[Beowulf] Nehalem and Shanghai code performance for our rzf example
landman at scalableinformatics.com
Mon Jan 19 14:18:19 PST 2009
Lux, James P wrote:
> On 1/19/09 10:07 AM, "Joe Landman" <landman at scalableinformatics.com> wrote:
> John Hearns wrote:
> > BTW, re the discussion on processor frequency scaling,
> > what finally did happen to Emitter Coupled Logic and gallium
> GaAs ... the material of the future. And always will be.
> > As a cub high energy physicist, I devoted many hours to learning about
> > ECL and Fastbus - it was (still is) the fastest switching technology,
> > therefore the Fastbus standard was based around it. Ditched now I
> > for VMEbus systems.
> > Sigh. Those big old crates were a reaql sight.
> > I wonder out loud if any exotic semiconductor plus ECL will be revived
> > in the future.
> The problems with GaAs are economic, and perceived safety (not to
> mention the technological problem of controlling defects in the material
> ... EL2 was/is a fun one). Economic, as Silicon has a huge installed
> base, and you can't simply switch materials in the same foundery ...
> different materials do require different processes and machines ...
> which increases costs. There are no economies of scale for GaAs.
> Perceived safety as the high temperature grown GaAs required 5+
> atmospheres of Arsenic gas at ~500C or thereabouts (quoting from memory)
> to maintain an equilibrated growth. Not too many people want to live
> near such a thing. There is a low temperature grown GaAs, though this
> has a number of defects you have to get better control over for devices.
> As if conventional silicon processes don’t use delightful gases like
> phosphine, arsine, and NF3, not to mention HF as an etchant? I doubt
> that it’s really a safety issue. If people actually knew how much
... which is why I used the word "perceived" ... :(
"There are chemicals in them thar water ... Hydrogen Hydroxide ... a
nasty solvent. Could cause cancer. Dissolves many things."
These days, perception == reality :(
> deadly stuff there is in the typical foundry/fab, there would be picket
> lines down the block outside.
> It’s all about yield. A) you can’t grow as big GaAs as Si crystals, so
As I remember, this is a technological problem. A solvable one (given
enough time/money ... etc.) See the economic argument as to why this
> wafer size is smaller. B) defects are more common.. Less of an issue for
Hmmm... I would say ... less well controlled. It took a long time to
understand how to control and therefore engineer Si crystals with
various properties. Ingots didn't start automagically growing in the
lab. GaAs has well known defects. Well, ok, EL2 took a long while to
understand (has it been finally put to rest?). It doesn't suffer from
some of Si's problems. But it has its own set.
> small die, where you can just throw away the ones with bad spots; tough
> for big dice, where the odds of getting a defect in every die on the
> wafer gets big
Yup, but again, this is a solvable problem given enough economic incentive.
> And, GaAs is not what you’d call a low power technology. No equivalent
> of CMOS for GaAs, it’s all bipolar, and in the case of GaAs ECL, non
> saturated, so it draws a lot of power. A saving grace is that you can
> run the junctions a lot hotter with GaAs, but even so, getting the heat
> out of the die is a challenge.
The other issue with GaAs (and lots of GaX and GaXY) are threading
defects. Took a long time to stabilize blue LEDs, originally Ga based.
The junctions got hot enough that it would start to drive defect
diffusion and threading dislocations.
> SiGe is probably more promising for real high speed stuff: you can
> integrate SiCMOS stuff with the SiGe on the same die.
> For real speed and power GaN is interesting.
... again it all comes down to a pure CBA. I wouldn't bet against Si.
(who bet against Si when he wrote his thesis ... on modeling defects in
GaAs some ... er ... years ago).
Joseph Landman, Ph.D
Founder and CEO
Scalable Informatics LLC,
email: landman at scalableinformatics.com
web : http://www.scalableinformatics.com
phone: +1 734 786 8423 x121
fax : +1 866 888 3112
cell : +1 734 612 4615
More information about the Beowulf