Core Count vs Cache Size for Manycore Architectures in the Cloud

The number of cores which fit on a single chip is growing at an exponential rate while off-chip main memory bandwidth is growing at a linear rate at best. This core count to off-chip bandwidth disparity causes per-core memory bandwidth to decrease as process technology advances. Continuing percore off-chip bandwidth reduction will cause multicore and manycore chip architects to rethink the optimal grain size of a core and the on-chip cache configuration in order to save main memory bandwidth. This work introduces an analytic model to study the tradeoffs of utilizing increased chip area for larger caches versus more cores. We focus this study on constructing manycore architectures well suited for the emerging application space of cloud computing where many independent applications are consolidated onto a single chip. This cloud computing application mix favors small, power-efficient cores. The model is exhaustively evaluated across a large range of cache and core-count configurations utilizing SPEC Int 2000 miss rates and CACTI timing and area models to determine the optimal cache configurations and the number of cores across four process nodes. The model maximizes aggregate computational throughput and is applied to SRAM and logic process DRAM caches. As an example, our study demonstrates that the optimal manycore configuration in the 32nm node for a 200mm die uses on the order of 158 cores, with each core containing a 64KB L1I cache, a 16KB L1D cache, and a 1MB L2 embedded-DRAM cache. This study finds that the optimal cache size will continue to grow as process technology advances, but the tradeoff between more cores and larger caches is a complex tradeoff in the face of limited off-chip bandwidth and the non-linearities of cache miss rates and memory controller queuing delay.