DNN accelerators are often developed and evaluated in isolation without considering the cross-stack, system-level effects in real-world environments. This makes it difficult to appreciate the impact of Systemon-Chip (SoC) resource contention, OS overheads, and programming-stack inefficiencies on overall performance/energy-efficiency. To address this challenge, we present Gemmini, an open-source, full-stack DNN accelerator generator. Gemmini generates a wide design-space of efficient ASIC accelerators from a flexible architectural template, together with flexible programming stacks and full SoCs with shared resources that capture system-level effects. Gemmini-generated accelerators have also been fabricated, delivering up to three orders-of-magnitude speedups over high-performance CPUs on various DNN benchmarks.

Linear algebra operations are at the heart of scientific computing solvers, machine learning and artificial intelligence. In this paper, LACore, a novel, programmable accelerator architecture for general-purpose linear algebra applications, is presented. LACore enables many of the architectural features typically available in custom supercomputing machines in an accelerator form factor that can be deployed in System-On-a-Chip (SoC) based designs. LACore has several architectural features including heterogeneous data-streaming LAMemUnits, a configurable systolic datapath that supports scalar, vector and multi-stream output modes, and a decoupled architecture that overlap memory transfer and execution. To evaluate LACore, we implemented its architecture as an extension to the RISC-V ISA in the gem5 cycle-accurate simulator. The LACore ISA was implemented in gcc, and a C-programming software framework, the LACoreAPI, has been developed for high-level programming of the LACore. Using the HPCC benchmark suite, we compare our LACore architecture against three other platforms: an in-order RISC-V CPU, a superscalar x86 CPU with SSE2, and a scaled NVIDIA Fermi GPU. The LACore outperforms the superscalar x86 processor in the benchmark suite by an average of 3.43x, and outperforms the scaled Fermi GPU by an average of 12.04x, within the same or less design area.