6.6.2 Cache Architecture

The FRAM controller uses a 2-way associative cache that has a 64-bit line size. The cache stores prefetched instructions. The function of the FRAM controller is to prefetch four instruction words depending on the current PC location. The actual execution of these instructions is carried out in the cache. When the end of the cache buffer is reached, the FRAM controller preserves the four current words in the same cache line and fetches the next four words. If a code discontinuity is encountered at the end of a 2-way associative cache line, the cache is refreshed and the following four instruction words are retrieved from FRAM. However, if the application code loops back to a location already present in the cache when the last instruction in the cache is reached, the relevant instruction is simply executed directly from the cache instead of fetching code from FRAM again.

Note that only FRAM accesses are subject to the 8-MHz access limitation. When executing from cache, a system clock of up to 16 MHz can be used. Thus the cache is useful in (1) overcoming the 8-MHz limitation and increasing the average system throughput and (2) reducing overall active power by ensuring that most instructions are executed from it. Note that this is an instruction-only cache; all data is fetched directly from FRAM and is not cached.

The cached execution of instructions in the FR59xx family is different from the F2xx family, in which every instructions is directly executed from flash with no prefetches or caching, providing a 1:1 relationship between MCLK and instruction execution. For example, at MCLK = 16 MHz, eight two-cycle instructions can be executed in 16 clocks. For the FR59xx family, this relationship is application-dependent. The 1:1 relationship holds true only up to MCLK = 8 MHz. For MCLK > 8 MHz, the number of inserted wait states (directly proportional to how many times FRAM is accessed) determines the MCLK:instruction-execution ratio.

To provide another application example, with MCLK = 16 MHz, a JMP $ instruction (single cycle) is executed at the same rate in both devices. This is because the FR59xx fetches this instruction once and stores it in cache where it can be executed at the maximum MCLK speed. However, a loop that has more than eight instruction words would require accessing the FRAM every time a cache refresh is needed. These FRAM accesses take place at MCLK / 2 = 8 MHz, thereby reducing the overall throughput of the system when compared to an F2xx device.