Chapter
Hardware Computation and MLIR with Chip Huyen
1:00:02 - 1:05:25 (05:23)
Chip Huyen, a machine learning engineer and teacher, talks about hardware computation and her work on MLIR, which is a project on the mid-level intermediate representation of computation and coordination of computers.
Clips
MLIR is a project under the graph about computation and coordination across a large number of potentially heterogeneous computers.
1:00:02 - 1:02:50 (02:48)
Summary
MLIR is a project under the graph about computation and coordination across a large number of potentially heterogeneous computers. Its goal is to represent computation and data movement between processing elements in a more efficient and clever manner.
ChapterHardware Computation and MLIR with Chip Huyen
Episode#162 – Jim Keller: The Future of Computing, AI, Life, and Consciousness
PodcastLex Fridman Podcast
The goal for PyTorch is to allow developers to compile and run their code seamlessly on hardware without needing to optimize it for performance.
1:02:50 - 1:04:28 (01:37)
Summary
The goal for PyTorch is to allow developers to compile and run their code seamlessly on hardware without needing to optimize it for performance. One example is turning big matrix multiplication into the right size chunks to use all processing elements without any extra tweaking.
ChapterHardware Computation and MLIR with Chip Huyen
Episode#162 – Jim Keller: The Future of Computing, AI, Life, and Consciousness
PodcastLex Fridman Podcast
Optimizing a GPU involves scheduling it just right, so as to avoid register conflicts, and is best handled by either an expert in micro-architecture, or by using pre-written libraries.
1:04:28 - 1:05:25 (00:57)
Summary
Optimizing a GPU involves scheduling it just right, so as to avoid register conflicts, and is best handled by either an expert in micro-architecture, or by using pre-written libraries. PyTorch makes it easier, as long as you have well-written code.