|Deletions are marked like this.||Additions are marked like this.|
|Line 27:||Line 27:|
|== Next Meeting: 2-3:30pm, Fri 30 Sept, University Place ==||== Next Meeting: 2-3:30pm, Fri 30 Sept, 1.219 University Place ==|
|Line 32:||Line 32:|
|* Will Thompson, Maths: image reconstruction|
University GPU Club
Software for GPUs inc. compiler/directives, maths libs & tools (debuggers and profilers)
The GPU Club is an initiative to bring people together to share ideas and experiences in order to more effectively overcome any barriers to getting their research done. Since November, the Club has set up a email list, held two meetings, set up a SpecialInterestGroup and negotiated with vendors regarding training and hardware provision.
Graphics Processing Units (GPUs) are fast becoming an option for undertaking intensive computation tasks without having to buy expensive compute nodes with fast interconnects. If one is to believe the sales pitches from the main vendors, NVIDIA and AMD/ATi, it’s possible to have your own personal supercomputer with 448 cores capable of peak performances of a teraflop for just £3,578 - about half the total performance of the University’s old “Horace” service. (The price is the online ex-VAT price for Amari 7046GT-TRF with a single NVIDIA C2050 card. Single precision performance.)
But there are many questions:
- Is this peak performance achievable for real life codes?
- What programming methods are best supported by these cards?
- How much effort is involved in porting?
- CUDA or OpenCL? NVIDIA or AMD/ATi? Or does it matter?
- Where can I get my hands on one to see if it is for me?
There is a lot of interest across all faculties on campus concerning GPUs. Many people are seriously considering their use for their future research. Several groups have already, or are in the process of, writing and running models on GPUs.
Join the Mailing List
We have a mailing list that you can join to share news and ask questions relating to GPUs. To join you need to log in to listserv and search for GPUCLUB (or click here), select the GPUCLUB email list then select Subscribe from the drop down list beneath and click on 'Submit Changes'. Full instructions are given at the bottom of the page.
Next Meeting: 2-3:30pm, Fri 30 Sept, 1.219 University Place
Confirmed speakers are
- Introduction, ITSfR: GPU research at UoM, forthcoming training
- Perri Needham, Chemisty: To C or not to C: GPU programming in FORTRAN
- Daniel Valdez Balderas, MACE: Smoothed Particle Hydrodynamics Simulations on Multi-GPU Clusters
- Will Thompson, Maths: image reconstruction
- Would Manchester like to host a GPU conference such as MRSC in 2012?
We're awaiting confirmation on a couple of additional speakers, talking on real-time analysis using GPUs, and we welcome suggestions for additional speakers.
Please login and add your suggested topics and speakers for future meetings:
- Is GPU HPC?
- GPU-enabled computational libraries
- Comparison of GPU technologies for research codes
Fri 15 July: Prof. Jack Dongarra "Emerging Technologies"
About 60 attended to hear Prof. Jack Dongarra speaking on Emerging Technologies.
Graphics Processing Units (GPUs) are fast becoming an option for undertaking intensive computational research without having to buy expensive computers with fast interconnects. The 100+ people at each University GPU Club pays testament to the interest in this area, but what are the barriers to using GPUs for High End Computing (HEC) and how will we overcome them?
GPU technology isn't standing still. Chip manufacturers are looking at combining elements of traditional CPUs, memory and GPUs, Field Programmable Gate Arrays (FPGAs) are in re-ascendence and research abounds in the hardware architecture field.
Hardware is only part of the issue: how to program new hardware efficiently will increasingly be key.
Prof. Jack Dongarra, Director of the Innovative Computing Laboratory at the University of Tennessee and Turing Fellow at the University of Manchester, has been bestowed with numerous prestigious awards in the fields of HEC and numerical libraries including IEEE Medal of Excellence in Scalable Computing, SIAM SIG Supercomputing Career Achievement and the 2011 IEEE IPDPS Charles Babbage Award. He is a Fellow of the AAAS, ACM, IEEE and SIAM.
Friday 18 March 2011
About 70+ people attended, with 48 signing attendance sheet (not fully circulated) of which 12 using Cuda and 5 using OpenCL with 34 asking for training. The following presentations were given:
IT Services for Research Team: Overview of GPU Computing at UoM covering GPU facilities available to UoM researchers, GPU enabled software and GPU training available to UoM researchers
Sam Proctor: Initial Experiences with OpenCL
Igor Kozin: Tools for GPU
Daniel Goodman: Experiences of porting C to CUDA
After the presentations the following points were discussed and actions agreed.
- re training: a show of hands indicated about 15 would attend OpenCL and about the same would attend CUDA training course (from attendance sheets: 34 people had indicated request for training)
On hardware issues:
regarding future GPU cards, University model is for researchers to pay for hardware (the contribution model: pay for some, get access to all facilities), but ITSfR has submitted proposal to AMD for some kit. Specific question about AMD Fusion, NVIDIA/Arm Tegra and Tilera chips, but same answer applies. CSF is a contribution model where equipment can be hosted. Researchers are encouraged to contribute. Talk to ITSfR.
- NVIDIA had gone lukewarm on providing training or hardware but ITSfR hopeful that the University's growing GPU cluster would entice increased interest from them in future, as per AMD's interest.
- ITSfR to ask again about how University could become an NVIDIA Research Centre, eg whether teaching an under/post-grad course using CUDA would suffice.
On applications, software and performance:
- ITSfR offered a talk from Mathworks on their CUDA-based Parallel Computing Toolbox (PCT). Sufficient hands went up for ITSfR to organise.
- "My OpenCL code is slower on my laptop than the CPU code. Can I get access to a bigger/faster GPU?". Yes, talk to ITSfR to get access to various GPU systems, both at the University, at DISCO and as part of the national HPC service, HECToR.
Future discussion topics for future meetings or SIGs:
Comparison of OpenMP/MPI/CUDA/OpenCL and of hardware, particularly for user codes. ITSfR to coordinate and everybody to make use of the Community Wiki. ITSfR have proposal in to EPSRC/HECToR to investigate multi-GPU implementations & Ben Rogers' group have MPI-Cuda version of their SPH work
CFD (OpenFoam, Code Saturne) and MD mentioned as key topics for future discussion
- Aqeel Al-Naser interested in integration of OpenGL and OpenCL.
Summary of agreed actions:
- ITSfR to organise talk by Mathwork about PCT
- ITSfR to enquire about NVIDIA Research Centre
- ITSfR to coordinate comparison of soft/hard-ware and different programming paradigms; everybody to put their experiences on the Community Wiki
ITSfR to organise next meeting for around June, possibly with Matt Harvey/ACEMD & Ben Rogers' group presenting
Inaugural Meeting: 26 November 2010
About 100 people turned up/expressed interest in the inaugural meeting of the University GPU Club at which the following presentations were given:
Dr. Michael Bane Introduction
Dr. Andrew Almond, "Exploring Carbohydrate Free Energy Landscapes Using GPUs"
Dr. Ben Rogers, “GPUs for the Particle Method SPH: Issues and Challenges”
De. Xuefei He, "GPU Monte Carlo Simulation for VaR Calculation"
Dr. Nick Dingle, "GPU-enabled Steady-State Solution of Large Markov Models"
Summary of agreed actions:
set up an email list for discussion eg of problems
- Community Wiki: everybody to contribute eg with solutions and their areas of interest
- regular meetings: agreed meetings are useful, 2-monthly frequency
- learning: people interest in CUDA and OpenCL courses