# Geo-distributed training works on Akash

- **Speaker:** Greg Osuri
- **Attribution:** attributed
- **Said:** 2023-08-31
- **Theme:** Decentralized AI — https://www.gregosuri.com/predictions/decentralized-ai/
- **Source:** Akash Mainnet 6 Livestream (Akash Network)
- **Watch at:** 01:20:48 — https://www.youtube.com/watch?v=PqKYTTQUwCM&t=4848s
- **Canonical:** https://www.gregosuri.com/predictions/decentralized-ai/2023-08-31-geo-distributed-training-works-on-akash-lkdu/
- **Cite as:** Greg Osuri, "Geo-distributed training works on Akash," Akash Mainnet 6 Livestream (Akash Network), 31 Aug 2023, 01:20:48. https://www.youtube.com/watch?v=PqKYTTQUwCM&t=4848s

## Quote — verbatim

> Yes, it is possible to distribute the training across different GPUs. In fact there is a proposal on Akash discussions talking about how they plan to use about 24,000 A100 GPU hours to train across a distributed cluster... so yes, it is possible to train on geographically distributed clusters on Akash.

## Context — site annotation, not spoken

Answering a chat question on multi-GPU training; Greg caveats it depends on batch sizes, parallelization strategy, gradient compression, and inter-cluster latency.