Design trade-offs for load/store buffers in embedded processing environments

Memory latency is a critical issue for conventional high-speed computing platforms, and it is becoming a common problem in embedded and CMP (chip multiprocessing) systems as well. Conventional processors typically adopt caches and a Load/Store Queue (LSQ) to address the processor-to-memory bottleneck. However, the conventional LSQ design, which has a large number of entries, is not appropriate for embedded systems due to its area and power hungry out-of-order speculation. A compact, low-power load/store buffer that also provides significant performance improvement is essential for such systems. In this paper, we propose an area-efficient WideWord Load/Store Buffer (WLSB) which supports both WideWord (256-bit) and scalar (32-bit) load/store instructions for a recently fabricated PIM (Processing-In-Memory) device [1]. Given its small size, the 4 entry WLSB yields a 57.33% load hit rate on SPEC2K [3] benchmarks. This result is 5.72% better as compared to a less area-efficient 32-entry fully associative scalar load/store buffer (SLSB). The WLSB was synthesized in IBM 90 nm technology, and the resulting implementation occupies less than a seventh of a square mm and is projected to run at 1.6ns cycle time with 15.72mW of dynamic power dissipation. This paper demonstrates how this very small-entry buffer can affect the load hit rate and quantifies the design trade-offs between wide small-entry and narrow large-entry buffers with respect to size, power, load hit ratio and clock speed. Although this WLSB has been specifically designed to benefit a PIM architecture, it is expected to be useful for other embedded processing platforms and CMPs due to emphasized area/power constraints.