| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_token_generation_latency_sync": 54.78649673461914, | |
| "base_token_generation_latency_async": 54.92506679147482, | |
| "base_token_generation_throughput_sync": 0.01825267282271962, | |
| "base_token_generation_throughput_async": 0.018206623285439767, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 53.236223983764646, | |
| "base_inference_latency_async": 52.28421688079834, | |
| "base_inference_throughput_sync": 0.01878420228123182, | |
| "base_inference_throughput_async": 0.01912623081416479, | |
| "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": 167.07427520751952, | |
| "smashed_token_generation_latency_async": 167.65917092561722, | |
| "smashed_token_generation_throughput_sync": 0.005985361892235776, | |
| "smashed_token_generation_throughput_async": 0.00596448136108018, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 173.8080261230469, | |
| "smashed_inference_latency_async": 141.2287473678589, | |
| "smashed_inference_throughput_sync": 0.0057534742342223765, | |
| "smashed_inference_throughput_async": 0.0070807113894120815, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |