Personal Supercomputing Survey Results

Print Friendly, PDF & Email

The polls are closed, the results are in and the votes from Florida have been reviewed.  We officially had sixty-one people respond to our survey on desktop supercomputing.  Sparked by recent product releases from Cray and SiCortex, John West and I were increasingly curious what the HPC market thought of such a computing paradigm?  Would these Top500 miniatures stand up to the industry criticism?  Why don’t we analyze the results.

In order to frame our results, we wanted to know what kind of an HPC individual they were [current user, admin, etc].  36% of respondents indicated they were currently working in an HPC center providing cycles to users.  30% said they were current users, 19.67% said they were not currently users and 14.75% said they not only worked at a center, but were also a user.  This was exactly the response we had hoped for.  It covers current users, future users and HPC decision makers.  Democracy in action!

Now we know who responded, lets take a look at how they responded.  73.33% of respondents said that a deskside machine would be a useful development tool in their daily work.  Wow.  Nearly the same percentage [67.80%] indicated that they currently have people on staff qualified to operate and utilize such a resource.  In my mind, this was one of the more surprising results.  I had always thought that one of the largest barriers to HPC adoption was the lack of human resources skilled in the art of HPC development and systems management.  These support resources tend to be difficult to find and expensive to hire and/or train.

This begs the question: Would the potential workloads fit well into a deskside supercomputer?  81.67% of respondents said that their applications are not physically or logically bound to their production-scale computational platform.  Applications tailored for platforms such as large shared memory machines [requiring TB’s of memory] or an IBM BlueGene would certainly have difficulty running on a deskside.  Otherwise, it seems that most of the respondents are running on commodity gear.  However, only 64.41% of respondents indicated that their applications would have the ability to utilize a parallel [read distributed memory] deskside platform similar to that of the Cray CX-1.  This is a bit puzzling.  This means that 35.59% percent of respondents are running applications that aren’t inherently parallelized past the cores they have in their current workstations.

The most interesting question and associated response was price.  We wanted to know what your threshold for pain was when it came to a deskside supercomputer.  47.46% of respondents indicated that $10,000+ was too high to justify purchasing a deskside unit.  25.42% said more than $20,000 would be too painful.  13.56% said more $50,000 would be too painful and the same percentage indicated that they didn’t really care, as long as it increased their productivity.  You heard it, nearly 50% of respondents want to pay less than the cost of a new Kia for their deskside super.  Whoa!  That’s going to be a tough number to hit for the vendor community.  I’ve seen several workstation offerings easily jump over $10k.  After engineering costs, packaging considerations and margin analysis, $10k is a tight rope to walk on.

We hope the survey results were useful to the community.  Personally, I would love to have a small super keeping my feet warm in the hot Texas summers.  If our small survey group reflects the thoughts of the industry at large, there is certainly a market for small, inexpensive cycles.  If our cost analysis is correct, SiCortex, HP and Cray should pay special attention to the likes of the Limulus Project.

For more info on the survey questions and results, check out the full report here [in pdf].


  1. […] John’s (West and Leidel) at did a nice study on personal supercomputing at the site. It is worth a […]


  1. I find this comment pretty ignorant:

    “Applications tailored for platforms … or an IBM BlueGene would certainly have difficulty running on a deskside.”

    The only thing needed to get tailor a code to run on a BlueGene effectively are paying attention to the memory usage per task and increasing maximum concurrency the code can run at.

    Neither of these things make a code difficult to run on a deskside. For example, QBOX, which ran @ 200 TF sustained on BG/L would work fine on such a machine.

  2. John Leidel says

    I believe there is a distinct difference between “running” and “running well.” Indeed, this phrase might be a bit misleading. Applications tailored to run on architectures whose per PE memory density is small, [BlueGene/L 256MB/PE] would certainly be able to “run” on a typical scalar deskside platform. There could be any number of inefficiencies moving to a [currently available] deskside platform with less than 64 PEs and 2GB/PE [depending upon the application]. Its simply a flip in platform architecture.

  3. You seem to be assuming that by reducing the memory needed per PE inefficiencies will be introduced. That is not necessarily the case.