A 19.4-nJ/Decision, 364-K Decisions/s, In-Memory Random Forest Multi-Class Inference Accelerator

Tree-based machine learning performed in-memory with memristive analog CAM

“…To test our system and compare results with previous benchmarks 27 , we implemented an RF for the classification of KUL Belgium traffic sign dataset 43 whose data processing is explained in Supplementary Information 4 . We mapped the RF into the analog CAM and RRAM arrays and evaluated the accuracy of inference on 200 samples 27 reaching 0.965, higher than the reference state of the art. These RF models are well matched to analog IMC implementations, showing strong resilience to variation and noise that can otherwise affect analog hardware.…”

“…For a fair comparison to the ASIC works, we evaluated our analog CAM hardware on a 65 nm technology by applying a constant field scaling procedure. We considered a maximum clock frequency of 1 GHz, as a practical case 27 , and evaluated the performance of our architecture compared with different results from literature 12 , 14 , 27 , 44 . The comparison is shown in Table 1 with analog CAM outperforming existing accelerators in throughput and energy per decision 27 .…”

“…We considered a maximum clock frequency of 1 GHz, as a practical case 27 , and evaluated the performance of our architecture compared with different results from literature 12 , 14 , 27 , 44 . The comparison is shown in Table 1 with analog CAM outperforming existing accelerators in throughput and energy per decision 27 . Moreover, the algorithm independent metrics, normalized by the number of nodes of each tree once again shows the strong improvement of accelerating RF with an analog CAM architecture compared to other works, with an energy-delay product per node 3 orders of magnitude lower the state of the art (see Supplementary Information note 2 ).…”

“…All tree-based models were trained in a Python environment with sklearn module. To match the benchmark, we trained RF with KUL Belgium traffic signs dataset 43 considering the same 8 class and training/testing set as in literature 27 , thus we used 2300 training and 200 testing images taken from the classes ‘No Overtaking’, ‘Children’, ‘Crossroads with a minor road’, ‘Priority road’, ‘Give Way’, Stop’, ‘No vehicles’, and ‘Maximum speed limit’. However, while the reference RF was trained with 64 trees and a maximum depth of 6, we optimized the hyperparameters namely maximum depth and number of trees reaching an accuracy of 96.5% when deployed to analog CAM, overcoming the given accuracy of 94%.…”

“…Other traditional CMOS accelerator approaches have been studied for these models, such as an RF IMC accelerator based on complementary-metal-oxide-semiconductor (CMOS) static random access memories (SRAM) in ref. 27 , but model inference at high throughput and low energy operation remains a challenge.…”