Data parallelism
Two-GPU data parallelism
Forward sends different batch shards through identical replicas. Backward averages local gradients using AllReduce.
global batch
B₀
B₁
GPU 0
replica
layer 1
layer 2
layer 3
GPU 1
replica
layer 1
layer 2
layer 3
Loss
L₀
Loss
L₁
no cross-GPU communication
local gradients
g₀
g₁
ALL-REDUCE
forward
backward