2 ms·
GPU Computing Era (2010 / Paywall)
- dragontamer 8y agoAlthough this article is behind the IEEE Paywall, it is a good piece of work that gently introduces GPGPUs to the technical reader. The concept of thread-blocks, shared memory, and high-speed specialized graphics-RAM applies just as well today in 2019 as it did back in 2010. Today's GPUs are bigger, faster, and more parallel than ever before, but the fundamental model hasn't changed dramatically in the past 10 years. In the past 10 years: CUDA has gotten easier, more C++ features have been added to the language. "Tensor" operations to accelerate deep-learning have been added. And I guess "ballot" and "cross-lane permute" operations have been added as well. But overall, the concept of using SIMD-cores to run thread-like code has proven to be highly successful. As such, this introductory article from 2010 is still very relevant.