Multi-FPGA platforms like Amazon Web Services F1 are perfect to accelerate multi-kernel pipelined applications, like Convolutional Neural Networks (CNNs). To reduce energy consumption, we propose to upload at runtime the best poweroptimized CNN implementation for a given throughput constraint. Our design method gives the best number of parallel instances of each kernel, their allocation to the FPGAs, the number of powered-on FPGAs and their clock frequency. This is obtained by solving a mixed-integer, non-linear optimization problem that models power and performance of each component, as well as the duration of the computation phases-data transfer between a host CPU and the FPGA memory (typically DDR), data transfer between DDR and FPGA, and FPGA computation. The results show that the power saved compared to simply clock gating the fastest implementation is obviously very high, but it is also much more significant than simply scaling the frequency of the fastest implementation or replicating the slowest implementation on multiple FPGAs.

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.

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.

Ever since transistor cost stopped decreasing, customized programmable platforms, such as field-programmable gate arrays (FPGAs), became a major way to improve software execution performance and energy consumption. While software developers can use high-level synthesis (HLS) to speed up registertransfer level (RTL) code generation from C++ or OpenCL source code, placement and routing issues, such as congestion, can still prevent achieving an FPGA programming bitstream or dramatically reduce the FPGA implementation performance. Congestion reports from physical design tools refer to thousands of RTL signal names instead of developer-accessible identifiers and statements, considerably complicating the developer understanding and resolution of the issues at the source level. We propose a high-level back-annotation flow that summarizes the routing congestion issues at the source level by analyzing the reports from the FPGA physical design tools and the internal debugging files of the HLS tools. Our flow describes congestion using comments back-annotated on the source code and identifies if the congestion causes are the on-chip memories or the DSP units (multipliers/adders), which are the shared resources very often associated with routing problems on FPGAs. We demonstrate on realistic large designs how the information provided by our flow helps to quickly spot congestion causes at the source level and to solve them using appropriate HLS directives.