Gemini 1.0 Ultra (Google DeepMind) and PaLM (540B) (Google Research) are both frontier AI models. Gemini 1.0 Ultra was published in December 2023 and PaLM (540B) in April 2022.
Gemini 1.0 Ultra was trained on 5 x 10^25 FLOP, about 19.8x the compute of PaLM (540B) at 2.5 x 10^24 FLOP. Training compute is the closest available proxy for how much was invested in a model, though it says nothing on its own about how well that compute was spent.
These two models share no benchmark on which both have been scored, so no direct performance comparison is possible here. The specification table below is a comparison of inputs, not of results.
| Field | Gemini 1.0 Ultra | PaLM (540B) |
|---|---|---|
| Organization | Google DeepMind | Google Research |
| Published | Dec 6, 2023 | Apr 4, 2022 |
| Training compute | 5 x 10^25 FLOP | 2.5 x 10^24 FLOP |
| Parameters | — | 540.4B |
| Dataset size | — | 780B |
| Training hardware | Google TPU v4 | Google TPU v4 |
| Chips used | 57,000 | 6,144 |
| Training time | 2.4K h | 1.5K h |
| Training cost (2023 USD) | $31M | $3M |
| Training power draw | 38.4 MW | 4.2 MW |
| Accessibility | API access | Unreleased |
| Open weights | No | No |
| Country | United States | United States |
