| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 77.12576599121094, | |
| "base_token_generation_latency_async": 76.90115496516228, | |
| "base_token_generation_throughput_sync": 0.01296583556932138, | |
| "base_token_generation_throughput_async": 0.013003705867005761, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 74.73766403198242, | |
| "base_inference_latency_async": 73.6828088760376, | |
| "base_inference_throughput_sync": 0.013380134540625605, | |
| "base_inference_throughput_async": 0.013571686737436665, | |
| "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": 193.7925048828125, | |
| "smashed_token_generation_latency_async": 193.7643963843584, | |
| "smashed_token_generation_throughput_sync": 0.005160158286847605, | |
| "smashed_token_generation_throughput_async": 0.005160906846974932, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 192.28262329101562, | |
| "smashed_inference_latency_async": 188.58113288879395, | |
| "smashed_inference_throughput_sync": 0.005200677954588343, | |
| "smashed_inference_throughput_async": 0.00530275741099561, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |