DART: A Framework for Determining Approximation Levels in an Approximable Memory Hierarchy

Software and hardware interfaces for approximate memories

“…In a scenario with an approximate DRAM, a precise cache also alleviates the impact of errors from this main memory because of the reduced number of accesses on the main memory. Although previous proposals mitigate approximation among more than one level in the memory hierarchy [77,130], our preliminary experiments, shown in Section 3.3.4, evidence that a precise cache can reduce the number of application-visible data errors from the last-level DRAM. Thus, the approximation only at the DRAM main memory could maximize its energy gains with less impact on the output quality using precise caches.…”

“…Static configurations assume that the approximation level for the application does not change in its execution phase. However, interfaces with dynamic configuration [78,95,112,130,129] usually adjust the approximation knob according to the workload or the different quality thresholds required per input. SmartApprox adjusts the knob according to the error changes at runtime, which may be affected by environment variables, such as temperature.…”

“…Approximate memories allow for adjusting parameters out of the guard-band margin while exposing stored data to errors through a redesign of the memory data interface to reduce energy consumption [37]. These errors are nondeterministic and may occur at any point on the exposed data according to a probability related to the configuration parameters, which provides control of the approximation [130]. Figure 1.2 shows the relative energy consumption collected from the execution of diverse applications with an instance of an approximate DRAM face to the error probability per access in this memory, based on an error characterization of the DRAM voltage [12] executing on a controlled environment with a median instance of error rate [30].…”

“…The configuration of an approximate memory is application-specific and depends on memory access patterns of the application [126]. Several configuration interfaces [56,77,112,129,130] also depend on annotations or on application-specific metrics to evaluate the acceptance of the results and adjust the parameters of the approximation as knobs. The additional complexity of the configuration is to tackle the error fluctuation that changes according to the hardware instance and environment, comprising an error scenario dependent on some variables, such as temperature, access delay, and fabrication process [13,65].…”

“…DART [130] is a framework for determining approximation levels among the memory hierarchy considering different approximation techniques on SRAM caches and DRAM main memory. This framework has an annotated application as input with specified maximum error magnitude for each approximate variable and extracts memory footprints to analyze data significance and infer approximation levels through a search tree.…”