| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 55.0278823852539, | |
| "base_token_generation_latency_async": 55.61535004526377, | |
| "base_token_generation_throughput_sync": 0.01817260553475296, | |
| "base_token_generation_throughput_async": 0.017980647414538038, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 53.40753936767578, | |
| "base_inference_latency_async": 52.58936882019043, | |
| "base_inference_throughput_sync": 0.018723948188581724, | |
| "base_inference_throughput_async": 0.019015250086364868, | |
| "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": 40.9294548034668, | |
| "smashed_token_generation_latency_async": 40.80178290605545, | |
| "smashed_token_generation_throughput_sync": 0.024432282443090307, | |
| "smashed_token_generation_throughput_async": 0.02450873292234464, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 51.92407073974609, | |
| "smashed_inference_latency_async": 40.842366218566895, | |
| "smashed_inference_throughput_sync": 0.01925889064076277, | |
| "smashed_inference_throughput_async": 0.024484379642661376, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |