Megatron-LM · 24 GPUs

3D Parallelism

Interactive visualization of how Tensor, Pipeline, and Data parallelism combine across 24 GPUs — hover to explore communication groups.

TP=4
Tensor Parallel
PP=3
Pipeline Stages
DP=2
Data Replicas
24
Total GPUs

GPU Communication Map

Click a parallelism type below, then hover over any GPU to see its communication group light up.