Abstract

G-GPU introduces a general-purpose, GPU-like ASIC accelerator designed for energy-efficient, high-performance parallel processing in Systems on Chip (SoCs). The core innovation is GPUPlanner, an open-source, fully-automated platform that generates these complex accelerators from RTL to tapeout-ready GDSII layout. Compared to the popular RISC-V architecture, the generated G-GPU designs demonstrate massive performance gains, achieving up to 223 times raw speed-up and 11 times better performance when normalized by area.

Report

Key Highlights

• Core Innovation: G-GPU is a general-purpose, GPU-like accelerator designed to improve energy efficiency and throughput in SoCs without being application-specific.
• Automation Solution: The paper introduces GPUPlanner, the first known open-source, fully-automated platform capable of generating GPU-like ASIC accelerators from RTL through to GDSII (tapeout-ready layout).
• Performance Metrics: G-GPU designs achieve exceptional performance relative to standard RISC-V CPU architectures, showing raw performance speed-ups of up to 223 times.
• Efficiency Metric: Performance derated by area shows gains of up to 11 times compared to RISC-V.
• Fabrication Ready: Tapeout-ready layouts of the G-GPU were successfully generated in 65nm CMOS technology.

Technical Details

• Platform Name: GPUPlanner (open-source generator).
• Design Scope: Handles the full design flow, addressing the gap in automated generation for GPU-like ASIC architectures.
• Optimization Strategy: Utilizes a design space exploration (DSE) approach to find optimal accelerator configurations.
• Architectural Methods: Achieves efficiency through a smart breakdown of the memory hierarchy and on-demand pipelining of the logic.
• Output Standard: Generates layouts that are GDSII compatible for tapeout.

Implications

• Democratization of Acceleration: GPUPlanner lowers the technical barrier for SoC designers to incorporate highly complex, customized, high-performance accelerators, shifting the focus from manual design to automated generation.
• Enhancing RISC-V Ecosystem: As RISC-V systems increasingly rely on specialized co-processors for efficiency, G-GPU provides a robust, pre-verified, and highly competitive general-purpose GPU solution that can seamlessly integrate into heterogeneous RISC-V dominated SoCs.
• High Performance Computing (HPC) on Chip: The substantial speed-ups (up to 223x raw) confirm that G-GPU offers a viable path toward integrating extremely high-throughput parallel processing capabilities directly into custom ASICs, suitable for demanding modern workloads.