ResNet-50

ResNet-50 is a convolutional neural network for image classification, introduced by researchers at Microsoft Research. Its memorable idea is the residual or skip connection: a block learns a change to its input while an identity path carries the input forward, helping very deep networks train. ’50’ refers to the conventional layer count, not classes or chips. Although residual connections now appear widely, this entry is the classic image classifier rather than a segmentation model. TT-Metal lists 224x224 images across n150 and n300 Wormhole, p150 Blackhole, Wormhole QuietBox at DP=8, and Wormhole Galaxy at DP=32, with batches from 16 to 512. Those rows collapse into one architecture and show broad single- to many-device scaling.