| { | |
| "base_current_gpu_type": "NVIDIA A100-PCIE-40GB", | |
| "base_current_gpu_total_memory": 40339.3125, | |
| "base_perplexity": 9.883513450622559, | |
| "base_token_generation_latency_sync": 36.09357814788818, | |
| "base_token_generation_latency_async": 35.24684049189091, | |
| "base_token_generation_throughput_sync": 0.027705759620247277, | |
| "base_token_generation_throughput_async": 0.028371337289935697, | |
| "base_token_generation_CO2_emissions": null, | |
| "base_token_generation_energy_consumption": null, | |
| "base_inference_latency_sync": 117.57680435180664, | |
| "base_inference_latency_async": 38.5784387588501, | |
| "base_inference_throughput_sync": 0.008505078918524242, | |
| "base_inference_throughput_async": 0.02592121485918335, | |
| "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": 10.60670280456543, | |
| "smashed_token_generation_latency_sync": 42.5255615234375, | |
| "smashed_token_generation_latency_async": 41.7684581130743, | |
| "smashed_token_generation_throughput_sync": 0.023515268562623468, | |
| "smashed_token_generation_throughput_async": 0.023941511015150005, | |
| "smashed_token_generation_CO2_emissions": null, | |
| "smashed_token_generation_energy_consumption": null, | |
| "smashed_inference_latency_sync": 160.99471435546874, | |
| "smashed_inference_latency_async": 69.73233222961426, | |
| "smashed_inference_throughput_sync": 0.0062113840445224, | |
| "smashed_inference_throughput_async": 0.014340550043661315, | |
| "smashed_inference_CO2_emissions": null, | |
| "smashed_inference_energy_consumption": null | |
| } |