CS 101.3: Hacking the GPU
Graphics Cards and the Future of Scientific Supercomputing
In this new seminar/project course we will bring together graphics,
numerical analysis, computer languages and computer architecture as we explore
the use of programmable graphics processors (GPUs) for
scientific computing tasks. The next generation of graphics cards will
feature fully programmable floating point pipelines well matched to a
streaming model of computation. Architecturally they are the first
broadly available, commercial examples of high performance
architectures well suited to scientific supercomputing.
We will explore this computing paradigm in theory and (bleeding edge) practice
covering background on graphics pipelines and how they came to be the way
they are; different architectures and their related compute and communication
trade-offs; programming languages and semantics for high performance computations;
computational kernels from different numerical simulation areas; and their
implementation in hard and software.
This class is open to grads and undergrads with a background in either graphics, numerical analysis, computer languages or computer architecture. There will be a few homework assignments requiring implementation as well as a significant group project. We will provide access to some hardware and emulation software. Students are expected to contribute actively to the class through presentations and as scribes. There will be a number of guest lectures by experts in the relevant areas.
9 units (3-3-3); first term.