|
|
Automated Multi-Objective Parallel Evolutionary Circuit Design and Approximation
Hrbáček, Radek ; Fišer, Petr (oponent) ; Trefzer,, Martin (oponent) ; Sekanina, Lukáš (vedoucí práce)
Recently, energy efficiency has become one of the most important properties of computing platforms, especially because of limited power supply capacity of battery-power devices and very high consumption of growing data centers and cloud infrastructure. At the same time, in an increasing number of applications users are able to tolerate inaccurate or incorrect computations to a certain extent due to the imperfections of human senses, statistical nature of data processing, noisy input data etc. Approximate computing, an emerging paradigm in computer engineering, takes advantage of relaxed functionality requirements to make computer systems more efficient in terms of energy consumption, computing performance or complexity. Error resilient applications can achieve significant savings while still serving their purpose with the same or a slightly degraded quality. Even though new design methods for approximate computing are emerging, there is a lack of methods for automated approximate HW/SW design offering a rich set of compromise solutions. Conventional methods often produce solutions that are far from an optimum. Evolutionary algorithms have been shown to bring innovative solutions to complex design and optimization problems. However, these methods suffer from several problems, such as the scalability or a high number of fitness evaluations needed to evolve competitive results. Finally, existing methods are usually single-objective whilst multi-objective approach is more suitable in the case of approximate computing. In this thesis, a new automated multi-objective parallel evolutionary algorithm for circuit design and approximation is proposed. The method is based on Cartesian Genetic Programming. In order to improve the scalability of the algorithm, a brand new highly parallel implementation was proposed. The principles of the NSGA-II algorithm were used to provide the multi-objective design and approximation capability. The performance of the implementation was evaluated in multiple different applications, in particular (approximate) combinational arithmetic circuits design, bent Boolean functions discovery and approximate logic circuits for TMR schema. In these cases, important improvements with respect to the state of the art were obtained.
|
|
|
Využití přibližné ekvivalence při návrhu přibližných obvodů
Matyáš, Jiří ; Lengál, Ondřej (oponent) ; Češka, Milan (vedoucí práce)
Tato práce je zaměřena na využití formálně verifikačních technik pro návrh funkčních aproximací kombinačních obvodů. Jsou zde důkladně prostudovány existující formální přístupy pro zkoumání přibližné ekvivalence a jejich použití při vývoji aproximovaných obvodů. V rámci této práce je navržena nová metoda, která integruje vybrané formální techniky do Kartézského genetického programování. Klíčovým bodem nového přístupu je využití prohledávací strategie, která vede evoluci směrem k řešením, která lze rychleji verifikovat. Navržený algoritmus byl implementován v rámci syntézního nástroje ABC. Jeho výkonnost byla otestována na vývoji funkčních aproximací násobiček a sčítaček s šířkami vstupních operandů 32, respektive 128 bitů. Dosažené výsledky ukazují výjimečnou škálovatelnost navržené metody.
|
|
|
Exploiting Approximate Arithmetic Circuits in Neural Networks Inference
Matula, Tomáš ; Mrázek, Vojtěch (oponent) ; Češka, Milan (vedoucí práce)
This thesis is concerned with the utilization of approximate circuits in neural networks to provide energy savings. Various studies showing interesting results already exist, but most of them were very application specific or demonstrated on a small scale. To take this further, we created a platform by nontrivial modifications of robust open-source framework Tensorflow allowing us to simulate approximate computing on known state-of-the-art neural networks e.g. Inception or MobileNet. We focused only on replacement of most computationally expensive parts of convolutional neural networks, which are multiplication operations in convolution layers. We experimentally demonstrated and compared various setups and even that we proceeded without relearning, we were able to obtain promising results. For example, with zero accuracy loss on Inception v4 architecture, we gained almost 8% energy savings which could be valuable, especially in low-power devices or in large neural networks with enormous computational demands.
|
|
|
Automated Multi-Objective Parallel Evolutionary Circuit Design and Approximation
Hrbáček, Radek ; Fišer, Petr (oponent) ; Trefzer,, Martin (oponent) ; Sekanina, Lukáš (vedoucí práce)
Recently, energy efficiency has become one of the most important properties of computing platforms, especially because of limited power supply capacity of battery-power devices and very high consumption of growing data centers and cloud infrastructure. At the same time, in an increasing number of applications users are able to tolerate inaccurate or incorrect computations to a certain extent due to the imperfections of human senses, statistical nature of data processing, noisy input data etc. Approximate computing, an emerging paradigm in computer engineering, takes advantage of relaxed functionality requirements to make computer systems more efficient in terms of energy consumption, computing performance or complexity. Error resilient applications can achieve significant savings while still serving their purpose with the same or a slightly degraded quality. Even though new design methods for approximate computing are emerging, there is a lack of methods for automated approximate HW/SW design offering a rich set of compromise solutions. Conventional methods often produce solutions that are far from an optimum. Evolutionary algorithms have been shown to bring innovative solutions to complex design and optimization problems. However, these methods suffer from several problems, such as the scalability or a high number of fitness evaluations needed to evolve competitive results. Finally, existing methods are usually single-objective whilst multi-objective approach is more suitable in the case of approximate computing. In this thesis, a new automated multi-objective parallel evolutionary algorithm for circuit design and approximation is proposed. The method is based on Cartesian Genetic Programming. In order to improve the scalability of the algorithm, a brand new highly parallel implementation was proposed. The principles of the NSGA-II algorithm were used to provide the multi-objective design and approximation capability. The performance of the implementation was evaluated in multiple different applications, in particular (approximate) combinational arithmetic circuits design, bent Boolean functions discovery and approximate logic circuits for TMR schema. In these cases, important improvements with respect to the state of the art were obtained.
|
|
|
Exploiting Approximate Arithmetic Circuits in Neural Networks Inference
Matula, Tomáš ; Mrázek, Vojtěch (oponent) ; Češka, Milan (vedoucí práce)
This thesis is concerned with the utilization of approximate circuits in neural networks to provide energy savings. Various studies showing interesting results already exist, but most of them were very application specific or demonstrated on a small scale. To take this further, we created a platform by nontrivial modifications of robust open-source framework Tensorflow allowing us to simulate approximate computing on known state-of-the-art neural networks e.g. Inception or MobileNet. We focused only on replacement of most computationally expensive parts of convolutional neural networks, which are multiplication operations in convolution layers. We experimentally demonstrated and compared various setups and even that we proceeded without relearning, we were able to obtain promising results. For example, with zero accuracy loss on Inception v4 architecture, we gained almost 8% energy savings which could be valuable, especially in low-power devices or in large neural networks with enormous computational demands.
|
|
|
Využití přibližné ekvivalence při návrhu přibližných obvodů
Matyáš, Jiří ; Lengál, Ondřej (oponent) ; Češka, Milan (vedoucí práce)
Tato práce je zaměřena na využití formálně verifikačních technik pro návrh funkčních aproximací kombinačních obvodů. Jsou zde důkladně prostudovány existující formální přístupy pro zkoumání přibližné ekvivalence a jejich použití při vývoji aproximovaných obvodů. V rámci této práce je navržena nová metoda, která integruje vybrané formální techniky do Kartézského genetického programování. Klíčovým bodem nového přístupu je využití prohledávací strategie, která vede evoluci směrem k řešením, která lze rychleji verifikovat. Navržený algoritmus byl implementován v rámci syntézního nástroje ABC. Jeho výkonnost byla otestována na vývoji funkčních aproximací násobiček a sčítaček s šířkami vstupních operandů 32, respektive 128 bitů. Dosažené výsledky ukazují výjimečnou škálovatelnost navržené metody.
|
host ::
RSS.