The reason there’s no report on the GP-GPU market (and
lordy, do we need a better name) is because it’s impossible to measure. Nvidia
and AMD don’t know who is buying chips for supercomputers, there’s no one (yet)
building “systems” except the labs, and the software that could be used on them
like Streaming on Cuda aren’t selling very well. Then you have the
chest-beating by AMD and Nvidia using IDC’s old HPC market numbers, which were
already ridiculously inflated and exaggerated, saying it’s a $20 billion
business. Well, if you count everything including the gas of the guy who brings
coffee to the folks building the house where the machines will reside you might
get to that number. As for the GP-GPU market, what part of it do you measure?
GPU sales? I can estimate that pretty closely, and guess what, less than
100,000 chips. With a blended ASP (includes the memory) of $250, you get $250
million. If you take an average of 100 chips per system, for 1,000 systems, and
add boxes (PSU, OS, and a display) you can add another $2 million. And then if
you add COTS SW, maybe we can pile on another $10 million. Are we rich yet?
The real cost, which will never get measured or reported, is
the slave labor of grad students and underpaid professors, and scientists in
national labs who write the real code and labor over the results.
So Dr. Pixels says there ain’t much there there in view of
the markets these giant companies are dealing in now, or, as our hedge fund
clients say, “Doesn’t move the needle.”
This would make a nice business for a startup, and if a
couple of companies could get serious about it we could actually create a
category and then we could measure it, and then those companies could interest
investors (assuming they can make money at it), and we could sell reports and
everyone would be happy.
And it’s not just GPUs from AMD and Nvidia—Mer-cury
Computing is offering GPU and Cell-based stream processors as well. They’re the
best candidate to make this into a category right now because they’re offering
more of a complete system solution.
However, we wondered at Siggraph last year if one couldn’t
take a bunch of Nvidia’s QuadraPlexes and build a dandy little supercomputer.
Yeah, a lot of folks are thinking that this is the year that
GPU Computing—Nvidia’s name for it—will do somewhere around $2 to $5 million.
Next year, we’re told we should see three to four times that. 2009 gets us to
$100 million. This is the whole market, with HW and SW and services. Double
precision and compilers are the hockey stick triggers.
Nvidia is running GPU Computing universities worldwide and
we hear they’re well attended. What’s driving this is the applications. The
users don’t care what hardware they have, as long as their applications run.
IBM’s seeding the hell out of Cell, but finding things slow.
So will dropping the cost per GFLOPS by a factor of 100
change the market? What do you think? If you can buy all of the computes you
can use, does it change what you do? 
 |
|
A |
*Colossus, The
Farbin Project: An apocalyptic sci-fi
movie based on Dennis Feltham Jones’ novel about a massive defense computer
that becomes sentient and takes control of the world.
It discovers a similar Soviet
project called “Guardian” that controls Soviet nuclear weapons. The computers
order a link and start exchanging messages, the communications become
increasingly complex, moving into mathematics unknown to mankind and finally
into a binary language scientists can’t interpret.
Eventually, Colossus
broadcasts to all countries, declaring itself the ruler of the world and saying
war will be abolished and problems such as famine, disease, and overpopulation
will be solved. “Obey me and live, or disobey and die,” spoke Colossus. And to
underline the point Colossus threatens to blow up nuclear warheads in the U.S.
desert and the Ukraine.