Dynamic RAM Refresh & Performance

2 posts · Nov 2018 – Jun 2022

What this trail covers

Cloudflare's exploration of low-level hardware behaviors and their impact on system performance has evolved to include the analysis of dynamic voltage and frequency scaling (DVFS) and its data-dependent nature. This vulnerability, Hertzbleed, demonstrates how CPU power consumption variations, influenced by data characteristics like Hamming weight and Hamming distance, can lead to data-dependent CPU frequencies and execution times. This has implications for cryptographic software security, partic

All Cloudflare trails

Jun 2022Hertzbleed explainedThis post explains the Hertzbleed vulnerability, a new type of side-channel attack that exploits data-dependent CPU frequency scaling. It details how Dynamic Voltage and Frequency Scaling (DVFS) on modern x86 processors can be influenced by the data being processed, leading to variations in CPU frequency and execution time. The post explains the concepts of P-states, Thermal Design Point (TDP), and how data-dependent power consumption (due to Hamming weight and Hamming distance) can cause different P-state distributions, ultimately making program execution time data-dependent. It highlights the practical threat to cryptographic software, especially SIKE, and notes that conventional cryptography is not currently known to be targeted.
Nov 2018Every 7.8μs your computer’s memory has a hiccupThis post details an experiment to measure the impact of Dynamic RAM refresh cycles on system performance. It describes the experimental setup, including a tight loop in C with cache flushing and timing measurements, and the challenges of noisy data. The post then explains the application of Fast Fourier Transform (FFT) to analyze the timing data, revealing a dominant frequency corresponding to the DRAM refresh interval (approximately 7.8μs). It also discusses the presence of harmonics and provides a command-line tool for others to replicate the experiment.
Advertisement