| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_perplexity": 21.553375244140625, | |
| "base_token_generation_latency_sync": 24.06592712402344, | |
| "base_token_generation_latency_async": 23.613459430634975, | |
| "base_token_generation_throughput_sync": 0.04155252340150924, | |
| "base_token_generation_throughput_async": 0.04234872924645034, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 40.53800964355469, | |
| "base_inference_latency_async": 22.180819511413574, | |
| "base_inference_throughput_sync": 0.024668206673017907, | |
| "base_inference_throughput_async": 0.045083996986018955, | |
| "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": 23.580896377563477, | |
| "smashed_token_generation_latency_sync": 36.70088005065918, | |
| "smashed_token_generation_latency_async": 36.72023508697748, | |
| "smashed_token_generation_throughput_sync": 0.02724730302433277, | |
| "smashed_token_generation_throughput_async": 0.027232941119013736, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 66.75742721557617, | |
| "smashed_inference_latency_async": 44.14482116699219, | |
| "smashed_inference_throughput_sync": 0.014979606640183328, | |
| "smashed_inference_throughput_async": 0.022652713807972485, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |