A note from Prof. Cox, builder of the Lego Pi Supercomputer
|
| Author | Content |
|---|---|
| Bob_Robertson Sep 12, 2012 2:31 PM EDT |
I asked him what the preliminary system speeds were, he replied with the following. Reprinted with permission: ===== Thank you for your interest in our work and for taking the time to drop me a mail. At the moment we have not done any systematic testing on the whole 64 processor system (yet) and, in reality, the processor is probably not the place to get performance from on the Pi - there is much more on the graphics side (and people are slowly getting to that). There is a some specific information on raw performance of each node at: http://www.raspberrypi.org/faqs "How powerful is it? The GPU provides Open GL ES 2.0, hardware-accelerated OpenVG, and 1080p30 H.264 high-profile decode. The GPU is capable of 1Gpixel/s, 1.5Gtexel/s or 24 GFLOPs of general purpose compute and features a bunch of texture filtering and DMA infrastructure. That is, graphics capabilities are roughly equivalent to Xbox 1 level of performance. Overall real world performance is something like a 300MHz Pentium 2, only with much, much swankier graphics." - which quantifies what I have said above. Also on small jobs such as the test calculation for Pi the time is dominated by the start of the MPI processes. As you can imagine, we were interested in how the price to get a system together has gone from 1,000,000s to 100,000s, to 10,000s and now down to 1000s of dollars/ pounds and that with Raspberry Pi you can really show (non virtualised) parallel programming in the raw at the 1000s dollar price point. That means (in principle) that it is accessible to students and school children and so brings a whole new generation that can be inspired and enabled to get into computational science and engineering. Also, at the moment, the core operating system images that are available are still being nicely tuned up with some boosts in performance between releases of the system images along with other kernel/ firmware updates and enhancements. There is also consideration of SD Card speed (we went for a fast 16Gb Class 10 in the end) and the fact that the 100Mb Ethernet actually hangs off USB2.0. This is all before one considers the normal challenges of getting ultimate performance in a supercomputer throught the switching fabric which connects the nodes together- all things you probably know a lot about! There is certainly quite a lot of interesting work to be done on performance - and something that anyone can get on with! We will be doing some more on this now we have the system up. Best Regards, Simon Prof Simon Cox University of Southampton |
| montezuma Sep 12, 2012 2:58 PM EDT |
Wow that is an impressive response in several ways Bob. Shall follow this project with interest. |
| Bob_Robertson Sep 12, 2012 3:46 PM EDT |
Monte, you are too kind. When, working at NASA, I saw in the numeric aerodynamic simulation center their Cray-2, a 64-CPU Connection Machine, their honest-to-Crom -CYBER- mainframe, from which we get Cyberspace, a Cray Y-MP, and many others. The cream of the crop from 40 years of computer development. Then came Linux and cheap clusters, and we all know what has happened to supercomputing since then. I can't find a reference to the EFF clustering of a bunch of 80286 mobos back in the early 1990s, which I distinctly recall started the entire clustering practice. Oh well. My organic memory is associative, so I cannot be certain how much is being "filled in" and how much is actual memory. Such is life! |
| montezuma Sep 12, 2012 6:01 PM EDT |
Yeah I use a big cheap personal cluster at work (NYU). 32 opteron cpus on 4 nodes. Runs Red Hat and is my dream machine. It is unbelievable how much more work I get done with this compared with the old group supercomputers (like the YMP) which I used to use in the 80s and 90s. It is so nice using linux on a laptop and a supercomputer. Incidentally the cluster only cost $7000 3 years ago. |
| Steven_Rosenber Sep 12, 2012 10:13 PM EDT |
I am very interested in how this Pi supercomputer is put together and how the coding needs to be changed to tap all at scale. I'm also interested in how these graphics capabilities are leveraged in normal Pi use and how they compare to other graphics chips. |
Posting in this forum is limited to members of the group: [ForumMods, SITEADMINS, MEMBERS.]
Becoming a member of LXer is easy and free. Join Us!