At 128 or more GPUs, FSDP training becomes communication-bound, making tensor and pipeline parallelism preferable, and extra GPUs yield diminishing throughput per watt.
Fully sharded data parallel: faster ai training with fewer gpus — engineering.fb.com.https://engineering.fb.com/2021/07/15/open-source/fsdp/,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2024 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Hardware Scaling Trends and Diminishing Returns in Large-Scale Distributed Training
At 128 or more GPUs, FSDP training becomes communication-bound, making tensor and pipeline parallelism preferable, and extra GPUs yield diminishing throughput per watt.