SRAM budgeting for on-device training: Accounting, feasible suffix selection, and a configuration-dependent crossover analysis

On-device training enables neural networks to adapt to new data directly on resource-constrained microcontroller units (MCUs) without requiring cloud connectivity. However, the severe memory limitations of MCUs, typically ranging from 256KB to 1MB of SRAM, pose significant challenges for implementing backpropagation-based training. Existing approaches either ignore memory constraints entirely or rely on ad-hoc heuristics for layer selection, leading to suboptimal memory utilization or unexpected out-of-memory failures during deployment. In this paper, we present a comprehensive SRAM budgeting framework that provides precise memory cost modeling for on-device training. Our framework decomposes training memory into four core components: forward activations, backward gradients, optimizer states, and optional buffers for quantized training scenarios. Given a target memory budget, our automated suffix selection algorithm determines the maximum number of trainable layers while guaranteeing the memory constraint is satisfied. Through extensive experiments on CIFAR-10 and Visual Wake Words benchmarks using ResNet-8 architecture, we demonstrate that our framework predicts training memory within 0.5% of measured usage on average for multi-block configurations under an explicitly stated mixed-precision deployment configuration (FP16 gradient and activation buffers, FP32 momentum). On VWW, our framework enables a 21.6 percentage point accuracy improvement (61.3% to 82.9%) when the budget increases from 256KB to 512KB, and 26.3 points at 768KB, by automatically selecting optimal suffix configurations. Our analysis reveals a critical crossover regime: below approximately 350KB under the deployment configuration (472KB under pure FP32), sparse update methods outperform contiguous suffix training, while above this threshold suffix-based training dominates. We release our framework as an open-source tool to enable reproducible on-device training research.