A new preprint estimates that Sigries could have a mean time to failure, or MTTF, of about one second under a modeled attack, compared with 13 years for PARA. The calculation uses a 4.2% victim-row failure probability for each vulnerability window and assumes that a window is open 70% of the time. Relative to PARA, the analysis reports an eight-orders-of-magnitude reduction.
The comparison combines analytical security accounting with DRAMSim3 performance simulation. Sigries’s production parameters are undisclosed; for the calculation, the analysis estimates or infers them and assumes a 32-entry TinyMG per sub-bank for a target TRHD of 3K.
The stated objective is to design a secure mitigation that avoids Sigries’s complexity. The paper evaluates FiRM-P and FiRM-D as alternatives within that design problem.
The exposure sits at a mode change
The reported weakness appears at a change between Sigries’s heavy and lite modes. At the PARA design point, the combined unmitigated activation budget at that heavy-to-lite transition reaches 1.5 × TRHD. TRHD is the threshold used in the security definition: between consecutive mitigation or refresh events, the attacker must get no more than TRHD activations.
That matters because the threat model gives the attacker knowledge of the mitigation algorithm and its parameters. The analysis therefore evaluates the design against an attacker who can plan around the stated rule.
Under the modeled Round-Robin Attack, exposure is measured across 6,144 sub-banks. A vulnerability window is open 70% of the time, compared with a 1/8,760 duty cycle for an individual sub-bank under the target. The system-level fraction is the figure used in the MTTF estimate.
A simpler tracker
FiRM is built around a shared activation budget. The design partitions TRHD between a filter and a fallback, then replaces the CAM tracker with an untagged, direct-mapped SRAM counter table. A CAM is a lookup structure that searches stored entries by matching their contents, while direct mapping sends each counter to a fixed location. This is the paper’s route to avoiding Sigries’s lookup complexity.
For a 3K activation budget, FiRM-P uses p1 = 1/25 during entry and bridge, p2 = 1/150 in steady state, and p3 = 1/60 at exit. Its security analysis reports no window in which the guarantee is weaker than the steady-state guarantee.
At equal entry count, FiRM-P uses 352 bytes per bank, compared with 864 bytes for Sigries, or 0.4 times Sigries’s storage. It uses direct-mapped lookup instead of Sigries’s 32-way CAM. The paper reports no hardware-area or power measurement, so this is a storage and lookup comparison from the design analysis.
Performance stays flat in benign tests
The performance results are based on DRAMSim3, with weighted speedup normalized to a no-mitigation baseline. On benign workloads, PARA averaged a 4.4% slowdown, while Sigries and FiRM-P recorded 0% across the tested workloads.
The workload set contained 12 SPEC2017 high-MPKI benchmarks, six GAP benchmarks and four STREAM benchmarks. Each application ran in 8-core rate mode for 250 million instructions. The results are therefore tied to the selected benchmark suite, rather than a claim about every workload.
When the simulations included attacks, Sigries and FiRM-P showed a 4.3% average slowdown and a 7.7% worst-case slowdown. Those figures are simulator results for the reported attack scenarios, not measurements of physical bit flips or field failures.
A deterministic alternative
FiRM-D offers a different trade-off. It is reported as deterministic and requires no random-number generator, uses 544 bytes per bank—0.63 times Sigries’s storage—and produces 0% benign-workload slowdown. In the co-running-attacker comparison, its slowdown is reported as twice the Sigries/FiRM-P average.
Taken together, the two FiRM variants are presented as lower-storage alternatives with different implementation choices: FiRM-P retains the reported security guarantee, while FiRM-D removes RNG dependence but carries the reported co-running-attacker cost.
What the preprint cannot settle
The central Sigries estimate depends on assumptions that remain unverified in production. The analysis uses inferred parameters, a 32-entry TinyMG per sub-bank and a target TRHD of 3K because deployed Sigries parameters are undisclosed. It also models the Round-Robin Attack rather than reporting that attack on a deployed system.
The document is an arXiv preprint, version v1, dated 26 Aug 2026. Its reported security, failure-probability, MTTF and performance numbers are analytical or simulator-derived results for the stated setup. The findings support design-level comparisons; they do not establish how the defenses would perform across other hardware, workloads or attack conditions.
Paper data and sources
Original title: From Fleet to Lab: Revisiting the Security and Complexity of Industrial Rowhammer Mitigation
Authors: Hritvik Taneja, Moinuddin Qureshi
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text