| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 59.103076171875, | |
| "base_token_generation_latency_async": 56.22761659324169, | |
| "base_token_generation_throughput_sync": 0.016919593103613507, | |
| "base_token_generation_throughput_async": 0.017784854855828897, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 53.950770950317384, | |
| "base_inference_latency_async": 53.47094535827637, | |
| "base_inference_throughput_sync": 0.018535416313529385, | |
| "base_inference_throughput_async": 0.018701745280537058, | |
| "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": 165.9166275024414, | |
| "smashed_token_generation_latency_async": 166.53981637209654, | |
| "smashed_token_generation_throughput_sync": 0.0060271234719093195, | |
| "smashed_token_generation_throughput_async": 0.0060045700889072694, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 172.5318145751953, | |
| "smashed_inference_latency_async": 140.17925262451172, | |
| "smashed_inference_throughput_sync": 0.005796032473559626, | |
| "smashed_inference_throughput_async": 0.00713372329554809, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |