RISC-V is a rapidly emerging open standard instruction set architecture (ISA), widely adopted in academia for its modularity and extensibility. Among its base ISAs, RV32IM and RV64IM share a common 32-bit instruction encoding but differ significantly in datapath width and associated hardware implications. While transitioning from a 32-bit to a 64-bit RISC-V core may appear straightforward, the hardware overhead remains largely unexplored, especially in interleaved multi-threaded (IMT) cores. To address this gap, we extended the Klessydra T03 core to 64 bit, with and without the M extension, and compared its hardware requirements on a Xilinx Virtex UltraScale VCU108 FPGA with the baseline 32-bit versions. The results show that widening the base architecture to 64 bits nearly doubles the LUTs and FFs utilization (1.98 \(\times \) and 1.95 \(\times \) , respectively), while reducing the maximum operating frequency by 13.5%. With the M extension enabled, LUT, FFs and DSP usage increases by 2.27 \(\times \) , 1.97 \(\times \) and 2.71 \(\times \) respectively, along with a 9.9% frequency drop. Our findings demonstrate that, while RV32IM and RV64IM share instruction formats, the hardware implications of XLEN widening in IMT cores are substantial.

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

Re-Architecting RISC-V: A 64-Bit Take on Interleaved Multi-threading

  • Andrea Marcelli,
  • Marco Pisani,
  • Rocco Martino,
  • Abdallah Cheikh,
  • Francesco Menichelli,
  • Mauro Olivieri

摘要

RISC-V is a rapidly emerging open standard instruction set architecture (ISA), widely adopted in academia for its modularity and extensibility. Among its base ISAs, RV32IM and RV64IM share a common 32-bit instruction encoding but differ significantly in datapath width and associated hardware implications. While transitioning from a 32-bit to a 64-bit RISC-V core may appear straightforward, the hardware overhead remains largely unexplored, especially in interleaved multi-threaded (IMT) cores. To address this gap, we extended the Klessydra T03 core to 64 bit, with and without the M extension, and compared its hardware requirements on a Xilinx Virtex UltraScale VCU108 FPGA with the baseline 32-bit versions. The results show that widening the base architecture to 64 bits nearly doubles the LUTs and FFs utilization (1.98 \(\times \) and 1.95 \(\times \) , respectively), while reducing the maximum operating frequency by 13.5%. With the M extension enabled, LUT, FFs and DSP usage increases by 2.27 \(\times \) , 1.97 \(\times \) and 2.71 \(\times \) respectively, along with a 9.9% frequency drop. Our findings demonstrate that, while RV32IM and RV64IM share instruction formats, the hardware implications of XLEN widening in IMT cores are substantial.