GPUs and CUDA - Ryan Bergmann

Attending

• Ryan Bergmann
• Katy Huff
• Jankai (Jack) YU
• Dan Wooten
• Prof. Max Fratoni
• Sandra Bogetic
• Christian DiSanzo
• Josh Howland
• Prof. Rachel Slaybaugh
• Kelly Rowland
• Nikola Radnovic

Lesson: GPUs and CUDA

Ryan Bergmann covered various features of GPUs and CUDA. Here is Ryan’s Tutorial.

Things we learned include:

• CUDA stands for Compute Unified Device Architecture.
• SIMD stands for Single Instruction Multiple Data.
• GPUs are good for turning compute-bound problems into memory-bound ones.
• CUDA cores aren’t really cores there are multiple cores per CUDA core.
• You have to use the SIMD lanes in order to get good performance out of a GPU system.
• Coalesced reading and writing means that your cores should be accessing adjacent pieces of memory simultaneously.
• The memory latency is higher for GPUs than CPUs, but the GPU hides this better the more threads you’re running.
• The host thread launches the GPU kernel
• Threads are organized into blocks
• Blocks are organzied into grids
• The grid is the kernel you have loaded.
• We learned how to launch a kernel for

Lightning Talks

• Katy gave a quick lightning talk on style guides for code.
• Kelly gave a more in-depth lightning talk on Laser Doppler Vibrometry.