| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_perplexity": Infinity, | |
| "base_token_generation_latency_sync": 34.3837194442749, | |
| "base_token_generation_latency_async": 32.92566556483507, | |
| "base_token_generation_throughput_sync": 0.029083531862243196, | |
| "base_token_generation_throughput_async": 0.03037144376112505, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 128.52879257202147, | |
| "base_inference_latency_async": 31.447863578796387, | |
| "base_inference_throughput_sync": 0.007780357848142447, | |
| "base_inference_throughput_async": 0.031798662490835995, | |
| "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": 55.48102798461914, | |
| "smashed_token_generation_latency_async": 55.78669644892216, | |
| "smashed_token_generation_throughput_sync": 0.018024179369517584, | |
| "smashed_token_generation_throughput_async": 0.017925420640664602, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 214.5728515625, | |
| "smashed_inference_latency_async": 109.41667556762695, | |
| "smashed_inference_throughput_sync": 0.004660421822789281, | |
| "smashed_inference_throughput_async": 0.00913937473252815, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |