2019 · 55 citations · 24 references
EngineeringComputer EngineeringComputer ScienceComputer ArchitectureDesign OptimizationLow-power DesignAccelerator ArchitectureHardware AccelerationMemory Hierarchy DesignMemory Subsystem DesignMicroarchitecture DesignPhysical DesignAcceleratorsDeep Learning AcceleratorsEnergy-efficient Hardware AccelerationModel CompressionOn-chip Memory DesignSystolic Array DesignSystolic ArraysTechnology Scaling
Deep neural network inference on mobile SoCs demands energy‑efficient accelerators, but area constraints limit on‑chip SRAM, forcing costly off‑chip DRAM usage. This work explores technology‑system co‑design for systolic‑array DNN accelerators to evaluate alternative on‑chip memory technologies. The study conducts a design‑space exploration of SRAM, eDRAM, MRAM, and 3D VRRAM memories, applying model compression and buffer‑scheduling optimizations to systolic arrays and benchmarking on ResNet‑50, MobileNet, and Faster‑RCNN. MRAM accelerators achieve up to 4.68× energy savings with 57% area overhead, while 3D VRRAM designs deliver 2.22× energy savings and 33% area reduction compared to an SRAM/DRAM baseline.
Deep neural network (DNN) inference tasks have become ubiquitous workloads on mobile SoCs and demand energy-efficient hardware accelerators. Mobile DNN accelerators are heavily area-constrained, with only minimal on-chip SRAM, which results in heavy use of inefficient off-chip DRAM. With diminishing returns from conventional silicon technology scaling, emerging memory technologies that offer better area density than SRAM can boost accelerator efficiency by minimizing costly off-chip DRAM accesses. This paper presents a detailed design space exploration (DSE) of technology-system co-design for systolic-array accelerators. We focus on practical/mature on-chip memory technologies, including SRAM, eDRAM, MRAM, and 3D vertical RRAM (VRRAM). The DSE employs state-of-the-art optimizations (e.g., model compression and optimized buffer scheduling), and evaluates results on important models including ResNet-50, MobileNet, and Faster-RCNN. Compared to an SRAM/DRAM baseline, MRAM-based accelerators show up to 4.68× energy benefits (57% area overhead), while a 3D VRRAM-based design achieves 2.22× energy benefits (33% area reduction).
24
In-Datacenter Performance Analysis of a Tensor Processing Unit
Norman P. Jouppi, Cliff Young, Nishant Patil et al. · 2017 · 4.4K citations
Optimizing FPGA-based Accelerator Design for Deep Convolutional Neural Networks
Chen Zhang, Peng Li, Guangyu Sun et al. · 2015 · 2.1K citations
Song Han, Xingyu Liu, Huizi Mao et al. · ACM SIGARCH Computer Architecture News · 2016 · 2K citations
Ali Shafiee, Anirban Nag, Naveen Muralimanohar et al. · ACM SIGARCH Computer Architecture News · 2016 · 1.5K citations
Tianshi Chen, Zidong Du, Ninghui Sun et al. · 2014 · 1.3K citations