Gerardo Hernandez Correa, Application Engineer with MathWorks, with be on campus offering a complementary MatLab seminar:
Parallel and GPU Computing in MATLAB
Date: Thursday, February 7
Time: 9:30 a.m. – 11:30 a.m. (Arrive between 9:15 – 9:30 for registration/sign in.)
Location: Bryan Research Building for Neurobiology, Room 101L (Directions)
Session Description & Registration Available online:
Questions: Contact Tom McHugh at (508) 647-7657 firstname.lastname@example.org.
This Powerpoint slides for today’s Condor talks are here.
The bdgpu GPGPU compute nodes are being upgraded from CUDA 2.3 to CUDA 4.1. This process should be complete by Sunday, April 8th. This does not affect CPU jobs running on these nodes. If you have problems with GPGPU applications not running in the new environment, please send an email to email@example.com.
We haven’t done a purchasing round in a while, partially because we were waiting to hear about the availability of the newest Intel CPUs. The latest word is that the new CPUs are still “coming soon” — rather than wait any longer, we’d like to offer the following set of options for new DSCR purchases:
These prices all represent a significant decrease from the last purchasing round; the larger memory sizes seem to be exceptionally good deals and we know that many of you are starting to run into memory-size issues with your applications. These prices do NOT include the Condo Service fee.
Please let us know if you are interested in making a purchase. We’d like to have fund-codes ready on Friday, December 16th to let our finance team get them into the system before the holidays.
We are pleased to announce that our new GPU-computing cluster is ready for general use.
For those who may not have heard of this before, Graphics Processing Units (GPUs) are now highly programmable — suitable for running general purpose computations, not just image/pixel processing algorithms. The newest GPUs are capable of nearly 900 GFLOPS of performance for single-precision and integer calculations, and over 70 GFLOPS for double-precision. Real-world applications are seeing 10x to 50x performance improvements relative to CPU-based programs.
Duke’s new GPU-cluster is a part of another new effort called the Blue Devil Grid (BDgrid) — a campus-wide computational grid that it is separate from the DSCR. There are 16 GPU-enabled machines in the BDgrid — 11 of them contain consumer-grade GPUs (GTX-275), 4 contain high-end Tesla GPUs, and one contains two Tesla GPUs. You should be able to login to the front-end machine with your NetID and password — just ssh to bdgpu-login-01.oit.duke.edu — and launch jobs onto the GPU-cluster using Condor.
For more information, see our wiki:
We are also offering an OpenMP seminar next Wednesday which will include a new approach to GPU programming — Portland Group’s “Accelerator” Framework — which may enable you to port your program to the GPU very quickly. See our Training page for more info:
On Dec 4th, Intel announced that they were canceling retail products based on the forthcoming Larrabee chip. The Larrabee was to be Intel’s answer to GPGPU computing (e.g. CUDA, OpenCL) but was apparently plagued by numerous delays. As recently as the Nov SC’09 conference, Intel was demo’ing Larrabee samples at 1 TFLOPS. These samples will not make it to your desktop though there are plans to produce some Larrabee “development platforms” — presumably to help software developers get ready for the next (?) Intel graphics/HPC chip — but not a fully support, retail product.
At the recent Supercomputing ’09 conference in Portland, OR, there were almost too many GPU-compute products and announcements to mention. All the usual suspects (system vendors) were showing off their new GPU-enabled hardware — Dell, Penguin, Microway, Silicon Mechanix, Appro, you name it — everyone’s got GPUs. Even Cray, a company once synonymous with vector-based computers, now sells GPU-enabled “deskside supercomputers”.
The SCSC has also been experimenting with GPU-computing for several months now. We now have 16 GPU-enabled servers that you can access for testing and benchmarking. 10 of those systems are single-CPU (quad-core) machines with NVIDIA GTX-275 (mid-range, consumer-grade graphics cards with 768MB of GPU-memory) and the remaining 6 are single-CPU (quad-core) machines with high-end Tesla C1060 cards (with 4GB of GPU-memory). The six Tesla cards were a generous donation from NVIDIA.
If you haven’t heard all the hype — GPUs can be 10x or 100x faster than a CPU — you should look at NVIDIA’s CUDA Zone webpage.
Contact us (scsc at duke) if you would like more information, or if you would like to try out the GPU-compute cluster.
Some GPU-computing news: