YuvrajSingh9886/bonsai-jetson-benchmark-25w
收藏资源简介:
这是一个在NVIDIA Jetson Orin Nano Super 8GB边缘AI设备上,以25W功率模式进行的语言模型推理性能基准测试数据集。数据集包含多个Bonsai模型(如1.7B、4B、8B)及其三元量化版本,在不同量化配置(如Q1_0、Q2_0)、提示令牌数(256、512、1024、2048)和生成令牌数(128、256、512)下的性能指标,如令牌每秒(Tok/s)、功率消耗(瓦特)和关键能效指标令牌每焦耳(Tok/J)。数据集还包括内存不足(OOM)和跳过的模型记录,以及热和功率摘要,旨在评估边缘AI场景下的模型能效和推理效率。
This is a benchmark dataset for language model inference performance on the NVIDIA Jetson Orin Nano Super 8GB edge AI device, operating in 25W power mode. The dataset includes multiple Bonsai models (e.g., 1.7B, 4B, 8B) and their ternary-quantized versions, with performance metrics under various quantization configurations (e.g., Q1_0, Q2_0), prompt token counts (256, 512, 1024, 2048), and generation token counts (128, 256, 512). Metrics include tokens per second (Tok/s), power consumption (watts), and the key energy efficiency metric tokens per joule (Tok/J). The dataset also records out-of-memory (OOM) and skipped models, along with thermal and power summaries, aiming to evaluate model energy efficiency and inference performance in edge AI scenarios.




