PRNGine: Massively Parallel Pseudo-Random Number Generation and Probability Distribution Approximations on AMD AI Engines

Generating large volumes of random numbers is essential for high-performance computing applications such as Monte Carlo simulations, machine learning, and dynamic game-play. Many of these applications require random number generation within a processing pipeline. Coarse-Grained Reconfigurable Archit...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:2025 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW) s. 91 - 98
Hlavní autoři: Bouaziz, Mohamed, Fahmy, Suhaib A.
Médium: Konferenční příspěvek
Jazyk:angličtina
Vydáno: IEEE 03.06.2025
Témata:
AI Engine
CGRA
Dataflow
Engines
Generators
High performance computing
Machine learning
Monte Carlo methods
Performance evaluation
Pipelines
PRNG
Probability distribution
Random number generation
Reconfigurable architectures
ISSN:2995-066X
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Generating large volumes of random numbers is essential for high-performance computing applications such as Monte Carlo simulations, machine learning, and dynamic game-play. Many of these applications require random number generation within a processing pipeline. Coarse-Grained Reconfigurable Architectures (CGRAs) are well-suited for this task, enabling efficient dataflow-based distribution across processing elements. This work explores efficient random number generation on AMD AI Engines (AIEs) through two execution models: a co-processor model and a standalone dataflow accelerator model. Key challenges in porting Pseudo-Random Number Generators (PRNGs) to AIEs, including the lack of support for certain operations, unsigned data types, and efficient vectorization, are identified and overcome. Additionally, the challenges of approximating a normal distribution on AIEs are analyzed and addressed. Optimized implementations of essential PRNG operations are presented, demonstrating linear complexity and enabling scalable random number generation. Performance evaluation provides insights into the suitability of both execution models for various applications.
ISSN:2995-066X
DOI:10.1109/IPDPSW66978.2025.00022