27 November 2024 in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
Skip connections have emerged as a key component of modern convolutional neural networks (CNNs) for computer vision tasks, allowing for the creation of more accurate and deeper models by addressing the vanishing gradient problem. However, the existing implementations of field-programmable gate array (FPGA)-based accelerators for ResNets and MobileNetV2 often experience decreased performance and increased computational latency due to the implementation of skip blocks. This paper presents a novel framework for developing deep learning models on FPGAs that focuses on skip connections, with a unique approach to reduce buffering overhead. This results in a more efficient utilization of resources in the implementation of the skip layer. The nn2fpga compiler follows a thorough set of high-level synthesis (HLS) design principles and optimization strategies, exploiting in novel ways standard techniques to effectively map skip connection-based networks into static dataflow accelerators. To maximize throughput and efficiently use the available resources, our compiler employs a fast and effective design space exploration method based on a binary integer programming model which accurately assigns FPGA resources to the network layers, to maximize global throughput under resource constraints and then minimize resources for the achieved maximum throughput. Experimental results on the CIFAR-10 and ImageNet datasets demonstrate substantial gains in throughput (3× to 7× on past HLS-based work) for ResNet8, ResNet20, and MobileNetV2 models deployed on various Xilinx FPGA boards. Notably, MobileNetV2 deployed on the ZCU102 achieves a throughput of 2115 FPS, representing even a 10% speedup over a state-of-the-art highly optimized manual RTL implementation, showing that HLS can actually improve over manual design, thanks to the faster exploration of the design space.
10 April 2023 in IEEE Access
Flip-flops are the most used sequential elements in synchronous circuits, but designs based on latches can operate at higher frequencies and occupy less area. Techniques to increase the maximum operating frequency of flip-flop based designs, such as time-borrowing, rely on tight hold constraints that are difficult to satisfy using traditional back-end optimization techniques. We propose Mix & Latch , a methodology to increase the operating frequency of synchronous digital circuits using a single clock tree and a mixed distribution of positive- and negative-edge-triggered flops, and positive- and negative-level-sensitive latches. An efficient mathematical model is proposed to optimize the type and location of the sequential elements of the circuit. We ensure that the initial registers are not moved from their initial location, although they may change type, thus allowing the use of equivalence checking and static timing analysis to verify formally the correctness of the transformation. The technique is validated using a 28nm CMOS FDSOI technology, obtaining 1.33X post-layout average operating frequency improvement on a broad set of benchmarks over a standard commercial design flow. Additionally, the circuit area was also reduced by more than 1.19X on average for the same benchmarks, although the overall area reduction is not a goal of the optimization algorithm. To the best of our knowledge, this is the first work that proposes combining mixed-polarity flip-flops and latches to improve the circuit performance.
31 January 2024 in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
Flip-flops (FFs) are the most commonly used sequential elements in synchronous circuits, but their timing requirements limit the operating frequency. Borrowing time with a latch-based approach can increase operating frequency, but traditional backend optimization tools struggle to manage hold time requirements. The Mix & Latch technique achieves higher frequencies and often lower area than commercial state-of-the-art retiming by exploiting four types of synchronous sequential gates, namely positive and negative edge-triggered FFs and positive and negative transparent latches, all using a single clock tree. In this paper we first significantly accelerate the Mix & Latch flow convergence with respect to past work, by using a post-synthesis-based timing analysis that eliminates the first placement and routing needed for post-layout timing analysis. Then, by adding tolerance margins to the timing model, the pessimism is reduced to improve both convergence speed and maximum frequency. Finally, we reduce the complexity of the problem by applying the methodology only to the sequential elements belonging to critical paths. The effectiveness of Mix & Latch is then demonstrated on a RISC-V processor core from the Pulp platform using 28nm CMOS FDSOI technology. The results are compared to both the original Mix & Latch flow and a retiming performed with a state-of-the-art tool, showing a 25% frequency improvement over the original flow and 7.5% over the retiming flow. Compared to the retiming flow, we achieve comparable or lower power and area, while preserving the original registers and allowing logic equivalence checking.