| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 57.667279434204104, | |
| "base_token_generation_latency_async": 58.03066436201334, | |
| "base_token_generation_throughput_sync": 0.017340856198027464, | |
| "base_token_generation_throughput_async": 0.017232268680601154, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 56.35584030151367, | |
| "base_inference_latency_async": 55.629873275756836, | |
| "base_inference_throughput_sync": 0.017744389838742956, | |
| "base_inference_throughput_async": 0.017975953226479737, | |
| "base_inference_CO2_emissions": null, | |
| "base_inference_energy_consumption": null, | |
| "smashed_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "smashed_current_gpu_total_memory": 40339.3125, | |
| "smashed_token_generation_latency_sync": 41.73652305603027, | |
| "smashed_token_generation_latency_async": 41.733165830373764, | |
| "smashed_token_generation_throughput_sync": 0.02395983006676249, | |
| "smashed_token_generation_throughput_async": 0.023961757515941703, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 51.78419189453125, | |
| "smashed_inference_latency_async": 42.34201908111572, | |
| "smashed_inference_throughput_sync": 0.01931091252783664, | |
| "smashed_inference_throughput_async": 0.023617201581348155, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |