Do you have a calculation that shows where that $2M number comes from, EXACTLY?

https://arxiv.org/pdf/2302.13971.pdf table 15. 1770394 A100-80GB hours to train the entire model suite at the going rate for cloud 8xA100-80GBs (~$12/hr if you could actually get capacity) is ~$2.6M, under extremely optimistic assumptions. YMMV on bulk pricing ;) "the more you buy the more you save"

Hmmm… the values in the 7B model seem feasible. An order of magnitude lower GPU hours, plus presumably the lower parameter count means it probably could fit on a 24GB Radeon RX 7900 XTX, which has higher single precision flops than the A100 and costs $1000 instead of $15,000.

An order of magnitude lower GPU-hour time, plus if you train it for 210 days instead of 21 days, means you could do a 7B model with 20 consumer GPUs which are $1000 apiece. $20k, not counting mainboard, etc. Really not bad. Might even be doable as a volunteer project.

I'm not aware of any efficient transformer training code for AMD cards.

Also most training is done using bfloat, not single precision (which is usually only used for accumulators)

Sure, you would need to rewrite the training code for AMD's ecosystem. If you're using mixed precision training, I suppose you're right about BF16. That puts the relative performance of A100 about 2.5x that of the Radeon RX 7900 XT. May be better to go with the Nvidia GeForce RTX 4090 with a $1600 retail.

It all works with pytorch and huggingface's transformers library out of the box with Rocm.

You would need to compile a few components from source for Navi 31 if you were to try it today, so out-of-the-box is perhaps an overstatement, but it's certainly doable.

Page 4 https://arxiv.org/abs/2302.13971

When training a 65B-parameter model, our code processes around 380 tokens/sec/GPU on 2048 A100 GPU with 80GB of RAM. This means that training over our dataset containing 1.4T tokens takes approximately 21 days.

At $4/GPU-hour per A100 80GB GPU, that's $4 * 2,048 * 21 * 24 = $4,128,768.

Hmmm… so a 7 billion parameter model could probably be trained on consumer GPUs for one or two orders of magnitude lower cost, particularly if you didn’t go well beyond Chinchilla-optimal training time.

The whole point of Llama is to go beyond Chinchilla optional:

> The objective of the scaling laws from Hoffmann et al. (2022) is to determine how to best scale the dataset and model sizes for a particular training compute budget. However, this objective disregards the inference budget, which becomes critical when serving a language model at scale. In this context, given a target level of performance, the preferred model is not the fastest to train but the fastest at inference, and although it may be cheaper to train a large model to reach a certain level of performance, a smaller one trained longer will ultimately be cheaper at inference.