Gemini 1.0 Ultra (Google DeepMind) and Llama 3.1-405B (Meta AI) are both frontier AI models. Gemini 1.0 Ultra was published in December 2023 and Llama 3.1-405B in July 2024.
Gemini 1.0 Ultra was trained on 5 x 10^25 FLOP, about 1.3x the compute of Llama 3.1-405B at 3.8 x 10^25 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 | Llama 3.1-405B |
|---|---|---|
| Organization | Google DeepMind | Meta AI |
| Published | Dec 6, 2023 | Jul 23, 2024 |
| Training compute | 5 x 10^25 FLOP | 3.8 x 10^25 FLOP |
| Parameters | — | 405B |
| Dataset size | — | 15.6T |
| Training hardware | Google TPU v4 | NVIDIA H100 SXM5 80GB |
| Chips used | 57,000 | 16,384 |
| Training time | 2.4K h | 2.1K h |
| Training cost (2023 USD) | $31M | $53M |
| Training power draw | 38.4 MW | 22.6 MW |
| Accessibility | API access | Open weights (restricted use) |
| Open weights | No | Yes |
| Country | United States | United States |
