| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_perplexity": Infinity, | |
| "base_token_generation_latency_sync": 19.835946083068848, | |
| "base_token_generation_latency_async": 20.04348710179329, | |
| "base_token_generation_throughput_sync": 0.0504135268271151, | |
| "base_token_generation_throughput_async": 0.049891518123636784, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 41.10581817626953, | |
| "base_inference_latency_async": 16.791272163391113, | |
| "base_inference_throughput_sync": 0.024327456412905117, | |
| "base_inference_throughput_async": 0.05955474905470433, | |
| "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": 23.63898696899414, | |
| "smashed_token_generation_latency_async": 23.484380170702934, | |
| "smashed_token_generation_throughput_sync": 0.042302997218605046, | |
| "smashed_token_generation_throughput_async": 0.04258149428391186, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 52.26373176574707, | |
| "smashed_inference_latency_async": 23.98686408996582, | |
| "smashed_inference_throughput_sync": 0.019133727466728395, | |
| "smashed_inference_throughput_async": 0.04168948455493687, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |