Google llc (20240104012). TOPOLOGICAL SCHEDULING simplified abstract
Contents
TOPOLOGICAL SCHEDULING
Organization Name
Inventor(s)
Lukasz Lew of Sunnyvale CA (US)
TOPOLOGICAL SCHEDULING - A simplified explanation of the abstract
This abstract first appeared for US patent application 20240104012 titled 'TOPOLOGICAL SCHEDULING
Simplified Explanation
The abstract describes a method for performing topological scheduling on a machine-learning accelerator with an array of tiles. The method involves performing operations at each time step of a plurality of time steps corresponding to columns within wide columns of the tile array, including multiplications using tiles, computing output results for each tile column, and storing the results in an output RAM within the same tile column.
- Explanation of the patent/innovation:
- Performing topological scheduling on a machine-learning accelerator with an array of tiles - Multiplying using tiles in respective tile columns for each time step - Computing output results for each tile column by summing the results of the multiplications - Storing the output results in an output RAM within the same tile column
Potential applications of this technology: - Accelerating machine-learning algorithms - Improving the efficiency of neural network computations
Problems solved by this technology: - Optimizing the scheduling of operations on machine-learning accelerators - Reducing latency in processing large-scale neural networks
Benefits of this technology: - Faster execution of machine-learning tasks - Enhanced performance of neural network models
Potential commercial applications of this technology: - AI accelerators for data centers - Edge computing devices for real-time inference tasks
Possible prior art: - Previous methods of scheduling operations on machine-learning accelerators - Existing techniques for optimizing neural network computations
Questions: 1. How does this method compare to traditional scheduling techniques on machine-learning accelerators? 2. What impact does topological scheduling have on the overall performance of machine-learning models?
Original Abstract Submitted
methods, systems, and apparatus, including computer programs encoded on computer storage media, for performing topological scheduling on a machine-learning accelerator having an array of tiles. one of the methods includes performing, at each time step of a plurality of time steps corresponding respectively to columns within each of a plurality of wide columns of the tile array, operations comprising: performing respective multiplications using tiles in a respective tile column for the time step, computing a respective output result for each respective tile column for the time step including computing a sum of results of the multiplications for the tile column, and storing the respective output result for the tile column in a particular output ram having a location within the same tile column and on a row from which the output result will be read by a subsequent layer of the model.