The John’s (West and Leidel) at InsideHPC.com did a nice study on personal supercomputing at the site. It is worth a read.
In short, they found people would find such boxen useful. But they don’t want to spend more than $10k for them.
This is interesting at many levels. Matches up very well with informal/anecdotal data we have from conversations with users.
We noticed that users wanted personal supers many years ago. They wanted something that could give immediate feedback, even if it wasn’t as fast as the heavy metal down the hall.
This is part of what drove us to build our Pegasus many-core systems. The purpose of these are to localize a significant amount of power. And they easily hit below the $10k price point, even for the 16 core systems.
While these are shared memory single units, you can run them with MPI as 8-16 way computing systems. We have quite a few customers doing exactly that. The desktops are so powerful, our customers are using them to run reasonable sized CFD and other simulations on them.
This said, distributed resources are more complex. You need N motherboards, N CPUs, and N memories. If you don’t mind inexpensive motherboards (aka desktop motherboards), and some strange cases, you can keep the cost per system pretty far down. Assume you can get a quad core MB + 4GB memory + CPU for about $800. The problem with most desktop motherboards are that they don’t include video out, and don’t take kindly to booting without this. You can find them though.
So for $800/unit, get 9 units, and put disks on one. Get a small switch. The issue is that the case and mounting issues will be hard. Your BOM costs would be close to $9000 when done. Under $10k for this type of home-brew system may be possible, but it would be hard.
Ok. What are the other options? This is why this is interesting. Accelerators.
With the right platform, you can add multiple units. With less complexity than the cluster.
I’ve been making the point for a while that if you give people the choice to manage a cluster or a single box as powerful as the cluster, the latter will win hands down.
And this is why the price profile of accelerators is so important.
You can create an awesome computational engine in well under $10k. We have. You can build a cluster in under $10k but you are going to need to skimp on resources to get there.
Most folks won’t want to build their own. Which is why the CX-1 and the SiCortex machines are interesting. And hopefully why Pegasus is interesting. We have been showing them for 2 years, and selling them for as long.
End users want many-core and parallel machines at their desk. This is known, and the study helps nail that point. They don’t want to spend more than $10k.
I am sure others will enter this market and get their products noted on HPCwire. Seems to happen a bit these days. Maybe we should figure out what they are doing, and do some of that ourselves. Our customers who know about Pegasus like them. Just like JackRabbit.