Résumé
Modern computer memories have shown to have reliability issues. The main memory is the target of a security threat called Rowhammer, which causes bit flips in adjacent victim cells of repeatedly activated aggressor rows [1]. This issue is becoming more important as DRAM technology scales down, with the required aggressor activations to corrupt a victim going from 130k for DDR3 [1] to around 10k for the most recent LPDDR4 memories [2]. Numerous countermeasures have been proposed, implemented either in software [3], [4] or in hardware [1], [5]-[10]. Among the hardware-based proposals, some rely on probability, randomly refreshing neighbors of activated rows [1], [5], [6], while others rely on row activation counters to detect aggressor rows before acting to prevent the corruption [7]-[9]. Counter-based hardware mitigation proposals offer the lowest performance overhead, as the mechanism only acts when an aggressor is detected and does not disturb the system for harmless applications. However, they require a lot of counters to track row activations. Considering the unrealistic amount of counters needed to track every rows, those mitigation exploit different most-frequent-elements detection algorithms to reduce the number of counters needed while keeping a complete protection and a minimal false positive rate. Most of them offer a bank-level attack detection, with a separate set of counters for each bank. In this talk, We will show you that by changing the counting granularity from bank-level to rank-level, we can further reduce the total required number for counters from 20% for DDR3 to 70% for DDR5, thus reducing the silicon area and energy overheads of such mitigations.