nVidia co-processing reduces power use by an order of magnitude

I had a briefing yesterday about Agilent’s ADS and EMPro software for signal integrity. The cool thing was that the demo was at nVidia, so I could check out the 3-D simulations on state-of-the-art hardware while looking at successes on real nVidia boards. Look for more information about this in my March 2010 article about signal integrity software.

There were a bunch of nVidia engineers at the demo, so talk naturally turned to their hardware. I am writing this blog post on my old magnificent Quadro FX 2000 professional dual-head CAD graphics card from nVidia, so I am already a big believer in their hardware. Check out what a late-model Quadro card can do— render 9 HD video feeds in real-time in Adobe Premier (YouTube). The CPU is loaded at 12% since the nVidia card is doing the hard work. I hope that my Sony Vegas will have the same performance, since I don’t like Premier and refuse to use Adobe products anyway. It might be time to drop a couple grand on a video workstation box.

What I learned from them in this meeting was that they have a whole division that sells their graphical processing units for use as coprocessors to do massive computing problems. They call this the Tesla product line, a nice name to a Croatian boy like me. The revelation for me was that just because these products are based on the nVidia GPU, they do not have to be used for graphics. You can slave them together to do rendering for your Hollywood movie or you can solve complex problems like oil field simulations, financial prediction, and weather modeling.

You can think of a Tesla card as an nVidia card with no monitor connectors. You just feed massively parallel problems to the card over the PCIe bus and it returns the answers back across the bus. You can slave dozens or hundreds or thousands of these processors together.

Things got really interesting when I asked for the power consumption per MIPs relative to a CPU. Mark Priscaro and Andy Walsh explained that the nVidia cards would easily save a factor of 10 over using multi-core CPUs to do the calculations. The speed is also scaled on that factor, since the nVidia cards can do a thousand parallel calculations if necessary, as opposed to a quad core doing four. Mark sent a pdf that shows real results from finance and oil geo-simulations. Performance per watt was 27 to 13 times better with GPUs, and performance per dollar was 20 to 10 times better.

The cool thing is that nVidia works closely with software companies to insure that applications can run on these massively parallel processors. I was also astonished to learn that some 3-D games I was watching were not 3D because the game designers created them that way. Instead, nVidia did the work to interpret 2-D screens and add depth so that the game looked like a true 3-D application. This is another example of how the gamer industry has really helped us engineers, whether it is designing in Solidworks or watching electromagnetic field simulations running in 3D on an Agilent ADS simulation. Thank you gamers. Thank you nVidia. Engineering is such fun, who would have thought?

I mentioned to Mark, a PR manager, what a shame it was that Silicon Graphics imploded once, since they had such leading-edge hardware. He smiled and admitted that nVidia had a lot of former SGI people. The brilliant things SGI did on the old days were not all lost. From what I remember, SGI’s mistake was that they got all starry-eyed about the Hollywood render-farm applications and dedicated the whole company to that, instead of focusing on the broader mechanical and architectural CAD market, which, lets face it, was not nearly as glamorous as those Hollywood projects. Thing is, once the Hollywood companies bought a dozen Octane boxes, they were set for a decade or so. So poor SGI was stuck with technology too expensive for gamers, and they also bet on UNIX over PCs. Rolling your own UNIX was not the right strategy, but who knew that back them?

Anyway, as to the fun, check out this great song about the Silicon Graphics boxes (mp3). The production values are great, and dorks that don’t understand engineering might think it is stupid to sing about crossbar switches, but that technology is really groundbreaking, and a testament to the SGI folks that they came up with these great concepts back in the 1990s.

And in honor of Agilent’s great 3-D presentation of electromagnetic field solvers, here is another old SGI song, a blues number about 3D graphics (mp3). Let the dweebs laugh; songs about 3D graphics are a lot cooler than 90% of the crap on the radio these days. “Am I seeing advanced 3D graphics, or just too much vermouth?” Now that’s a blues song, I don’t care who you are. Even if Silicon Graphics misread the market back then, the engineers were doing great things, and SGI still exists as a server and storage company. Check them out, if only because my buddy Dave’s brother still works there.