[Beowulf] First 96-Node Transmeta Desktop Cluster Ships

[Jim Lux]

At 10:08 AM 5/4/2005, Vincent Diepeveen wrote:
>Yes it is very interesting.
>
>However their sales price of $100k they initially quoted i found a bit
>expensive.

I suspect that they picked the price point according to typical "purchasing 
authority" thresholds.




>I don't want to be rude, but these processors are quite slow if you just
>look to 1 processor and they go back in performance after a while when they
>go hot.
>
>So a fully loaded system will have a far smaller performance in the long
>run than it has in the short run.

But then, this isn't a device designed to be churning 24/7 serving multiple 
users.  It's for someone like me to shove under my desk and use in a very 
"bursty" fashion.  Much like the processor utilization on my desktop 
computer now, where it runs around 5% most of the time, and bursts to 90% 
when I fire up Matlab and start crunching.  But then, I go to lunch, and it 
drops back down.



>Of course that has its charm too, a system that's fast if you just use it
>real short at a "critical" moment.
>
>So the quoted performance is still impressive then.
>
>The network is not so impressive what each node is connected to another
>node. That is not relevant for embarrassingly parallel software though.
>
>Relevant is the price per gflop IMHO.

Look at TOTAL Cost per USEFUL gflop.


>For a NEW product that tries to fight itself into the market, you must
>simply be factors cheaper than the competition. $100k i find a tad expensive.

Or, find a new and different market that has different requirements, not 
currently well addressed by the existing marketplace.  This product will 
NEVER compete against a rack of 4U chassis with a team of gradstudents 
installing and maintaining it for (virtually) free.  Nor will it compete 
against any "home built" system, where labor is free.  Nor will it compete 
against the 1024 processor dreadnought class clusters.


They'll not sell it on dollars/GFLOP.  They'll sell it on the basis of 
increased productivity of expensive people.  The structural analyst with a 
fully burdened cost of $250K/yr who becomes 15% more productive, for 
instance, will pay for it in the usual 3 year amortization 
interval.  Justifying that cost/benefit will be the real challenge, but 
it's no different (conceptually) than justifying a $5K wordprocessor for a 
$15K/yr typist. (1985 GS-3 Typist II.. probably after you add burden, 
around $30K/yr cost)
Here's a quote from a 1989 business magazine:"
When all is said and done, however, forget about features; forget about 
price. Instead, choose the program that you can get help with at midnight 
on a Sunday when you have a proposal due at eight o'clock on Monday morning.
"
Another report from that era mentions that factories capitalize their 
workers (in 1981) at $25K/worker (that is, they invested $25K in plant and 
machinery for that worker), while clerical staff was capitalized at $2.9K.



>So to speak you can buy 10 quad dual core opterons 1.8Ghz for that price.
>
>that's 80 processors x 3.6 gflop = 300 gflop too.
>
>So effectively 10 quad opterons are faster in floating point than this
>orion machine.
>
>It's true however that the 10 quads eat a bigger power bill.
>
>But if you can afford paying $100k that won't be a problem either.


Au contraire...  The problem with power isn't the electricity cost.  It's 
the removal of the heat and other infrastructure costs.  Those 10 quad dual 
core opterons will probably fill a rack and have a mighty roar of cooling 
fans, and I won't be able to plug it in in my office built back in the 
60's.  (assuming I have ear plugs)

Lest you say, well, just use it remotely, and park it in the machine room; 
I'd just point out that the same argument was made when PCs were 
introduced.  Why use a PC?  We'll give you a perfectly good VT-100 or ADM3, 
and we'll take care of the machine down the hall, and then, we can all 
share the massively bigger processor and achieve economies of scale.  We've 
even set up a very slick batch job utility and your printouts will be 
delivered to a box down the hall within an hour.  (There is old serial 
cabling protruding from the raceways behind my desk dating from that very 
era, a time of Gandalf "short haul" modems and current loop interfaces).

>Nevertheless if we consider the fact that this company just starts to make
>clustered systems and already so quickly puts such an impressive product in
>the market they sure have my compliments.
>
>IMHO they are the right type of company to make a clustered cell type
>processor system (if ibm wants to deliver them cpu's). Their system needs a
>fast processor that uses little power. The philosophy i really like.
>
>Please note it would perform horrible for my own software, as it is latency
>sensitive software, so anything i write down here is meant for software
>other than my searching software.

Perhaps.... they DO have the advantage of a very wide/fast bus internally 
among the processors, with short runs. Since they don't have to go "outside 
the box", they can control the interfaces: all those line 
drivers/receivers, and just connector costs add up.  Decent shielded 
connectors are several bucks apiece, and when you have to buy 100 of them, 
thats a significant fraction of the purchase price.  Not to mention power 
distribution, power supplies, and all those things.





James Lux, P.E.
Spacecraft Radio Frequency Subsystems Group
Flight Communications Systems Section
Jet Propulsion Laboratory, Mail Stop 161-213
4800 Oak Grove Drive
Pasadena CA 91109
tel: (818)354-2075
fax: (818)393-6875