Robert Myers wrote:

IBM did eventually develop a specialized processor for Cray

Are you sure? For Cray? Wasn't that DEC?

Mayan Moudgill wrote:
Robert Myers wrote:

IBM did eventually develop a specialized processor for Cray

Are you sure? For Cray? Wasn't that DEC?

Don't think so. The only thing I can remember IBM doing for DEC was to
fab the Alpha EV7 processor when Alpha made the move to SOI. Since
Alpha's coffin had already been securely nailed shut by that point there
wasn't much down-side for IBM and reportedly they needed the fab
business - but the design and at least most of the development was at
DEC (IBM probably did help with knowledge of how to use the process
well: my impression is that they were being magnanimous given the
circumstances).

IBM spent a lot, repeat lot, of money on steve chen and SSI. Not so
much for cray to my best recollection.

SSi died anyway.

Del Cecchi wrote:
And something about the computing with GPU thing that just crossed
my mind. I have decided that it is a kluge for those without the
resources to design a chip of their own. After all the existing
GPU designs are not optimized for use in large scale computing.
However something much like them certainly could be, (before I am
attacked).

Most supercomputers are built and for "those without the resources
to design a chip of their own".

See http://www.top500.org/stats/list/33/procfam
87.2% by count is x86. Not really optimized for HPC.

GPUs' role in supercomputing is like x86: which is the most cost
efficient mass market processor to put in a supercomputer?

--

I must admit that I am puzzled as to how and why IBM stays in
supercomputing. My guess is that the US Govt funds Blue*, and that
IBM management is smart enough to see the payback. Even though it
is hard to make it pencil out using conventional accounting rules.

Related, I am puzzled as to why IBM persists in building
supercomputers whose business relevance is quite remote. I guess,
in for a penny, in for a pound.

---

I remember fondly working on supercomputing at Gould. Even
occasionally at Intel, although it is yet to be seen if Intel will
ever get serious, or practical, about HPC. If they were, I might
have stayed.

The behavior of the US over the last ten years and looking at where we
are now indicates how poorly federal decisionmakers who panicked over
Japan, Inc. understood the business. Neither Japan nor the Federal
government can compete with the flow of money coming from consumers.

a slight topic drift in these recent posts mentioning "business
ethics" is an *oxymoran*
http://www.garlic.com/~lynn/2009o.html#47 U.S. begins inquiry of IBM in mainframe market
http://www.garlic.com/~lynn/2009o.html#52 Revisiting CHARACTER and BUSINESS ETHICS
http://www.garlic.com/~lynn/2009o.html#57 U.S. begins inquiry of IBM in mainframe market

Robert Myers <rbmyers... at (no spam) gmail.com> writes:
The behavior of the US over the last ten years and looking at where we
are now indicates how poorly federal decisionmakers who panicked over
Japan, Inc. understood the business.  Neither Japan nor the Federal
government can compete with the flow of money coming from consumers.

that showed up even more in the dept. of commerce meetings over hdtv
... it included apparently constantly fiddling the standard ... the
theory was if foreign company(s) won the hdtv market ... that would be
significantly larger higher performance chip funding than the measely
personal computer market.

early in hsdt projecthttp://www.garlic.com/~lynn/subnetwork.html#hsdt

"Andy \"Krazy\" Glew" <ag-news at (no spam) patten-glew.net> writes:

Joe Pfeiffer wrote:
Virtually all semiconductors of any stripe are made for those without
the resources to design a chip of their own -- but GPUs seem like a
pretty extreme special case of something heavily optimized for one
application that turns out to be useful someplace else.
Then I think that you have not been looking closely at GPUs recently.

You're right, I haven't -- sounds like I need to. Are there some good
references, or am I better off just googling?

So Tesla is used as a bag on the side of a convention system?

Memory is/was insufficient, but it was a start. The real problem is
software. And software needs memory because among other things, the
amount of data is growing because interests and needs (applications are
growing). Basic theory for simulation can usually be small, but
reading briefly what Paul and Joe posted is basically right on.

Eugene Miya wrote:
The AI types remain largely clueless to parallelism.

Does that explain, or is a consequence of, the fate of the 5th gen fiasco?

On Oct 30, 12:49=A0pm, eug... at (no spam) cse.ucsc.edu (Eugene Miya) wrote:
The people who thought highly of the CM-1 were the AI types who also
liked Symbolics 3600s. =A0I also saw and have pieces of the MPP (now at
the Smithsonian near Dulles). The AI types remain largely clueless to
parallelism. =A0They don't constitute a viable market.

The real ho-hum part wasn't the speed. The problem for real machines
with real problems was that they lacked sufficient amounts of real memory=
.

That's an interesting claim. It implies that all you have to do is to
build a knock-off of CM-1, give it sufficient memory, and people would
no longer regard Thinking Machines as the exercise in hubris that they
now do. I think of Thinking Machines as a major contributor to moving
computing out of Kendall Square forever.

I read Danny Hillis' PhD thesis after conversations with you. I don't
remember much of it because I'm pretty sure I missed the really,
really smart part of it. Mostly, it seemed really, really obvious.
Perhaps I'm just not really, really smart enough myself.

In any case, you have offered one of your as-always tantalizing
claims. Is there any substantiating evidence, or are you just
repeating what the really, really smart people were saying at the
time?

Probably, I'll do better just to go to AI seminars at MIT and see what
I can find out for myself. I'm sure many of the players are still
around.

I'm in a strange field (X-ray crystallography) where the problems have
stayed the same size, and moved from being big enough to require
really baroque coding to run on the biggest available computers, to
the scale where I can run sixteen independent protein refinements on
sixteen threads on a Mac Pro and get through a week's depositions
(that is, every protein structure published that week) inside 24
hours.

People around here (Bay Area, SF) certainly have those kinds of problems.
The problem is that for any compute node, more memory is requested of any
processing as time advances as curiosity drives thinking.  And some of
it doesn't pan out.  One still has to pay for the bills.