| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 52.916421508789064, | |
| "base_token_generation_latency_async": 52.918823808431625, | |
| "base_token_generation_throughput_sync": 0.01889772534663756, | |
| "base_token_generation_throughput_async": 0.018896867466670124, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 53.00623397827148, | |
| "base_inference_latency_async": 51.94547176361084, | |
| "base_inference_throughput_sync": 0.018865705501921223, | |
| "base_inference_throughput_async": 0.01925095616708839, | |
| "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": 39.14347877502441, | |
| "smashed_token_generation_latency_async": 38.77364415675402, | |
| "smashed_token_generation_throughput_sync": 0.025547039540033224, | |
| "smashed_token_generation_throughput_async": 0.025790714846332264, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 52.293735122680665, | |
| "smashed_inference_latency_async": 39.81616497039795, | |
| "smashed_inference_throughput_sync": 0.019122749554110228, | |
| "smashed_inference_throughput_async": 0.025115427383412443, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |