YuvrajSingh9886/bonsai-jetson-benchmark-maxn
收藏资源简介:
Bonsai Jetson Benchmark数据集是一个性能基准测试数据集,专注于在NVIDIA Jetson Orin Nano Super 8GB边缘AI平台上评估Bonsai系列语言模型(包括1.7B、4B、8B参数版本及其Ternary变体)的推理效率。测试在MAXN_SUPER电源模式下进行,使用llama.cpp后端和CUDA加速,覆盖多种量化配置(如Q1_0、Q2_0)、输入token长度(256、512、1024、2048)和输出token长度(128、256、512)。数据集的核心指标是每焦耳输出token数(Tok/J),用于衡量能效,同时记录时间到首个token(TTFT)、token到token延迟(T2T)、吞吐量(Tok/s)、请求延迟和功率消耗等详细性能数据。数据以表格形式呈现,包括完整结果表和每个模型的最佳配置摘要,旨在为边缘设备上的LLM推理优化提供参考。
The Bonsai Jetson Benchmark dataset is a performance benchmarking dataset focused on evaluating the inference efficiency of the Bonsai series language models (including 1.7B, 4B, 8B parameter versions and their Ternary variants) on the NVIDIA Jetson Orin Nano Super 8GB edge AI platform. Tests are conducted in MAXN_SUPER power mode using the llama.cpp backend with CUDA acceleration, covering various quantization configurations (e.g., Q1_0, Q2_0), input token lengths (256, 512, 1024, 2048), and output token lengths (128, 256, 512). The core metric of the dataset is tokens per joule (Tok/J) to measure energy efficiency, along with detailed performance data such as time to first token (TTFT), token-to-token latency (T2T), throughput (Tok/s), request latency, and power consumption. The data is presented in tabular form, including a full results table and a summary of optimal configurations per model, aiming to provide references for LLM inference optimization on edge devices.




