Hacker News Reader

If I just needed a GPU for learning purposes, is 2xGPUs necessary? Would a single 24GB GPU significantly bottleneck training with any publicly available datasets? Just need something faster than my laptop, but if it takes twice as long, not really an issue.

View on HN

Training what on what - Resnet50 on imagenet? Yeah sure a single 4090 is fine. Will take a bit.

A 1.5B parameter LLM? That’s a few weeks with 64 V100s - on a small dataset.

Training something Lllama 7b class? (Not using lora)? Weeks with the same number of A100s.

With lora? Back to a single 4090 - depending on your dataset. It still might take weeks to go through 2000 examples for finetuning with a large context size.

Reply on news.ycombinator.com