDS-2106X-EZ-LITE, what you get, and what does it cost.
> (I'm still waiting for a 'C6x EV board TI was musing about some months ago).
I believe TI is better than AMD.
> And never, never underestimate Intel.
As per Intel, do you know how much an embeddable 386 costs nowadays? 486?
To summarize, I believe that for special purpose supercomputing, the MISC
approach will be 10 times the performance/price of the general purpose
approach. I remains to be seen whether this is enough for somebody to
make money on that.
--
Penio Penev <Penev@pisa.Rockefeller.edu> 1-212-327-7423