Program acceleration using nearest distance associative search

ALook

“…In [21], the authors propose a configurable associative memory which relaxes computation by applying VOS to the non-volatile associative memory to trade accuracy for energy savings. Work in [23] designed a novel associative memory based on non-volatile memory which supports searches for nearest absolute distance, rather than Hamming distance similarity. All existing work provide low hit rate on practical applications because their stored values are static and cannot change to adapt to a running application.…”

“…Placing a small lookup table next to each floating point unit enables approximate computational reuse [21,23]. Figure 1a) shows the pre-run and profiling stage of a static table [23]. During the pre-run stage, the user must flag the section of OpenCL code they wish to approximate.…”

“…Two lookup tables are placed alongside each floating point unit to enable parallel lookup. We implement ALOOK with content addressable memory (CAM) introduced in [23]. The lookup table identifies nearest distance matches for each incoming set of input values.…”

“…When a search provides a poor match for an operation the FPU computes the result exactly and the table updates dynamically using this value. Figure 3 shows the architecture of an enhanced FPU using a single lookup table to store highly frequent patterns [23]. The table is placed adjacent to the first pipeline stage of the FPU block and search operations occur during this stage.…”

“…Our evaluation shows that ALOOK provides 3.6× EDP (Energy Delay Product) and 32.8% performance speedup, compared to an unmodified GPU, for applications accepting less than 5% output error. The proposed ALOOK architecture improves the GPU performance by 2.0× as compared to state-of-the-art computational reuse methods [22,23] for the same level of output error.…”