Higher Performance and Capacity with OMI Near Memory

Storage interconnects such as CXL OpenCAPI, and Gen-Z promise to change computing by providing greater amounts of heterogeneous memory and allowing disaggregation of memory and computing. However, these Far Memory (higher latency) approaches do not address the needs of low latency, Near Memory, currently provided by parallel interface Double Data Rate (DDR), Low Power DDR (LPDDR) or High Bandwidth Memory (HBM) technologies. This paper will look at the current and future capabilities of DDR and HBM Near Memory and compare these conventional approaches to a low latency serial approach, the Open Memory Interface (OMI). OMI can provide higher Near Memory capacities than DDR or HBM, with performance that comes close to that available from HBM and at lower cost than HBM. OMI also requires significantly less processor die area than DDR, LPDDR or HBM to support high bandwidth, can reduce total memory power consumption and allows caching and new types of Near Memory (similar to what CXL provides for Far Memory). OMI has significant promise to help provide Near Memory to support large data processing applications.