• Karra.Greeshma
• K.Jhansi Rani

Deep neural network (DNN) accelerators must balance throughput, power and area, and physically implementing every candidate configuration to explore that trade-off is prohibitively expensive. This work builds an ML-assisted flow that establishes actual, physically measured runtime and energy for two implemented MAERI configurations, MAERI-64 and MAERI-128, and predicts these same metrics for a third, considerably larger MAERI-256 configuration by combining a hardware-trend extrapolation from MAERI-64/128 with independent architectural simulation, avoiding the need for repeated synthesis, physical design or simulation on every new query. MAERI-64 and MAERI-128 were synthesized at 45 nm (Cadence Genus) and physically implemented at 90 nm (Cadence Innovus), giving actual ResNet-50 runtimes of 239.96 ms and 120.54 ms and energies of 40.96 mJ and 41.10 mJ. Cycle-level simulation across five CNN workloads (ResNet-50, ResNet-34, AlexNet, GoogLeNet, MobileNetV1) produced a 355-sample, 26-feature dataset used to train polynomial regression, ElasticNet and gradient boosting models, later extended with XGBoost and ExtraTrees. The extended models reach mean errors as low as 2.52% (runtime) and 5.49% (energy) for ResNet-34, though irregular workloads like GoogLeNet and MobileNetV1 remain harder to fit (14–24% error). For MAERI-256, frequency (0.25 GHz) and power (681.61 mW) were linearly extrapolated from the 64/128 trend and combined with real simulated cycle counts, predicting a ResNet-50 runtime of 63.73 ms and energy of 43.44 mJ. MAERI-128 roughly halves ResNet-50 runtime relative to MAERI-64 at the cost of higher power, and this trend extrapolates consistently to MAERI-256, confirming that the proposed ML-assisted flow can reliably guide design-space exploration for larger, unimplemented MAERI configurations without the time and cost of a full physical-design run.

MAERI, deep neural network (DNN) accelerator, application-specific integrated circuit (ASIC), synthesis, Cadence Genus,Cadence Innovus, design space exploration (DSE), machine learning, polynomial regression, XGBoost,ExtraTrees, MAERI-256 extrapolation, runtime and energy prediction

"Machine Learning Assisted Design Space Exploration and Performance Prediction of MAERI DNN Accelerators", International Journal of Emerging Technologies and Innovative Research (www.jetir.org), ISSN:2349-5162, Vol.13, Issue 10, page no.b371-b382, October-2026, Available :http://www.jetir.org/papers/JETIR2610138.pdf

"Machine Learning Assisted Design Space Exploration and Performance Prediction of MAERI DNN Accelerators", International Journal of Emerging Technologies and Innovative Research (www.jetir.org | UGC and issn Approved), ISSN:2349-5162, Vol.13, Issue 10, page no. ppb371-b382, October-2026, Available at : http://www.jetir.org/papers/JETIR2610138.pdf