home .. forth .. misc mail list archive ..

Re: MISC-d Digest V97 #28

On Wed, 23 Jul 1997, Jeff Fox wrote:

Thank you for your iTV update, Jeff. Things have been pretty quiet lately.
> Dear MISC readers:
> [...]
> iTV is negociating with companies that can provide access to even
> smaller and more sophisticated fabrication processes that would
> provide more speed and access to large amounts of on chip memory.
> As people have noted here before the on chip memory is needed to
> make full use of the CPU and coprocessor speed on these chips.

xref core on die: Analog's ADSP-21061 (esp. 62 and 63). Their 
price/performance ratio looks especially interesting. 

Concerning Beowulfs: current Pentium clones can be had for few 10 $, last
time I looked MuP21 sold for roughly $40. Motherboards cost next to
nothing, and the rest of the components (RAM, HD, etc) are also needed for
F21 nodes. L3/L4 microkernel, a message-passing microkernal with Linux
personality fits in 12 kBytes (i.e. in K6's cache). Beowulf nodes don't
run X, it is just needed for visualization on one machine. There's lots 
of MPI scientific software, which can be ported to Beowulf by just one 
"make". This code requires good float performance, which F21 can't do.

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. (While Beowulfs 
offer one order of magnitude better price/performance ratio as sci 
supercomputers, a SuperDSP box can top that by another order of 
magnitude. However, beware of the Merced, which will be a C6x clone, and 
run at 0.5-1 GHz).

Eugene

> 
> >We have heard of the stories of internet packages that 
> >can archeive limited graphical service withing 640k, why not.  I would 
> >like to here Jeff's veiw on this, how much space does the ITV box code 
> >take up for what services?
> 
> Yes, we can run with 1Mx4 or 256kx4 DRAM chips.  With the smaller chips
> we have 256k 20 bit words, or 640k bytes of memory.  We could get by
> with less if we had to.  The biggest single requirement is video memory.
> Even in 384x482x15 NTSC composite some 60k words are needed by the
> video output coprocessor.
> 
> Our CPU code is very dense.  4os provides us with a multi-threaded
> Forth environment with a dynamic memory manager with garbage collection,
> GUI with support for the video coprocessor, support for the on
> chip keyboard and serial i/o coprocessors, video interrupt and real
> time clock support, flash memory support, and TCP/IP networking code.
> 
> I don't have permission to disclose too many details of our code, but 
> I like to point out that our GUI requires one K word of memory for code
> executed by the CPU.  The video coprocessor requires 60k words of
> video memory for the display, but the CPU code is only 1K.  I like to
> compare this to the executable code in other GUIs.
> 
> >> I am also looking forward to the F21. But notice how much has occured 
> >> interim, especially the advent of Beowulfs, and 1 Gbit ethernet. Mass 
> >> production makes even inefficient monster hardware ridiculously cheap, a 
> >> large handicap for MISC.
> 
> I think we are out of sync, but I like the lead in.
> 
> One of the changes I am pleased with is the fall in memory prices.  The
> cost of an F21 node is starting to look like it could be ridiculously
> cheap to me.  We know that if it is produced in large quantity F21 could 
> be really cheap.  Since it only requires memory to make a node the nodes
> can be really cheap.  My low end estimate for an F21 node is now about
> like buying a happy meal for a kid.  How many "ridiculously cheap" 
> Beowulf nodes can one buy for that kind of bucks?  Anyone care to speculate 
> on the minimal cost of a node with the fastest possible Pentium, motherboard
> chipset controllers, cache, DRAM, pcb, and multiple high speed ethernet
> controllers?
> 
> I read up on Beowulf computers on the web.  These are nothing more than
> workstation farms using Pentium PCs and multiple high speed ethernet
> connections.  It is very much like the idea behind F21.
>  
> I would not say that Beowulf computers are ridiculously cheap!  The spec
> says they should have the fastest Pentium available, 256k cache, 32M
> DRAM, and multiple high speed ethernet controllers per node.  So
> although we are talking about similar performance per node compared to
> an F21 system we are still talking about a HUGE price difference, like
> maybe 100 to 1!  When comparing two systems where one is 100x more
> expensive I would hardly call the more expensive one ridiculously cheap.
> 
> And of course there is the issue of software bloat.  It is bad enough
> to have to pay for all the extra memory to hold all the extra software
> needed to meet the Beowulf spec, but I think the real problem with
> bloat is the bizzare idea that virtual memory will solve the problem.
> We all know that we give up an order of magnitude in performace when
> we have too much stuff to keep on chip and must go to external memory.
> We also know that we give up another order of magnitude when we have
> too much stuff to keep in cache (or high speed memory) and must go to 
> slower memory.  We also know we give up several more orders of magnitude
> when we have too much stuff to fit into memory at once and must
> page it in and out of disk memory.  You know when you are forced to give
> up many orders of magnitude in performance because of bloat you really
> do notice it. 
> 
> Beowulf computers use Linux and X so the bloat factor may not be quite
> as bad as it is with some other OS and GUI, but I really don't think
> X will ever fit into 1k of memory.
> 
> Then there are other things like active messages on F21 vs conventional
> message passing on Beowulf to get the parallelism.  Layers and layers
> of software inserted below the application's parallelism will also 
> be a factor.  That high speed network connection gets wasted when
> you build a parallel system on top of conventional message passing
> on top of ethernet.
> 
> >Question? how hard (extra equipement) would it take to drive 
> >ethernet off the F21 chip. 
> 
> There are people looking into ethernet on MISC chips, but that is
> about all I can say about it.
> 
> This Saturday Skip Inskeep, John Rible, and Chuck Moore will be doing
> presentations at the SV FIG meeting.  Skip will talk about the tethered
> development environments that he has done for i21 at iTV.  John will
> talk about Rochester, and Chuck will talk about how he has added Forth
> to OKAD among other things.  You never know what else Chuck might cover.
> I like to hear his talks so I can find out what information he has
> released publicly so I know what I can safely say about work at iTV.
> I will try to attend so that I can post an unofficial transcript to
> the MISC list.
> 
> Jeff
> Jeff Fox 
> jfox@dnai.com    Ultra Technology Inc. 
> http://www.dnai.com/~jfox/ 
> 
> 
> 

______________________________________________________________________________
|mailto:ui22204@sunmail.lrz-muenchen.de |transhumanism >H, cryonics,         |
|mailto:Eugene.Leitl@uni-muenchen.de    |nanotechnology, etc. etc.           |
|mailto:c438@org.chemie.uni-muenchen.de |"deus ex machina, v.0.0.alpha"      |
|icbmto:N 48 10'07'' E 011 33'53''      |http://www.lrz-muenchen.de/~ui22204 |