Simultrain - Solution

[ T_\textseq = T_\textsend + T_\textforward + T_\textbackward + T_\textrecv ]

where ( T_\textsend ) and ( T_\textrecv ) depend on bandwidth, and ( T_\textforward, T_\textbackward ) on model size. For large models (e.g., ResNet-50), ( T_\textsend \gg T_\textforward ) on typical 4G/5G networks. simultrain solution

SimulTrain sends activations (lower dimension than raw data but higher than gradients). However, it enables bidirectional overlap , reducing total bandwidth-time product by 65% compared to SyncSGD. | Dataset | Centralized | SyncSGD | FedAvg (5 local steps) | SimulTrain | |-------------|-------------|---------|------------------------|------------| | UCF-101 | 84.2% | 83.9% | 81.1% | 83.7% | | WISDM | 91.5% | 91.3% | 88.9% | 91.1% | However, it enables bidirectional overlap , reducing total

[ w_t+1 = w_t - \eta \nabla \ell(w_t; x_t, y_t) ] This is analogous to CPU pipelining but applied

[ \tilde\nabla_k = \nabla \ell(w^(e)_k; x_k) + \alpha \cdot (w^(c)_k - w^(e)_k) ]

In edge-cloud setting, data is at edge, compute is in cloud. The sequential round-trip time is:

of SimulTrain is that the forward pass of one batch and the backward pass of a previous batch can overlap in time, if we carefully manage parameter versions and gradients. This is analogous to CPU pipelining but applied to distributed training across heterogeneous compute nodes.