Pipeline parallelism assigns contiguous blocks of layers to different chips and passes activations down the line, so the only communication is one activation tensor per stage boundary. That makes it the cheap way to span slow links between racks and pods.
The cost is pipeline bubbles: chips idle while the pipeline fills and drains each step. Micro-batching and interleaved forward/backward schedules shrink the bubbles, and frontier recipes like DeepSeek-V3’s 16-way pipeline stack it with expert and data parallelism.
