Pages

February 25, 2008

Teraflop to Petaflop GPGPUs and artificial neurons for artificial vision and smell

Nvidia introduced the Cuda development tool in June 2007 to allow scientists to tap into the GPU's power. More than 50,000 users have downloaded the software

By 2012, three of the top five supercomputers in the world will have graphics processors using parallel computing applications to crunch numbers at a clip that's not possible on standard CPU-only set-ups, predicts Nvidia chief scientist David Kirk.

Since other supercomputers are projected to be at several petaflops in speed then the GPGPU enhanced machines must be targeting 5-10 petaflops in performance by 2012.

Evolved machines is using GPGPUs for enhancing neural network and neuron simulation for visual systems and for artificial sense of smell

ATI's Radeon HD 3870 X2, released last month, can hit around 1TFLOPS (one trillion floating-point operations per second). Its 320 stream processors combine for massive parallel computation.

Nvidia could have tighter co-processor coupling between its GPGPUs and apple computers.



4 comments:

Anonymous said...

One thing I'm interested to see is how the market for supercomputers expands in the coming years. I remember reading maybe 3 years ago that the pharmacuetical industry can make use of supercomputers over 1 petaflop. And the more powerful they get from there, the more simulating they can do on the supercomputer.

You figure a company like Pfizer spends 7 billion a year on research, a 100 or 200 million dollar supercomputer really isn't that much in light of their total expenses. Especially if it replaces money currently being spent elsewhere.

And there are a number of pharma companies with annual research budgets in the 4+ billion range.

--aa2

bw said...

The price of a petaflop is coming down. It is $100-200 million now, but the GPGPU version of 2010-2012 might only be $200K-3 million.

In terms of the number of companies that need whatever the highest-end of computing power is one for each trillion of normal economy and for some parts of the economy like oil and pharma could need one for every 50-100 billion of economy. At least a $50 billion for enough research budget. although like China a government could boost the high end with a national supercomputer network and increasing demand and access.

More industries are becoming information base and driven which should also increase the supercomputer to economy ratio. I also foresee the new Dwave quantum computer being part of the computing arsenal of many companies. Giving billions or more boost using quantum annealing and computing over conventional computing.

bw said...

Looks like I had a lower estimate of the cost of the current systems.

Here is a good article that discusses cost and power (operating costs) for petaflop systems and using FPGAs for semi-custom systems that would have lower cost and lower energy usage and high performance.

http://www.hpcwire.com/hpc/2112632.html

Anonymous said...

Thats a good point that the price of a petaflop will be declining each year, maybe 40%. If they decided they were going to spend 200 million a year on supercomputers they'd be able to buy more petaflop supercomputers by 2012, or buy a more powerful system.

I'd like to see the pharma companies go after some of the SENS aging obstacles. Like design twenty different molecules or enzymes that break up different kinds of junk that accumulates over a lifetime outside of cells, as one example. Doing it through simulations and where the computer searches for solutions.

The market for such anti-aging medicine would be enormous, allowing them to deploy yet more supercomputer resources.

--aa2