Nvalt: Nonvolatile Approximate Lookup Table for GPU Acceleration
Mohsen Imani, Daniel Peroni, Tajana Rosing · IEEE Embedded Systems Letters · 2017
In this letter, we design a nonvolatile approximate lookup table, called Nvalt, to significantly accelerate general public utilities (GPUs) computation. Our design stores high frequency input patterns within an approximate Nvalt to model each application's functionality. Nvalt searches for and returns the stored data best matching the input data to produce an approximate output. We define a similarity metric, appropriate for binary representation, by exploiting the analog characteristics of the nonvolatile content addressable memory. Our design controls the ratio of the application running between the approximate Nvalt and accurate GPU cores in order to tune the level of accuracy necessary for user requirements. Our evaluations on seven general GPU applications shows that, Nvalt can improve energy computation by 4.5× and performance by 5.7× on average while providing less than 10% average relative error as compared to the baseline GPU.