| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 55.25741424560547, | |
| "base_token_generation_latency_async": 54.7143429517746, | |
| "base_token_generation_throughput_sync": 0.018097118977649745, | |
| "base_token_generation_throughput_async": 0.018276743282495476, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 53.19772186279297, | |
| "base_inference_latency_async": 52.31618881225586, | |
| "base_inference_throughput_sync": 0.018797797442890316, | |
| "base_inference_throughput_async": 0.019114542223032403, | |
| "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": 166.15823211669922, | |
| "smashed_token_generation_latency_async": 166.91399328410625, | |
| "smashed_token_generation_throughput_sync": 0.006018359651886896, | |
| "smashed_token_generation_throughput_async": 0.005991109435012368, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 173.4734848022461, | |
| "smashed_inference_latency_async": 141.62616729736328, | |
| "smashed_inference_throughput_sync": 0.005764569733178336, | |
| "smashed_inference_throughput_async": 0.007060842068121245, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |