Load instructions often limit instruction-level parallelism (ILP) in modern processors due to data and resource dependences they cause. Prior techniques like Load Value Prediction (LVP) and Memory Renaming (MRN) mitigate load data dependence by predicting the data value of a load instruction. However, they fail to mitigate load resource dependence as the predicted load instruction gets executed nonetheless (even on a correct prediction), which consumes hard-to-scale pipeline resources that otherwise could have been used to execute other load instructions. Our goal in this work is to improve ILP by mitigating both load data dependence and resource dependence. To this end, we propose a purely-microarchitectural technique called Constable, that safely eliminates the execution of load instructions. Constable dynamically identifies load instructions that have repeatedly fetched the same data from the same load address. We call such loads likely-stable. For every likely-stable load, Constable (1) tracks modifications to its source architectural registers and memory location via lightweight hardware structures, and (2) eliminates the execution of subsequent instances of the load instruction until there is a write to its source register or a store or snoop request to its load address. Our extensive evaluation using a wide variety of 90 workloads shows that Constable improves performance by $5.1 \%$ while reducing the core dynamic power consumption by $3.4 \%$ on average over a strong baseline system that implements MRN and other dynamic instruction optimizations (e.g., move and zero elimination, constant and branch folding). In presence of 2-way simultaneous multithreading (SMT), Constable’s performance improvement increases to $8.8 \%$ over the baseline system. When combined with a state-of-the-art load value predictor (EVES), Constable provides an additional $3.7 \%$ and $7.8 \%$ average performance benefit over the load value predictor alone, in the baseline system without and with 2-way SMT, respectively.

Deep Neural Networks (DNN) are used in a variety of applications and services. With the evolving nature of DNNs, the race to build optimal hardware (both in datacenter and edge) continues. General purpose multi-core CPUs offer unique attractive advantages for DNN inference at both datacenter [60] and edge [71]. Most of the CPU pipeline design complexity is targeted towards optimizing general-purpose single thread performance, and is overkill for relatively simpler, but still hugely important, data parallel DNN inference workloads. Addressing this disparity efficiently can enable both raw performance scaling and overall performance/Watt improvements for multi-core CPU DNN inference.We present REDUCT, where we build innovative solutions that bypass traditional CPU resources which impact DNN inference power and limit its performance. Fundamentally, REDUCT’s "Keep it close" policy enables consecutive pieces of work to be executed close to each other. REDUCT enables instruction delivery/decode close to execution and instruction execution close to data. Simple ISA extensions encode the fixed-iteration count loop-y workload behavior enabling an effective bypass of many power-hungry front-end stages of the wide Out-of-Order (OoO) CPU pipeline. Per core performance scales efficiently by distributing light-weight tensor compute near all caches in a multi-level cache hierarchy. This maximizes the cumulative utilization of the existing architectural bandwidth resources in the system and minimizes movement of data.Across a number of DNN models, REDUCT achieves a 2.3× increase in convolution performance/Watt with a 2× to 3.94× scaling in raw performance. Similarly, REDUCT achieves a 1.8× increase in inner-product performance/Watt with 2.8× scaling in performance. REDUCT performance/power scaling is achieved with no increase to cache capacity or bandwidth and a mere 2.63% increase in area. Crucially, REDUCT operates entirely within the CPU programming and memory model, simplifying software development, while achieving performance similar to or better than state-of-the-art Domain Specific Accelerators (DSA) for DNN inference, providing fresh design choices in the AI era.

Capability architectures for memory safety have traditionally required expanding pointers and radically changing microarchitectural structures throughout processors, while only providing superficial hardening. We hence propose Cryptographic Capability Computing (C3) - the first memory safety mechanism that is stateless to avoid requiring extra metadata storage. C3 retains 64-bit pointer sizes providing legacy binary compatibility while imposing minimal touchpoints. Pointers are encrypted to unforgeably (within cryptographic bounds) reference each object. Data is encrypted even in caches and entangled with pointers for both spatial and temporal object-granular protection. Pointers become like unique keys for each allocation. C3 deploys a novel form of prediction for address translation that mitigates performance overheads even when addresses are partially encrypted. Use of a low-latency, low-area cipher from the NIST Lightweight Cryptography project avoids delaying loads by readying a data keystream by the time data is returned from the L1 cache. C3 is compatible with legacy binaries. Simulated performance overhead on SPEC CPU2006 is negligible with no memory overhead, which is a big leap forward compared to the overheads imposed by past memory safety approaches. C3 effectively replaces inefficient metadata with efficient cryptography.

Scaling off-chip bandwidth is challenging due to fundamental limitations, such as a fixed pin count and plateauing signaling rates. Recently, vendors have turned to 2.5D and 3D stacking to closely integrate system components. Interestingly, these technologies can integrate a logic layer under multiple memory dies, enabling computing capability inside a memory stack. This trend in stacking is making PIM architectures commercially viable. In this work, we investigate the suitability of offloading kernels in scientific applications onto 3D stacked PIM architectures. We evaluate several hardware constraints resulted from the stacked structure. We perform extensive simulation experiments and in-depth analysis to quantify the impact of application locality in TLBs, data caches, and memory stacks. Our results also identify design optimization areas in software and hardware for HPC scientific applications.

Whereas emerging non-volatile memories (NVMs) are low power, dense, scalable alternatives to DRAM, the high latency and low endurance of these NVMs limit the feasibility of NVM-only memory systems. Smart hybrid memories (SHMs) that integrate NVM, DRAM, and on-module processor logic are an efficient means to bridge the latency and endurance gaps between NVM-only and DRAM-only memory systems. However, these SHMs are vulnerable to data confidentiality and integrity attacks that can be executed on the unsecure NVM, DRAM, and/or memory buses. STASH is the first comprehensive end-to-end SecuriTy Architecture for SHMs that integrates (i) counter mode encryption for data confidentiality, (ii) low overhead page-level Merkle Tree (MT) authentication for data integrity, (iii) recovery-compatible MT updates to withstand power/system failures, and (iv) page-migration-friendly security meta-data management. For security guarantees equivalent to the closest state-of-the-art security solution extensible to SHMs, STASH reduces memory overhead by 12.7 ×, increases system performance by 65%, and improves NVM lifetime by 2.5 ×.

Modern memory systems are susceptible to data confidentiality attacks that leverage memory access pattern information to obtain secret data. Oblivious RAM (ORAM) is a secure cryptographic construct that effectively thwarts access-pattern-based attacks. However, in Path ORAM (state-of-the-art efficient ORAM for main memories) and its variants, each memory request (read or write) is transformed to an ORAM access, which is a sequence of read and write operations, increasing the latency of the memory requests and degrading system performance. In practice, the ORAM access for a read request is on the critical path of program execution, blocked by ORAM accesses for older write requests. Although modern memory controllers (MCs) realize read prioritization through write buffering, the ORAM access translation of each memory request to multiple memory read and write operations results in frequent MC write buffer overflow, decreasing its efficiency. ReadPRO (Read Prioritization) scheduling in ORAM addresses this challenge by promoting read requests over write requests in the ORAM controller prior to their ORAM access translation, while preserving all data dependencies. ReadPRO complements read promotion with staggered writes, wherein ORAM accesses for write requests can be paused securely to serve ORAM accesses for read requests. Full-system evaluations on composite SPEC CPU2006 workloads show that ReadPRO decreases the average ORAM read latency by 75%, improving system performance by 40%.

Data tampering threatens data integrity in emerging non-volatile memories (NVMs). Whereas Merkle Tree (MT) memory authentication is effective in thwarting data tampering attacks, it drastically increases cell writes and memory accesses, adversely impacting NVM energy, lifetime, and system performance (instructions per cycle (IPC)). We propose ASSURE, a low overhead, high performance Authentication Scheme for SecURE energy efficient (ASSURE) NVMs. ASSURE synergistically integrates (i) smart message authentication codes (SMACs), which eliminate redundant cell writes by enabling MAC computation of only modified words on memory writes, with (ii) multi-root MTs (MMTs), which reduce MT reads/writes by constructing either high performance static MMTs (SMMTs) or low overhead dynamic MMTs (DMMTs) over frequently accessed memory regions. Our full-system simulations of the SPEC CPU2006 benchmarks on a triple-level cell (TLC) resistive RAM (RRAM) architecture show that on average, SMMT ASSURE (DMMT ASSURE) reduces NVM energy by 59% (55%), increases memory lifetime by 2.36x (2.11x), and improves IPC by 11% (10%), over state-of-the-art MT memory authentication.

Data persistence in emerging non-volatile memories (NVMs) poses a multitude of security vulnerabilities, motivating main memory encryption for data security. However, practical encryption algorithms demonstrate strong diffusion characteristics that increase cell flips, resulting in increased write energy/latency and reduced lifetime of NVMs. State-of-the-art security solutions have focused on reducing the encryption penalty (increased write energy/latency and reduced memory lifetime) in single-level cell (SLC) NVMs; however, the realization of low encryption penalty solutions for multi-/triple-level cell (MLC/TLC) secure NVMs remains an open area of research. This work synergistically integrates zero-based partial writes with XOR-based energy masking to realize Smartly EnCRypted Energy efficienT, i.e., SECRET MLC/TLC NVMs, without compromising the security of the underlying encryption technique. Our simulations on an MLC (TLC) resistive RAM (RRAM) architecture across SPEC CPU2006 workloads demonstrate that for 6.25% (7.84%) memory overhead, SECRET reduces write energy by 80% (63%), latency by 37% (49%), and improves memory lifetime by 63% (56%) over conventional advanced encryption standard-based (AES-based) counter mode encryption.

Monolayer heterojunction FETs based on vertical heterogeneous transition metal dichalcogenides (TMDCFETs) and planar black phosphorus FETs (BPFETs) have demonstrated excellent subthreshold swing, high I ON I OFF , and high scalability, making them attractive candidates for post-CMOS memory design. This article explores TMDCFET and BPFET SRAM design by combining atomistic self-consistent device modeling with SRAM circuit design and simulation. We perform detailed evaluations of the TMDCFET/BPFET SRAMs at a single bitcell and at SRAM array level. Our simulations show that at low operating voltages, TMDCFET/BPFET SRAMs exhibit significant advantages in static power, dynamic read/write noise margin, and read/write delay over nominal 16nm CMOS SRAMs at both bitcell and array-level implementations. We also analyze the effect of process variations on the performance of TMDCFET/BPFET SRAMs. Our simulations demonstrate that TMDCFET/BPFET SRAMs exhibit high tolerance to process variations, which is desirable for low operating voltages.

Vertical monolayer heterojunction FETs based on transition metal dichalcogenides (TMDCFETs) and planar black phosphorus FETs (BPFETs) have demonstrated excellent sub-threshold swing, high ION/IOFF, and high scalability, making them attractive candidates for post-CMOS memory design. This paper explores TMDCFET and BPFET SRAM design by combining atomistic self-consistent device modeling with SRAM circuit design and simulation. Our simulations show that at low operating voltages, TMDCFET and BPFET SRAMs exhibit significant advantages in static power, dynamic read/write noise margin, and read/write delay over both nominal and read/write-assisted 16nm CMOS SRAMs.