Lecture 13: Computer Architecture

Introduction

• Technical Issues: Faced some initial technical issues.
• Guest Lecturer: Professor Musler is traveling; today's and tomorrow's lectures will be taught by a guest lecturer and colleague John.
• Lecture Structure: Two lectures today focusing on "Memory Controllers" and "Emerging Memory Technologies".

Memory Controllers

Heterogeneous Systems

• Current SoC Architectures: Have a lot of heterogeneity (large and small CPU cores, GPUs, hardware accelerators, DMA).
• Memory Scheduling: Discussed how memory controllers manage resources to mitigate interference and provide predictable performance.
• Memory Scheduling Techniques: Stage memory scheduling for CPU-GPU Integrated Systems.
• Latency Sensitivity: GPU applications are bandwidth-sensitive but error-tolerant. CPU applications are latency-sensitive.

Interference Control Techniques

• Resource Allocation: Techniques to mitigate interference include source throttling, prioritization, and data mapping.
• Predictable Performance: Allocation of resources to heterogeneous agents to mitigate interference.
• Source Throttling: Throttling GPUs to reduce requests when CPU performance degrades.
• QoS Techniques: Include memory channel partitioning to reduce harmful interference.

Emerging Memory Technologies

Overview

• Focus: Emerging memory technologies that can potentially replace DRAM.
• Types: Phase Change Memory (PCM), Spin Transfer Torque Magnetic RAM (STT-RAM), etc.
• Advantages: More scalable than DRAM, non-volatile.
• Challenges: Shortcomings like endurance issues, high latency, and energy consumption.

Phase Change Memory (PCM)

• Technology: Uses phase change material to store data; has amorphous and crystalline states.
• Operations: 'Set' and 'Reset' operations achieved through current injection.
• Opportunities: Better scalability, non-volatility, and potential for higher density.
• Challenges: Higher latency and active energy compared to DRAM; endurance issues.

Hybrid Memory Systems

• Hybrid Approach: Combining DRAM with PCM for better performance and energy efficiency.
• Data Placement: Techniques discussed for optimal data placement to improve performance.
• Simulation Results: Hybrid systems show potential in balancing performance and cost.

Additional Opportunities

Processing Near Memory

• Processing Using Memory (PUM): Utilizes emerging memory technologies for computation purposes (e.g., dot products in crossbar arrays).

Merging Memory and Storage

• Unified Interface: Potential to eliminate the traditional separation of memory and storage systems.
• Persistent Memory: Leverages NVM characteristics for new applications and system designs.

Challenges and Considerations

• Security and Privacy: Issues with data remaining in non-volatile memory after power off.
• Endurance and Reliability: Technologies face wear-out issues affecting longevity.
• Virtual Memory Systems: Rethinking the need for virtual memory in systems with large NVM capabilities.

Conclusion

• Future Prospects: Despite challenges, emerging memory technologies offer potential for significant improvements.
• Research Opportunities: Vast opportunities exist across system layers to optimize and integrate these technologies effectively.