qGAMMA: A Framework for Optimal DNN Mapping on Multi-Precision Accelerators
Authors: Leandro Fiorin, Luigi Altamura, Cristina Silvano
Organizations: Dipartimento di Elettronica e Informazione, Politecnico di Milano, Milano, Italy. · Department of Computer Science and Engineering, Chalmers University of Technology and University of Gothenburg, SE-41296 Gothenburg, Sweden
This work describes qGAMMA, a flexible framework for exploring and optimizing the mapping of mixed-precision Deep Neural Network (DNN) models onto general-purpose spatial accelerators that support a range of fixed- and floating-point precisions. qGAMMA navigates the design space using a domain-specific genetic algorithm and an analytical performance model based on hardware synthesis results. The framework identifies the optimal mapping and accelerator configuration that minimizes target metrics. Extensive evaluations across DNN workloads show that qGAMMA offers valuable insights for designing mixed-precision applications, enabling better trade-offs between performance, energy efficiency, and resource utilization.
Figures & tables
Figure 1: An example of (a) CNN computation and (b) dataflow structure.
Framework
Exploration method
Mult. precision models
Mixed-precision
Timeloop [ 12 ]
Random sampling
Yes
No
GAMMA [ 9 ]
Genetic algorithm
No
No
LOMA [ 14 ]
Constrained brute-force
No
No
CoSA [ 6 ]
Mixed ILP formulation
No
No
SALSA [ 8 ]
Simulated annealing
No
No
This work
Genetic algorithm + heuristics
Yes
Yes
Table 1: Comparison of mapping frameworks’ characteristics in previous works and this work. qGAMMA is the first tool to support mixed-precision mapping on multi-precision spatial architectures.
Figure 3: Overview of the proposed framework: The red boxes highlight the key technical contributions of this study.
Figure 4: (a) Example of encoding for a 3-level genome and corresponding dataflow; (b) dataflow modelling of an INT8 operation in the proposed functional unit model.
Figure 5: (a) Latency and (b) number of generations observed for the mixed-precision ResNet18 optimized for minimum latency, when using the heuristics implemented in qGAMMA, and without them (original GAMMA search algorithm). The graphs include the generations needed to implement the early-stop. Per-layer precision is shown in Figure 5(a) .
Figure 6: (a) EDP and (b) latency of a ResNetV1 optimized for minimum EDP, considering three uniform floating-point precisions and per-layer optimization.
Figure 7: (a) Energy and (b) latency of a MobileNetV1 optimized for minimum energy, considering three uniform integer fixed-point precisions and single-cluster optimization.
Figure 8: Comparison of (a) latency and (b) energy consumption between a mixed-precision (INT8, INT4) ResNet18 and a single-precision (INT8) ResNet18, both optimized for minimum latency with single clustering optimization.
Ericsson Research, Ericsson AB, Stockholm, Sweden · KTH Royal Institute of Technology, Stockholm, Sweden · RISE Research Institutes of Sweden, Stockholm, Sweden