Back from SC08. Recovery from the HPC discovery is in order.
The annual supercomputing (SC) show is always the week before the US Thanksgiving holiday. This schedule is a “Good Thing” as I always need a slow week at home to digest all the new and exciting news. I want to go on record stating “taking it easy week” has nothing to do with the late night activities that seem to be occurring in Austin last week. As matter of fact, for us writer types, it takes some time to wade though all the announcements and press releases that emerge during the show. I often wonder if maybe the biggest breakthrough is often lost in the deluge of announcements that rain down during the show. Correspondingly, there seems to be a drought in the months leading up to the show as companies withhold news in anticipation of SC.
While I am talking about press releases, I find many of them quite funny. There are generally two kinds. Those that actually have some news and those that seem to fall into the putting cosmetics on certain farm animals category. There is one rule, however, you must have a press release each year at SC. You must drive the attendees to the booth for the latest and greatest, or was it the free T-shirts. In any case, it is part of the “event.” The one thing that I find most amusing is the opening line of most press releases. It often goes something like this,
The Acme Widget Company, the leading widget company in the whole universe, announces …
I often wonder how there can be so many leaders in the market. Doesn’t anybody follow anymore? Just once I would like to see something like the following,
The Other Widget Company, a company that is not the leader, but has nicer people and is coming along quite well thank you very much, announces …
If someone wrote a press release like that, why, why it would be news. People might actually look at it and conclude that this company has certain honesty about it and maybe show some interest in the rest of the release. So much for my Andy Rooney moment.
Let’s get back to the show. I always preface any discussion of my “SC Experience” with the condition that I only talk about what I saw or touched. I do not re-quote press releases, although I do read them. With this in mind, I will undoubtedly miss some news or event that you or your company deem worthwhile. I suppose it is the luck of the draw. In any case, this year the emphasis seemed to be on co-processors — both GPGPU and FPGA types. There was of course the announced availability of the new (and better performing) Shanghai 45nm processor from AMD, but they were also pushing the “Fusion” concept at the show. AMD has a good white paper about “heterogeneous processors” or Fusion idea. Toward the end they even devote a small text box to “The Looming Crisis” (that would be the parallel software crisis for those that don’t follow my ranting about this issue). There was also much buzz about Intel Nehalem as well.
I manged to stop by the IBM booth to see a fully populated iDataPlex. Most importantly, I got to see the rear door heat exchanger. This thing is cool, no pun intended. Think of it as a large car radiator on the back of your rack. My prediction is that this will enable iDataPlex clusters to live in places that were never before possible, provided you can run chilled water.
As I mentioned, the GPGPUs were popping up everywhere. Nvidia Tesla and CUDA solutions were out in full force — including many desktop Tesla systems from hardware vendors. As I understand, Scott Morton of Hess Corporation had some nice geophysics performance numbers to report at the CUDA tutorial held on Monday. AMD/ATI was also making some announcements about their FireStream GPU boards. For programming, they are pushing Brook+ (AMDâ€™s modified Brook open source compiler) for now, but seem to be talking up OpenCL as the way forward. OpenCL is a language for programming heterogeneous data and task parallel computing across GPUs and CPUs. I sense a software battle looming in the near future.
While we are talking software, I want to mention the most interesting thing I saw at the show. I often visit Henry Dietz at the Aggregate.org booth to see some of the more leading edge research in HPC. Professor Dietz and his crew have developed a way to program SIMD GPGPU processors using a MIMD model. If you don’t know what SIMD or MIMD are, just remember, MIMD is what we want, but SIMD is what video cards give us. I’ll be watching this closely.
There were many more software companies touting tools for co-processor programming/computing. I’ll look at some of these in future columns, but one new company that caught my attention was Convey — which is supposed to sound a lot like Convex. For those who don’t remember, Convex was a supercomputer company co-founded by Steve Wallach. It turns out that Convey is a kind of a new Convex and is made up of mostly people from ex-Convex (ex-cons) including Wallach (Convex was purchased by HP in 1995). Their new product is an integrated hardware and software solution called the HC-1 or as they call it a “hybrid-core computer.” Essentially the HC-1 is a standard Xeon motherboard and a set of FPGAs that function like a vector processor. This vector unit connects directly into a standard Intel Xeon processor socket and provides a cache-coherent shared virtual memory. While other similar designs have been proposed, the unique aspect of this system is the integrated compiler suite that automatically vectorizes code. That is, a user can take a standard Fortran, C or C++ code, recompile it, and run it on the HC-1 for additional performance (up to 16X). Of course it all depends on how well the code vectorizes, but the idea is to minimize program changes and allow a much simpler and well known programming model.
Of course, there was plenty of “Cloud Computing Ready” hardware and software at the show. If you peal back the cloud computing labels you may see those “Grid Ready” lables that were so popular in years past. I’ll have more on the cloud thing in the future. For now, I’ll settle back and enjoy reading the market leading news from the market leading companies that have market leading products.
Douglas Eadline is the Senior HPC Editor for Linux Magazine.