The Technique for Data Parallelism in Neural Processing Units
| Autor: | V. A. Romanchuk, Ruslan Bazhenov |
|---|---|
| Rok vydání: | 2019 |
| Předmět: |
Dynamic programming
0209 industrial biotechnology 020901 industrial engineering & automation Optimization problem Computer science Data parallelism 0202 electrical engineering electronic engineering information engineering Executable compression Code (cryptography) Complexity class 020201 artificial intelligence & image processing 02 engineering and technology Parallel computing |
| Zdroj: | Advances in Artificial Systems for Medicine and Education II ISBN: 9783030120818 |
| DOI: | 10.1007/978-3-030-12082-5_4 |
| Popis: | In this paper, the authors propose a technique for efficient data parallelism in neural processing units through different dimensional data subsets and redistribution of similar operations between code segments that are executed in parallel. The authors observe a combined approach to optimize a solution of the one-dimensional optimization problem. The authors also consider a category of the neural processor bit depth, based on dynamic programming methods. Empirical study proves that the application of the method offered can improve significantly overall program instruction per second by 5–14%, depending on a complexity class of decision problem and the degree of operation homogeneity. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|