3D NoC-Enabled Heterogeneous Manycore Architectures for Accelerating CNN Training

Joardar, Biresh Kumar; Choi, Wonje; Kim, Ryan; Doppa, Janardhan Rao; Pande, Partha Pratim; Marculescu, Diana; Mărculescu, Radu

doi:10.1145/3130218.3130219

Cited by 15 publications

(8 citation statements)

References 24 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Hence, designing the 3D NoC for heterogeneous systems is more complicated than homogeneous systems; this aspect has not been explored adequately. On top of this, 3D ICs suffer from thermal issues due to higher power density [8], [19]. One of the common methodologies for reducing the peak temperature in a 3D architecture includes proper core placement to prevent high power consuming cores from being placed on top of each other [19].…”

Section: D Heterogeneous Nocsmentioning

confidence: 99%

“…For a given workload, application-specific NoCs are known to outperform conventional architectures, e.g., mesh NoCs [7]. A MOO formulation for 3D NoCs is presented in [8] for accelerating deep learning workloads. In [22], the authors have explored heterogeneous NoC design for multimedia applications.…”

Section: D Heterogeneous Nocsmentioning

confidence: 99%

“…One of the common methodologies for reducing the peak temperature in a 3D architecture includes proper core placement to prevent high power consuming cores from being placed on top of each other [19]. However, it is not possible to implement such a strategy for heterogeneous systems with many GPU cores [8]. Other techniques to reduce temperature include suitable floorplanning [20] and temperature-aware task scheduling [21].…”

Section: D Heterogeneous Nocsmentioning

confidence: 99%

“…Basic MOO algorithms such as genetic algorithms (GA), e.g., NSGA-II [9], or simulated annealing-based algorithms, e.g., AMOSA [10], have been used in different optimization problems. AMOSA has been demonstrated to be superior to GAs or simulated annealing [10] and has been applied for the problem of heterogeneous NoC design in [7], [8]. However, since AMOSA is based on simulated annealing, it needs to be annealed slowly to ensure a good solution, which does not scale well with the size of the search space.…”

Section: Multi-objective Optimization Algorithmsmentioning

confidence: 99%

“…Therefore, the design process must consider reducing temperature hotspots as an additional objective. Overall, the design of a 3D heterogeneous manycore architecture needs to consider each of these objectives and satisfy all of them simultaneously [8]. Hence, 3D heterogeneous manycore design can be formulated as a multi-objective optimization (MOO) problem.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Learning-Based Application-Agnostic 3D NoC Design for Heterogeneous Manycore Systems

Joardar

Kim

Doppa

et al. 2019

IEEE Trans. Comput.

Self Cite

View full text Add to dashboard Cite

The rising use of deep learning and other big-data algorithms has led to an increasing demand for hardware platforms that are computationally powerful, yet energy-efficient. Due to the amount of data parallelism in these algorithms, high-performance threedimensional (3D) manycore platforms that incorporate both CPUs and GPUs present a promising direction. However, as systems use heterogeneity (e.g., a combination of CPUs, GPUs, and accelerators) to improve performance and efficiency, it becomes more pertinent to address the distinct and likely conflicting communication requirements (e.g., CPU memory access latency or GPU network throughput) that arise from such heterogeneity. Unfortunately, it is difficult to quickly explore the hardware design space and choose appropriate tradeoffs between these heterogeneous requirements. To address these challenges, we propose the design of a 3D Networkon-Chip (NoC) for heterogeneous manycore platforms that considers the appropriate design objectives for a 3D heterogeneous system and explores various tradeoffs using an efficient machine learning (ML)-based multi-objective optimization (MOO) technique. The proposed design space exploration considers the various requirements of its heterogeneous components and generates a set of 3D NoC architectures that efficiently trades off these design objectives. Our findings show that by jointly considering these requirements (latency, throughput, temperature, and energy), we can achieve 9.6% better Energy-Delay Product on average at nearly iso-temperature conditions when compared to a thermally-optimized design for 3D heterogeneous NoCs. More importantly, our results suggest that our 3D NoCs optimized for a few applications can be generalized for unknown applications as well. Our results show that these generalized 3D NoCs only incur a 1.8% (36-tile system) and 1.1% (64-tile system) average performance loss compared to application-specific NoCs.

show abstract

Section: D Heterogeneous Nocsmentioning

confidence: 99%