GPkit: A Human-Centered Approach to Convex Optimization in Engineering Design

Burnell, Edward; Damen, Nicole B.; Hoburg, Warren

doi:10.1145/3313831.3376412

Cited by 33 publications

(19 citation statements)

References 52 publications

(56 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We implemented our allocation heuristics in C++ and linked it to an existing GP solver [20]. To validate our optimization method, we use several widely used CNNs: AlexNet [6], VGG-net [7], YOLO [4] and ResNet [8].…”

Section: Resultsmentioning

confidence: 99%

CNN-on-AWS: Efficient Allocation of Multikernel Applications on Multi-FPGA Platforms

Shan

Lazarescu

Cortadella

et al. 2021

IEEE Trans. Comput.-Aided Des. Integr. Circuits Syst.

View full text Add to dashboard Cite

Multi-FPGA platforms, like Amazon AWS F1, can run in the cloud multi-kernel pipelined applications, like Convolutional Neural Networks (CNNs), with excellent performance and lower energy consumption than CPUs or GPUs. We propose a method to efficiently map these applications on multi-FPGA platforms to maximize the application throughput. Our methodology finds, for the given resources, the optimal number of parallel instances of each kernel in the pipeline and their allocation to one or more among the available FPGAs. We obtain this by formulating and solving a mixed-integer, nonlinear optimization problem, in which we model the performance of each component and the duration of the phases in which the accelerated computation can be split into, namely: 1) data transfer from a host CPU to the DDR memory of each FPGA, 2) data transfer from FPGA DDR to FPGA on-chip memory, 3) kernel computation on the FPGA, 4) data transfer from FPGA on-chip memory to FPGA DDR, 5) data transfer from FPGA DDR to host. Finding the optimal solution using a Mixed-Integer Non-Linear Programming (MINLP) solver is often highly inefficient. Hence, we provide a fast heuristic method that according to our experiments can be much more efficient than the MINLP solver and finds comparable results. For larger problems (more CNN layers), our heuristic method can quickly find (several thousand times faster) much better solutions than the MINLP solver, even if we run the latter for a very long time.

show abstract

Section: Resultsmentioning

confidence: 99%

CNN-on-AWS: Efficient Allocation of Multikernel Applications on Multi-FPGA Platforms

Shan

Lazarescu

Cortadella

et al. 2021

IEEE Trans. Comput.-Aided Des. Integr. Circuits Syst.

View full text Add to dashboard Cite

show abstract

“…We implemented our allocation heuristic in C++ and linked it to an existing efficient GP solver [2]. To validate our optimization method we used two widely used CNNs, AlexNet [5] and VGG [7].…”

Section: Resultsmentioning

confidence: 99%

“…Since we want to guarantee that at least one CU is instantiated per kernel, i.e.N k ≥ 1, it is possible thatn k =N k /F be less than one 2 . Kernel execution time and II becomê…”

Section: Firstmentioning

confidence: 99%

Exact and Heuristic Allocation of Multi-kernel Applications to Multi-FPGA Platforms

Shan

Casu

Cortadella

et al. 2019

Proceedings of the 56th Annual Design Automation Conference 2019

View full text Add to dashboard Cite

FPGA-based accelerators demonstrated high energy efficiency compared to GPUs and CPUs. However, single FPGA designs may not achieve sufficient task parallelism. In this work, we optimize the mapping of high-performance multi-kernel applications, like Convolutional Neural Networks, to multi-FPGA platforms. First, we formulate the system level optimization problem, choosing within a huge design space the parallelism and number of compute units for each kernel in the pipeline. Then we solve it using a combination of Geometric Programming, producing the optimum performance solution given resource and DRAM bandwidth constraints, and a heuristic allocator of the compute units on the FPGA cluster.

show abstract

“…Note that many interior point methods used to solve geometric and signomial programs do not require the user to provide an initial guess, although signomial programming heuristics typically require an initial guess for variables that appear in signomial constraints [3]. Initial guesses used in this paper are based on solutions found a priori using GPkit [11] with MOSEK [12] (academic license) as the backend solver.…”

Section: Choice Of Solversmentioning

confidence: 99%

The Power of Log Transformation: A Comparison of Geometric and Signomial Programming with General Nonlinear Programming Techniques for Aircraft Design Optimization

Kirschen

Hoburg

2018

2018 AIAA/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference

View full text Add to dashboard Cite

Geometric and signomial programming are emerging as promising methods for aircraft design optimization, having both been demonstrated to reliably and quickly find optimal solutions to aircraft design problems. To better understand how they perform compared with more conventional alternatives, this work presents a direct comparison with a general nonlinear programming approach. The crux of geometric programming, and by extension signomial programming, is in the formulation and the logarithmic transformation that makes the problem convex. Starting with the same problem formulation we assess the difference in speed and effectiveness achieved by performing the transformation. Two relatively small aircraft design problems, one a geometric program, the other a signomial program, are solved using the interior point and sequential quadratic programming alogrithms implemented in MATLAB's fmincon function, both with and without performing the log transformation first. Results show that performing the log transformation consistently yields the same optimal solution, independent of initial guess, whereas applying a general nonlinear programming technique directly to the un-transformed problem, at best, takes significantly longer and, at worst, terminates at an infeasible solution. The results also show that the general approach is highly sensitive to the initial guess whereas geometric and signomial programming approaches are not.

show abstract

GPkit: A Human-Centered Approach to Convex Optimization in Engineering Design

Cited by 33 publications

References 52 publications

CNN-on-AWS: Efficient Allocation of Multikernel Applications on Multi-FPGA Platforms

CNN-on-AWS: Efficient Allocation of Multikernel Applications on Multi-FPGA Platforms

Exact and Heuristic Allocation of Multi-kernel Applications to Multi-FPGA Platforms

The Power of Log Transformation: A Comparison of Geometric and Signomial Programming with General Nonlinear Programming Techniques for Aircraft Design Optimization

Contact Info

Product

Resources

About