DRAM Row Hammer Mitigation via Victim Row Randomization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
As device geometries decrease, the number of activations required to cause a row hammer event in DRAM devices decreases, making traditional row hammer mitigation techniques impractical and increasing the likelihood of data loss due to simultaneous victim row starvation across multiple DRAM devices.
Innovation Solution
Implementing randomization in the selection of victim rows for row hammer mitigation within DRAM devices, using a unique value stored in a register, such as a PDA enumeration ID, to pseudo-randomly select rows for refresh, ensuring different rows are refreshed across devices, thereby reducing the likelihood of data loss.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional row hammer mitigation techniques are used in DRAM devices, then row hammer events can be addressed, but the complexity of mitigation increases and data loss occurs due to simultaneous victim row starvation across multiple devices
Solution Approach 1:
The patent performs row hammer mitigation refreshes in advance before victim rows become corrupted. By proactively identifying and refreshing potential victim rows based on recent activate patterns, the system prevents row hammer events rather than reacting to them after damage occurs. This preliminary action reduces the need for complex real-time mitigation and prevents data loss.
Solution Approach 2:
The patent dynamically adjusts mitigation strategies based on actual device behavior and workload patterns. Rather than applying static refresh rates to all rows, the system adapts refresh operations to target specific rows that show signs of being aggressors or victims, optimizing mitigation effectiveness while reducing overall complexity and overhead.
2Device complexity
If all DRAM devices apply the same mitigation algorithm, then mitigation can be simplified, but victim rows are starved out resulting in data loss
Solution Approach 1:
The patent introduces asymmetry by assigning unique identifiers to individual DRAM devices and using these to create device-specific refresh patterns. Instead of all devices following the same algorithm, each device operates with a slightly different refresh schedule determined by its unique ID, ensuring that not all devices target the same victim rows simultaneously. This asymmetric approach prevents universal starvation of specific rows across the memory system.
Solution Approach 2:
The patent applies local quality by customizing mitigation behavior at the individual device level rather than using a uniform approach across all devices. Each DRAM device tailors its refresh operations based on its unique characteristics and workload, allowing localized optimization that prevents global starvation while maintaining overall system reliability.
3Speed
If device geometries decrease to increase capacity and access speeds, then performance improves, but row hammer events occur more frequently
Solution Approach 1:
The patent performs row hammer mitigation refreshes in advance before victim rows become corrupted. By proactively identifying and refreshing potential victim rows based on recent activate patterns, the system prevents row hammer events rather than reacting to them after damage occurs. This preliminary action is especially critical in scaled devices where fewer activations are needed to cause row hammer, allowing prevention before charge leakage becomes irreversible.
Solution Approach 2:
The patent implements feedback mechanisms that monitor activate patterns and use this information to guide refresh operations. By tracking which rows are frequently accessed (potential aggressors) and which rows show signs of disturbance (potential victims), the system dynamically adjusts refresh operations to target the right rows at the right time, effectively countering the increased row hammer susceptibility of scaled devices.
Data Source
AI summary
A memory device with internal row hammer mitigation includes randomization for selection of victim rows to refresh for row hammer mitigation. When memory devices connected in groups all use the same probabilistic determination of which row to select for row hammer mitigation, all memory devices could miss refreshing the same victim row, resulting in data loss. With randomization of the selection, the memory devices are more likely to select different potential victim rows for refresh, reducing the risk of data loss. The memory device performs row hammer mitigation during a refresh operation on a row selected based on a recent activate command. Selection of the victim row can be performed with a pseudo-random computation based on a value unique to the memory device in the group.