runux-tpu-v5e-benchmarks
收藏资源简介:
RunuX-AI TPU v5e推理基准数据集是一个综合性基准测试集合,用于评估和比较大型语言模型(LLM)在Google TPU v5e硬件上的推理性能。该数据集旨在提供可独立验证的结果,以促进节能、高效AI推理的协作与研究。核心内容包括对五种前沿LLM(Qwen 2.5 0.5B、DeepSeek R1 1.5B、Mistral 7B v0.3、Gemma 2 9B和Gemma 2 27B)在五种不同推理框架(PyTorch、TF/JAX、JetStream、vLLM以及RunuX-AI)下的性能对比,评估指标全面覆盖推理效率的多个维度:吞吐量(每秒生成的令牌数)、能源效率(每令牌消耗的焦耳数)、硬件利用率(MXU利用率)、环境影响(基于不同地区电网碳强度的每千令牌CO₂排放量)以及经济成本(每百万令牌的云服务成本)。数据生成严格遵循MLPerf推理基准的方法论,在单芯片TPU v5e(v5litepod-1)硬件上,使用BF16精度,以固定的输入令牌(512)和解码令牌(128)长度进行测量,涵盖不同批次大小(1, 8, 32),并报告多次迭代的中值结果。能耗和碳排放模型基于保守的上限估算,为评估AI计算的可持续性提供量化依据。该数据集适用于文本生成任务的性能基准测试、推理框架的对比研究、AI系统能效与碳足迹分析,以及绿色计算和可持续AI领域的研究与开发。数据集以JSON格式(benchmark_results.json)提供完整的基准测试结果,并附有Apache-2.0许可的复现脚本,确保透明性和可重复性。
The RunuX-AI TPU v5e Inference Benchmark Dataset is a comprehensive benchmark suite for evaluating and comparing the inference performance of large language models (LLMs) on Google TPU v5e hardware. It aims to provide independently verifiable results to foster collaboration and research in energy-efficient and high-performance AI inference. The core content includes performance comparisons of five cutting-edge LLMs (Qwen 2.5 0.5B, DeepSeek R1 1.5B, Mistral 7B v0.3, Gemma 2 9B, and Gemma 2 27B) across five different inference frameworks (PyTorch, TF/JAX, JetStream, vLLM, and RunuX-AI). Evaluation metrics comprehensively cover multiple dimensions of inference efficiency: throughput (tokens generated per second), energy efficiency (joules per token), hardware utilization (MXU utilization), environmental impact (CO₂ emissions per thousand tokens based on regional grid carbon intensity), and economic cost (cloud service cost per million tokens). The dataset is generated strictly following the MLPerf inference benchmark methodology, measured on a single-chip TPU v5e (v5litepod-1) hardware with BF16 precision, using fixed input token (512) and decoding token (128) lengths. The benchmark covers different batch sizes (1, 8, 32) and reports median results from multiple iterations. Energy consumption and carbon emission models are based on conservative upper-bound estimates, providing quantitative basis for assessing the sustainability of AI computing. This dataset is suitable for performance benchmarking of text generation tasks, comparative studies of inference frameworks, analysis of AI system energy efficiency and carbon footprint, as well as research and development in green computing and sustainable AI. The dataset provides complete benchmark results in JSON format (benchmark_results.json) and includes reproducible scripts under the Apache-2.0 license, ensuring transparency and repeatability.




