| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 32.687651824951175, | |
| "base_token_generation_latency_async": 32.9969834536314, | |
| "base_token_generation_throughput_sync": 0.030592592130973414, | |
| "base_token_generation_throughput_async": 0.030305800571292753, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 31.924736404418944, | |
| "base_inference_latency_async": 30.68556785583496, | |
| "base_inference_throughput_sync": 0.03132367288274877, | |
| "base_inference_throughput_async": 0.032588609886515324, | |
| "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": 85.15609741210938, | |
| "smashed_token_generation_latency_async": 85.28711386024952, | |
| "smashed_token_generation_throughput_sync": 0.011743140308092582, | |
| "smashed_token_generation_throughput_async": 0.011725100718481206, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 84.63370208740234, | |
| "smashed_inference_latency_async": 84.725022315979, | |
| "smashed_inference_throughput_sync": 0.011815623981180532, | |
| "smashed_inference_throughput_async": 0.011802888599669352, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |