| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 70.8256935119629, | |
| "base_token_generation_latency_async": 70.77690288424492, | |
| "base_token_generation_throughput_sync": 0.014119169900272052, | |
| "base_token_generation_throughput_async": 0.014128903063694272, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 70.20144691467286, | |
| "base_inference_latency_async": 70.00713348388672, | |
| "base_inference_throughput_sync": 0.014244720642516976, | |
| "base_inference_throughput_async": 0.014284258621018476, | |
| "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": 120.27656784057618, | |
| "smashed_token_generation_latency_async": 115.46138767153025, | |
| "smashed_token_generation_throughput_sync": 0.008314171396422593, | |
| "smashed_token_generation_throughput_async": 0.008660904049108131, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 117.87663421630859, | |
| "smashed_inference_latency_async": 117.6546573638916, | |
| "smashed_inference_throughput_sync": 0.008483445482206064, | |
| "smashed_inference_throughput_async": 0.008499451040914778, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |