| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_perplexity": Infinity, | |
| "base_token_generation_latency_sync": 29.71763496398926, | |
| "base_token_generation_latency_async": 29.542125388979912, | |
| "base_token_generation_throughput_sync": 0.033650053283572644, | |
| "base_token_generation_throughput_async": 0.033849968031515756, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 127.26333312988281, | |
| "base_inference_latency_async": 30.81209659576416, | |
| "base_inference_throughput_sync": 0.007857722844485118, | |
| "base_inference_throughput_async": 0.032454785960182704, | |
| "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_perplexity": Infinity, | |
| "smashed_token_generation_latency_sync": 36.59109954833984, | |
| "smashed_token_generation_latency_async": 35.64547896385193, | |
| "smashed_token_generation_throughput_sync": 0.027329050297570805, | |
| "smashed_token_generation_throughput_async": 0.02805404862182101, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 181.29192962646485, | |
| "smashed_inference_latency_async": 64.09378051757812, | |
| "smashed_inference_throughput_sync": 0.005515965338669002, | |
| "smashed_inference_throughput_async": 0.015602137866181004, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |