RowHammer has gotten worse with every DRAM shrink. Activation thresholds keep dropping, and modern multi-bank hammering patterns laugh at TRR, ECC, and refresh-based defenses. The numbers are ugly: bitflip thresholds (N_BO) are sinking into the double digits, and existing mitigations either need expensive DRAM-side changes or wreck memory throughput.
Enter ARTA — an adaptive reinforcement-learning throttling agent that lives entirely in the memory controller. No DRAM modifications, no offline training, just a Q-learning governor that watches access behavior inside the refresh window (t_REFW) and adjusts core frequency to keep hammering in check. The hardware cost: a small per-core, per-bank FIFO queue and a compact Q-table, all in SRAM.
Zero Bitflips at N_BO 64, 22,000× Reduction at N_BO 20
The results justify the approach. ARTA eliminates every bitflip at N_BO 64 — the point where even trimmed DRAM starts throwing errors. At a more aggressive N_BO of 20, it cuts bitflips by a factor of 22,000 compared to an unprotected system. That's not a marginal improvement; it's a regime change.
73.6% Performance Edge Over State-of-the-Art Without Hardware Surgery
The real trick is performance. Traditional throttling mechanisms hammer the entire memory bus when they detect any aggression, wasting bandwidth on innocent traffic. ARTA learns which banks are being pounded and applies frequency scaling only where needed. Over existing mitigations, it improves throughput by up to 73.6% — not by adding more hardware, but by being smarter about when to back off.
The adaptive RL governor is the key. Q-values are updated online from memory access patterns, so the agent adapts to workload behavior without manual tuning. That means it works on future DRAM generations without a firmware respin.
Putting an RL agent in the memory controller to fight RowHammer is exactly the kind of software-hardware co-design that scales. ARTA proves you can kill bitflips at low N_BO values without nuking performance or waiting for the next DRAM standard. Expect to see this kind of approach land in real memory controllers sooner than most think.
Source: ARTA: Adaptive Reinforcement-Learning-Based Throttling Agent for RowHammer Vulnerabilities
Domain: arxiv.org
Comments load interactively on the live page.