Increasing Microprocessor Performance with Tightly-Coupled Reconfigurable Logic Arrays

Conventional approaches to increase the performance of microprocessors often do not provide the performance boost one has hoped for due to diminishing returns. We propose the extension of a conventional hardwired microprocessor with a reconngurable logic array, integrating both conventional and reconngurable logic on the same die. Simulations have shown that even a comparatively simple and compact extension allows performance gains of 2{4 times over conventional RISC processors of comparable complexity, making this approach especially interesting for embedded microprocessors. The integration scale in the semiconductor industry is incessantly increasing. This trend allowed the implementation of increasingly complex systems on a single die. Dynamic instruction tracing and cache hit rate measurements show that the eeectiveness of increased cache size and higher superscalarity is leveling oo beyond a certain threshold. For instance, doubling the cache size from 64 to 128 KByte has been shown to increase the hit rate by only 0.4% 1] under certain loads. Average CPI values of commercial superscalar processors remain poor. A CPI of 1.2{3.0 was reported for quad-issue Alpha 21164 at 300MHz 2], a quad-issue HaL SPARC at 118MHz achieves 0.83 CPI under standard load 3]. These numbers can at least partly be attributed to bottlenecks in the memory subsystem , however the question arises whether the cost of larger caches and additional superscalar function units can be justiied by these rather lackluster performance improvements. Consequently, alternative approaches have been investigated, of which only a few can be mentioned here. Since memory is a major factor of system performance and the processor-memory gap is widening, integration of memory or additional special-purpose logic on the same die with a microprocessor core looks attractive. Projects like IRAM 2] and PPRAM 4] have shown promising results from this venue. Other researchers are abandoning the conventional computing