| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 57.67717323303223, | |
| "base_token_generation_latency_async": 56.299011036753654, | |
| "base_token_generation_throughput_sync": 0.017337881590689525, | |
| "base_token_generation_throughput_async": 0.017762301354586327, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 55.80523529052734, | |
| "base_inference_latency_async": 54.78930473327637, | |
| "base_inference_throughput_sync": 0.01791946570593073, | |
| "base_inference_throughput_async": 0.018251737357649813, | |
| "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": 164.09751892089844, | |
| "smashed_token_generation_latency_async": 164.3142594024539, | |
| "smashed_token_generation_throughput_sync": 0.006093937352471733, | |
| "smashed_token_generation_throughput_async": 0.006085899079219328, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 170.6112030029297, | |
| "smashed_inference_latency_async": 127.96695232391357, | |
| "smashed_inference_throughput_sync": 0.005861279812808238, | |
| "smashed_inference_throughput_async": 0.0078145175909853, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |