| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 58.179527282714844, | |
| "base_token_generation_latency_async": 58.61021466553211, | |
| "base_token_generation_throughput_sync": 0.017188176781510228, | |
| "base_token_generation_throughput_async": 0.017061872332436393, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 55.57422065734863, | |
| "base_inference_latency_async": 54.24995422363281, | |
| "base_inference_throughput_sync": 0.017993954538123948, | |
| "base_inference_throughput_async": 0.018433195277506274, | |
| "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": 166.82664489746094, | |
| "smashed_token_generation_latency_async": 167.07237381488085, | |
| "smashed_token_generation_throughput_sync": 0.00599424630648566, | |
| "smashed_token_generation_throughput_async": 0.005985430009559915, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 173.52499237060547, | |
| "smashed_inference_latency_async": 141.24810695648193, | |
| "smashed_inference_throughput_sync": 0.005762858631131664, | |
| "smashed_inference_throughput_async": 0.007079740900938917, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |