| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 74.4272232055664, | |
| "base_token_generation_latency_async": 74.3584843352437, | |
| "base_token_generation_throughput_sync": 0.013435943958812238, | |
| "base_token_generation_throughput_async": 0.013448364486445426, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 73.60645141601563, | |
| "base_inference_latency_async": 72.3996639251709, | |
| "base_inference_throughput_sync": 0.013585765659970607, | |
| "base_inference_throughput_async": 0.013812218811313212, | |
| "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": 162.6752166748047, | |
| "smashed_token_generation_latency_async": 164.0027405694127, | |
| "smashed_token_generation_throughput_sync": 0.006147217876457767, | |
| "smashed_token_generation_throughput_async": 0.006097459082256975, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 170.59573669433593, | |
| "smashed_inference_latency_async": 131.87716007232666, | |
| "smashed_inference_throughput_sync": 0.00586181119984109, | |
| "smashed_inference_throughput_async": 0.0075828141844392185, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |