Accelerating Next-Generation Applications Via Secure and Reliable Compute-in-Memory Systems

点击翻译按钮获取中文摘要

Computing systems execute important applications that form the backbone of the global economy, e.g., machine learning, data and graph analytics, transaction processing, and high-performance computing applications. Emerging next-generation applications include genome sequencing, precision medicine, virtual/augmented reality, autonomous vehicles, and smart homes/cities. These "Big Data" applications have fast-growing data requirements, leading to an ever-increasing memory footprint. Application performance is dependent on, and limited by, the memory system. Frequent data movements between processors and memory lead to slower execution and energy inefficiency. Slower execution could be the difference between online (real-time) and offline processing, making some applications impractical. Energy inefficiency leads to higher energy costs and a growing carbon footprint for computer systems. Compute in Memory (CIM) systems reduce data movement by performing computations closer to data, leading to significant speedups and energy savings. However, commercial adoption has been hindered by significant system challenges. My research program's long-term goal is to enable breakthrough improvements in performance, security, and energy efficiency for future memory systems. This benefits the Canadian economy by enabling real-time execution of key next-generation applications, and training highly qualified personnel for future careers in academia and industry. This proposal addresses two critical but mostly unexplored challenges for CIM systems: Security and reliability. CIM Security/Privacy. Memory is vulnerable to physical, covert and side-channel attacks that could leak programs' private data. To avoid leaks, memory data is encrypted, and is only decrypted when it enters the trusted execution domain (processors and caches). To maintain data integrity and prevent corruption, message authentication codes (MAC) are used. Since CIM computes on unencrypted data, decryption and checking integrity's overheads reduce or eliminate CIM benefits. CIM Reliability. Error rates in memory systems are increasing due to memory capacity scaling (higher number of more vulnerable bit cells), and malicious software that corrupts data. Errors can manifest as failures due to detectable but uncorrectable errors or undetectable errors. Memories deploy error-correcting codes (ECC) to protect against failures. Unfortunately, slow ECC (to offset higher error rates) reduces CIM performance and energy gains. Despite their importance for CIM's commercial adoption, little research has been done to address both challenges. This proposal targets novel software, systems and architecture mechanisms to fill this crucial gap with the following objectives: (1) Modeling security and reliability features for CIM systems; (2) Accelerating computations on encrypted data; (3) Exploring and mitigating side-channel attacks on CIM systems; (4) Accelerating reliable computations on unreliable data.