Quantum Error Correction (QEC) codes form the foundation of Fault-Tolerant Quantum Computing (FTQC) and predominantly use the Clifford+T gate set. Recently, Clifford operations have become the key performance bottleneck in implementing QEC. While state-of-the-art approaches like Pauli-Based Compilation (PBC) reduce Clifford overhead by transforming Clifford gates into Pauli measurements, they do so at the cost of gate-level parallelism, inflating circuit depth and execution times. To overcome these limitations, we introduce TACO, a Transpiler-Architecture Co-design framework that tackles the Clifford bottleneck through circuit and architectural optimization. TACO uses FTQC insights to guide hardware-aware Clifford gate elimination and circuit restructuring, and leverages the resulting optimized circuits to refine architectural design. TACO applies FTQC-specific transformations to aggressively reduce Clifford overhead from rotation synthesis and Toffoli decompositions, while preserving gate-level parallelism. The resulting architecture is optimized for the locality and data-movement patterns of these circuits, enabling high-throughput, resource-efficient execution. Our evaluation across diverse benchmarks shows that TACO achieves up to 21.9x (mean 4.4x) reduction in execution time compared to the state-of-the-art baseline.

For several decades, Dynamic Random Access Memory (DRAM) scaling has enabled exponential growth in computing by supporting larger memory capacities and reducing cost per bit, while fostering the perception of main memory as a passive, reliable, and trusted substrate. However, this paradigm is shifting. As DRAM devices increase in density and heterogeneity, memory systems are now characterized not only by capacity and bandwidth, but also by diverse failure modes, refresh overheads, disturbance errors, side channels, security concerns, and the stringent requirements of modern applications.

Recommendation models play a crucial role in delivering accurate and tailored user experiences. However, training such models poses significant challenges regarding resource utilization and performance. Prior research has proposed an approach that categorizes embeddings into popular and non-popular classes to reduce the training time for recommendation models. We observe that, even among the popular embeddings, certain embeddings undergo rapid training and exhibit minimal subsequent variation, resulting in saturation. Consequently, updates to these embeddings become redundant. This paper presents Slipstream, a software framework that identifies stale embeddings on the fly and skips their updates to enhance performance. Our experiments demonstrate Slipstream's ability to maintain accuracy while effectively discarding updates to non-varying embeddings. This capability enables Slipstream to achieve substantial speedup, optimize CPU-GPU bandwidth usage, and eliminate unnecessary memory access. SlipStream showcases training time reductions of 2 x, 2.4 x, 1.2 x, and 1.175 x across real-world datasets and configurations, compared to Baseline XDL, Intel-optimized DRLM, FAE, and Hotline, respectively.

JEDEC has introduced the Per Row Activation Counting (PRAC) framework for DDR5 and future DRAMs to enable precise counting of DRAM row activations. PRAC enables a holistic mitigation of Rowhammer attacks even at ultra-low Rowhammer thresholds. PRAC uses an Alert Back-Off (ABO) protocol to request the memory controller to issue Rowhammer mitigation requests. However, recent PRAC implementations are either insecure or impractical. For example, Panopticon, the inspiration for PRAC, is rendered insecure if implemented per JEDEC’s PRAC specification. On the other hand, the recent UPRAC proposal is impractical since it needs oracular knowledge of the ‘top- N ‘ activated DRAM rows that require mitigation.This paper provides the first secure, scalable, and practical RowHammer solution using the PRAC framework. The crux of our proposal is the design of a priority-based service queue (PSQ) for mitigations that prioritizes pending mitigations based on activation counts to avoid the security risks of prior solutions. This provides principled security using the reactive ABO protocol. Furthermore, we co-design our PSQ, with opportunistic mitigation on Refresh Management (RFM) operations and proactive mitigation during refresh (REF), to limit the performance impact of ABO-based mitigations. QPRAC provides secure and practical RowHammer mitigation that scales to Rowhammer thresholds as low as 71 while incurring a $0.8 \%$ slowdown for benign workloads, which further reduces to $0 \%$ with proactive mitigations.

RowHammer vulnerabilities pose a significant threat to modern DRAM-based systems, where rapid activation of DRAM rows can induce bit-flips in neighboring rows. To mitigate this, state-of-the-art host-side RowHammer mitigations typically rely on shared counters or tracking structures. While these optimizations benefit benign applications, they are vulnerable to Performance Attacks (Perf-Attacks), where adversaries exploit shared structures to reduce DRAM bandwidth for co-running benign applications by increasing DRAM accesses for RowHammer counters or triggering repetitive refreshes required for the early reset of structures, significantly degrading performance. In this paper, we propose secure hashing mechanisms to thwart adversarial attempts to capture the mapping of shared structures. We propose DAPPER, a novel low-cost tracker resilient to Perf-Attacks even at ultra-low RowHammer thresholds. We first present a secure hashing template in the form of DAPPER-S. We then develop Dapper-H, an enhanced version of DapperS, incorporating double-hashing, novel reset strategies, and mitigative refresh techniques. Our security analysis demonstrates the effectiveness of DAPPER-H against both RowHammer and Perf-Attacks. Experiments with 57 workloads from SPEC2006, SPEC2017, TPC, Hadoop, MediaBench, and YCSB show that, even at an ultra-low RowHammer threshold of 500, DapperH incurs only a 0.9% slowdown in the presence of Perf-Attacks while using only 96 KB of SRAM per 32GB of DRAM memory.

Autoregressive Transformers rely on Key-Value (KV) caching to accelerate inference. However, the linear growth of the KV cache with context length leads to excessive memory consumption and bandwidth constraints. Existing methods drop distant tokens or compress states in a lossy manner, sacrificing accuracy by discarding vital context or introducing bias. We propose ${MorphKV}$, an inference-time technique that maintains a constant-sized KV cache while preserving accuracy. MorphKV balances long-range dependencies and local coherence during text generation. It eliminates early-token bias while retaining high-fidelity context by adaptively ranking tokens through correlation-aware selection. Unlike heuristic retention or lossy compression, MorphKV iteratively refines the KV cache via lightweight updates guided by attention patterns of recent tokens. This approach captures inter-token correlation with greater accuracy, which is crucial for tasks like content creation and code generation. Our studies on long-response tasks show 52.9% memory savings and 18.2% higher accuracy on average compared to state-of-the-art prior works, enabling efficient deployment.

To realize the full potential of quantum computers, we must mitigate qubit errors by developing noise-aware algorithms, compilers, and architectures.Thus, simulating quantum programs on highperformance computing (HPC) systems with different noise models is a de facto tool researchers use.Unfortunately, noisy simulators iteratively execute a similar circuit for thousands of trials, thereby incurring significant performance overheads.To address this, we propose a noisy simulation technique called Tree-Based Quantum Circuit Simulation (TQSim) 1 .TQSim exploits the reusability of intermediate results during the noisy simulation, reducing computation.TQSim dynamically partitions a circuit into several subcircuits.It then reuses the intermediate results from these subcircuits during computation.Compared to a noisy Qulacsbased baseline simulator, TQSim achieves a speedup of up to 3.89× for noisy simulations.TQSim is designed to be efficient with multinode setups while also maintaining tight fidelity bounds.

Sparse Matrix Multiplication (SpMM) is essential in various scientific and engineering applications but poses significant challenges due to irregular memory access patterns.Many hardware accelerators have been proposed to accelerate SpMM.However, they have yet to focus on on-chip memory utilization.In this paper, we highlight the underutilization of the on-chip memory in the SpMM accelerators.Then we propose Avalanche, a novel hardware accelerator that optimally utilizes the on-chip memory to efficiently cache both matrices 𝐵 and 𝐶.Avalanche incorporates three key techniques: Matrix Reordering (Mat-Reorder), Dead-Product Early Eviction (DP-Evict), and Reuse Distance-Aware Matrix Caching (RM-Caching).Mat-Reorder enhances data locality by reordering the columns of matrix 𝐴, ensuring early completion of computations for matrix 𝐶.DP-Evict optimizes on-chip memory usage by promptly evicting fully computed (dead) products from on-chip memory.RM-Caching maximizes data reuse by caching frequently accessed elements of matrix 𝐵 based on their reuse distance.Experimental results demonstrate that Avalanche achieves an average performance improvement of 1.97× compared to the state-of-theart SpMM accelerator, with a chip area of 6.15 mm 2 CCS Concepts• Computer systems organization → Special purpose systems.

Graph analytics and sparse linear algebra applications heavily rely on indirect memory access (IMA).IMAs are characterized by poor temporal and spatial locality, which causes frequent high-latency DRAM accesses.While dedicated hardware prefetchers for IMA have been explored, they target narrow access patterns and tend to introduce significant hardware complexity.Software prefetching offers a promising alternative, leveraging compiler analysis to prefetch indirection patterns.However, existing software prefetchers struggle with sparse applications due to limited loop iterations and complex IMA patterns across nested loops.We propose Magellan, a novel loop-guided software prefetcher designed to detect and schedule IMA prefetches efficiently.Magellan introduces two key innovations: (1) extracting dependence graphs across loop levels to detect complex IMA patterns and (2) capturing inner-outer loop semantics to prefetch for both current and future iterations.We evaluate Magellan on 14 memory-intensive benchmarks using real-world datasets from social networks and web graphs.Compared to the best existing IMA software prefetcher, Magellan reduces cache misses by 25% and dynamic instruction counts by 14% on average.This results in a 1.14× average speedup, with performance gains of up to 1.41×.

Per Row Activation Counting (PRAC) has emerged as a robust framework for mitigating RowHammer (RH) vulnerabilities in modern DRAM systems.However, we uncover a critical vulnerability: a timing channel introduced by the Alert Back-Off (ABO) protocol and Refresh Management (RFM) commands.We present PRACLeak, a novel attack that exploits these timing differences to leak sensitive information, such as secret keys from vulnerable AES implementations, by monitoring memory access latencies.To counter this, we propose Timing-Safe PRAC (TPRAC), a defense that eliminates PRAC-induced timing channels without compromising RH mitigation efficacy.TPRAC uses Timing-Based RFMs, issued periodically and independent of memory activity.It requires only a single-entry in-DRAM mitigation queue per DRAM bank and is compatible with existing DRAM standards.Our evaluations demonstrate that TPRAC closes timing channels while incurring only 3.4% performance overhead at the RH threshold of 1024. CCS Concepts• Security and privacy → Security in hardware.

Diffusion Transformers (DiTs) achieve state-of-the-art results in text-to-image, text-to-video generation, and editing. However, their large model size and the quadratic cost of spatial-temporal attention over multiple denoising steps make video generation computationally expensive. Static caching mitigates this by reusing features across fixed steps but fails to adapt to generation dynamics, leading to suboptimal trade-offs between speed and quality. We propose Foresight, an adaptive layer-reuse technique that reduces computational redundancy across denoising steps while preserving baseline performance. Foresight dynamically identifies and reuses DiT block outputs for all layers across steps, adapting to generation parameters such as resolution and denoising schedules to optimize efficiency. Applied to OpenSora, Latte, and CogVideoX, Foresight achieves up to 1.63x end-to-end speedup, end-to-end speedup, while maintaining video quality.

The Quantum Approximate Optimization Algorithm (QAOA) addresses combinatorial optimization challenges by converting inputs to graphs. However, the optimal parameter searching process of QAOA is greatly affected by noise. Larger problems yield bigger graphs, requiring more qubits and making their outcomes highly noise-sensitive. This paper introduces Red-QAOA, leveraging energy landscape concentration via a simulated annealing-based graph reduction.

Emerging deep learning workloads urgently need fast general matrix multiplication (GEMM). Thus, one of the critical features of machine-learning-specific accelerators such as NVIDIA Tensor Cores, AMD Matrix Cores, and Google TPUs is the support of mixed-precision enabled GEMM. For DNN models, lower-precision FP data formats and computation offer acceptable correctness but significant performance, area, and memory footprint improvement. While promising, the mixed-precision computation on error resilience remains unexplored. To this end, we develop a fault injection framework that systematically injects fault into the mixed-precision computation results. We investigate how the faults affect the accuracy of machine learning applications. Based on error resilience characteristics, we offer lightweight error detection and correction solutions that significantly improve the overall model accuracy by 75% if the models experience hardware faults. The solutions can be efficiently integrated into the accelerator's pipelines.

Recommendation models rely on deep learning networks and large embedding tables, resulting in computationally and memory-intensive processes. These models are typically trained using hybrid CPU-GPU or GPU-only configurations. The hybrid mode combines the GPU’s neural network acceleration with the CPUs’ memory storage and supply for embedding tables but may incur significant CPU-to-GPU transfer time. In contrast, the GPU-only mode utilizes High Bandwidth Memory (HBM) across multiple GPUs for storing embedding tables. However, this approach is expensive and presents scaling concerns. This paper introduces Hotline, a heterogeneous acceleration pipeline that addresses these concerns. Hotline develops a dataaware and model-aware scheduling pipeline by leveraging the insight that only a few embedding entries are frequently accessed (popular). This approach utilizes CPU main memory for nonpopular embeddings and GPUs’ HBM for popular embeddings. To achieve this, Hotline accelerator fragments a mini-batch into popular and non-popular micro-batches ($\mu$-batches). It gathers the necessary working parameters for non-popular $\mu$-batches from the CPU, while GPUs execute popular $\mu$-batches. The hardware accelerator dynamically coordinates the execution of popular embeddings on GPUs and non-popular embeddings from the CPU’s main memory. Real-world datasets and models confirm Hotline’s effectiveness, reducing average end-to-end training time by $2.2 \times$ compared to Intel-optimized CPU-GPU DLRM baseline.

Quantum computers face challenges due to limited resources, particularly in cloud environments. Despite these ob-stacles, Variational Quantum Algorithms (VQAs) are considered promising applications for present-day Noisy Intermediate-Scale Quantum (NISQ) systems. VQAs require multiple optimization iterations to converge on a globally optimal solution. Moreover, these optimizations, known as restarts, need to be repeated from different points to mitigate the impact of noise. Unfortunately, the job scheduling policies for each VQA task in the cloud are heavily unoptimized. Notably, each VQA execution instance is typically scheduled on a single NISQ device. Given the variety of devices in the cloud, users often prefer higher-fidelity devices to ensure higher-quality solutions. However, this preference leads to increased queueing delays and unbalanced resource utilization. We propose Qoncord, an automated job scheduling framework to address these cloud-centric challenges for VQAs. Qoncord leverages the insight that not all training iterations and restarts are equal, Qoncord strategically divides the training process into exploratory and fine-tuning phases. Early exploratory iterations, more resilient to noise, are executed on less busy machines, while fine-tuning occurs on high-fidelity machines. This adaptive approach mitigates the impact of noise, optimizes resource usage, and reduces queuing delays in cloud environments. Qoncord also significantly reduces execution time and minimizes restart overheads by eliminating low-performance iterations. Thus, Qoncord offers similar solutions 17.4 × faster. It also provides 13.3% better solutions for the same time budget as the baseline.

A unified address space is vital for heterogeneous systems as it enables efficient data sharing between CPUs and GPUs. However, GPU address translation faces challenges due to high TLB pressure, particularly with irregular and memory-intensive applications. Compared to an ideal scenario, we observe that address translation overheads cause a slowdown of up to 34.5% in modern heterogeneous systems. This paper introduces Avatar, a novel framework to accelerate address translation in GPUs. Avatar comprises two key components: Contiguity-Aware Speculative Translation (CAST) and In-Cache Validation (CAVA) mechanisms. Avatar identifies the potential for predicting virtual-to-physical address mapping by monitoring contiguous pages that lie in both virtual and physical address spaces. Leveraging this insight, CAST speculatively translates virtual addresses into physical addresses. This speculative address translation enables immediate data fetching into GPUs while addressing translation occurs in the background, reducing TLB-miss overhead. Unfortunately, modern GPUs lack support for speculative execution, which limits CAST's performance gain. Data fetched from speculated physical addresses is unusable until validation. CAVA addresses this limitation by quickly validating speculated physical addresses. To this end, CAVA embeds page mapping information into each 32B sector of 128B cache lines. Thus, CAVA enables fetching a sector block from memory for a speculated address and rapidly validating the speculative translation using the embedded mapping information. Our experiments show that Avatar achieves a 90.3% (high) speculation accuracy and improves GPU performance by 37.2% (on average).

Graphics Processing Units (GPUs), while offering exceptional performance for CNN inference tasks, are susceptible to both transient and permanent hardware faults due to the integration of numerous processing elements and advancements in technology scaling. This paper proposes a novel and cost-effective fault mitigation technique, called Sparsity-aware Fault Tolerance (SparseFT), to ensure reliable CNN inference on GPUs. SparseFT leverages inherent sparsity in the activation maps to detect and correct errors on the processing elements without hardware redundancy. By exploiting the characteristic of dot-products, where multiplications with zero operands are ineffectual, SparseFT dynamically duplicates an effectual computation (i.e., a multiplication with non-zero operands) to the processing element initially assigned to the ineffectual one. It then compares the duplicated computation results to detect errors. Experimental results demonstrate that SparseFT achieves more than 97% error detection coverage with less than 1% performance overhead for the state-of-the-art CNN models.

Deep learning models are a valuable “secret sauce” that confers a significant competitive advantage. Many models are never visible to the user and even publicly known state-of-the-art models are either completely proprietary or only accessible via access-controlled APIs. Increasingly, these models run directly on the edge, often using a low-power DNN accelerator. This makes models particularly vulnerable, as an attacker with physical access can exploit side channels like off-chip memory access volumes. Indeed, prior work has shown that this channel can be used to steal dense DNNs from edge devices by correlating data transfer volumes with layer geometry.

As Dynamic Random Access Memories (DRAM) scale, they are becoming increasingly susceptible to Row Hammer. By rapidly activating rows of DRAM cells (aggressor rows), attackers can exploit inter-cell interference through Row Hammer to flip bits in neighboring rows (victim rows). A recent work, called Randomized Row-Swap (RRS), proposed proactively swapping aggressor rows with randomly selected rows before an aggressor row can cause Row Hammer.Our paper observes that RRS is neither secure nor scalable. We first propose the ‘Juggernaut attack pattern’ that breaks RRS in under 1 day. Juggernaut exploits the fact that the mitigative action of RRS, a swap operation, can itself induce additional target row activations, defeating such a defense. Second, this paper proposes a new defense Secure Row-Swap mechanism that avoids the additional activations from swap (and unswap) operations and protects against Juggernaut. Furthermore, this paper extends Secure Row-Swap with attack detection to defend against even future attacks. While this provides better security, it also allows for securely reducing the frequency of swaps, thereby enabling Scalable and Secure Row-Swap. The Scalable and Secure Row-Swap mechanism provides years of Row Hammer protection with 3.3× lower storage overheads as compared to the RRS design. It incurs only a 0.7% slowdown as compared to a not-secure baseline for a Row Hammer threshold of 1200.

Federated Learning (FL) allows machine learning models to train locally on individual mobile devices, synchronizing model updates via a shared server. This approach safeguards user privacy; however, it also generates a heterogeneous training environment due to the varying performance capabilities across devices. As a result, “straggler” devices with lower performance often dictate the overall training time in FL. In this work, we aim to alleviate this performance bottleneck due to stragglers by dynamically balancing the training load across the system. We introduce Invariant Dropout, a method that extracts a sub-model based on the weight update threshold, thereby minimizing potential impacts on accuracy. Building on this dropout technique, we develop an adaptive training framework, Federated Learning using Invariant Dropout (FLuID). FLuID offers a lightweight sub-model extraction to regulate computational intensity, thereby reducing the load on straggler devices without affecting model quality. Our method leverages neuron updates from non-straggler devices to construct a tailored sub-model for each straggler based on client performance profiling. Furthermore, FLuID can dynamically adapt to changes in stragglers as runtime conditions shift. We evaluate FLuID using five real-world mobile clients. The evaluations show that Invariant Dropout maintains baseline model efficiency while alleviating the performance bottleneck of stragglers through a dynamic, runtime approach.

The advent of High-Performance Computing has led to the adoption of Convolutional Neural Networks (CNNs) in safety-critical applications such as autonomous vehicles. However, CNNs are vulnerable to DRAM errors corrupting their parameters, thereby degrading their accuracy. Existing techniques for protecting CNNs from DRAM errors are either expensive or fail to protect from large-granularity, multi-bit errors, which occur commonly in DRAMs.

Row Hammer is a fault-injection attack in which rapid activations to a single DRAM row causes bit-flips in nearby rows. Several recent defenses propose tracking aggressor-rows and applying mitigating action on neighboring victim rows by refreshing them. However, all such proposals using victim-focused mitigation preserve the spatial connection between victim and aggressor rows. Therefore, these proposals are susceptible to access patterns causing bit-flips in rows beyond the immediate neighbor. For example, the Half-Double attack causes bit-flips in the presence of victim-focused mitigation.

Row-Hammer (RH) is a DRAM data-disturbance failure that occurs when a row is activated frequently, which causes bit-flips in nearby rows. Row-Hammer is a significant security threat as an attacker can exploit the bit-flips to do privilege escalation and leak confidential data. While several solutions aim to mitigate RH, such solutions depend on the RH threshold and adversarial access patterns. Solutions developed for a given threshold become ineffective for newer devices with lower thresholds, and new attack patterns continue to break existing mitigations. Currently, there is no guaranteed solution for RH, which means that the system remains vulnerable to security threats even in the presence of RH mitigation.In this paper, we contend that simply relying on RH mitigation is insufficient to provide security in the presence of reducing threshold and motivated attackers. We propose SafeGuard, which equips the system with low-cost integrity protection as a defense against potential attacks that break the RH mitigation. As SafeGuard can detect arbitrary failures, it converts the problem of RH bit-flips from a security threat (silent consumption of corrupted data) to a reliability concern (detectable uncorrectable errors caused by integrity violation). We develop SafeGuard for systems that employ ECC modules and show that SafeGuard can provide strong detection to both SECDED (46-bit MAC per cache-line) and Chipkill (32-bit MAC per cache-line) while retaining the correction capability of conventional designs. SafeGuard avoids incurring any storage overheads in DRAM by simply reorganizing the ECC code to operate at a cache-line granularity (64 bytes) instead of a word granularity (8 bytes). Our evaluations show that SafeGuard has a negligible impact on both the system performance (0.7%) and the system reliability due to naturally occurring errors while still providing a strong defense against the security risk of RH by detecting arbitrary bit-flips.

DRAM systems continue to be plagued by the Row-Hammer (RH) security vulnerability. The threshold number of row activations (TRH) required to induce RH has reduced rapidly from 139K in 2014 to 4.8K in 2020, and TRH is expected to reduce further, making RH even more severe for future DRAM. Therefore, solutions for mitigating RH should be effective not only at current TRH but also at future TRH. In this paper, we investigate the mitigation of RH at ultra-low thresholds (500 and below). At such thresholds, state-of-the-art solutions, which rely on SRAM or CAM for tracking row activations, incur impractical storage overheads (340KB or more per rank at TRH of 500), making such solutions unappealing for commercial adoption. Alternative solutions, which store per-row metadata in the addressable DRAM space, incur significant slowdown (25% on average) due to extra memory accesses, even in the presence of metadata caches. Our goal is to develop scalable RH mitigation while incurring low SRAM and performance overheads.

Rowhammer allows an attacker to induce bit flips in a row by rapidly accessing neighboring rows. Rowhammer is a severe security threat as it can be used to escalate privilege or break confidentiality. Moreover, the threshold of activations needed to induce Rowhammer continues to reduce and new attacks like Half-Double break existing solutions that refresh victim rows. The recently proposed Randomized Row-Swap (RRS) scheme is resilient to Half-Double as it provides mitigation by swapping an aggressor row with a random row. However, to ensure security, the threshold for triggering a row-swap must be set much lower than the Rowhammer threshold, leading to a significant performance loss of 20% on average, at a Rowhammer threshold of 1K. Furthermore, the SRAM overhead for storing the indirection table of RRS becomes prohibitively large – 2.4MB per rank at a Rowhammer threshold of 1K. Our goal is to develop a scalable Rowhammer mitigation that incurs negligible performance and storage overheads.To this end, we propose AQUA, a Rowhammer mitigation that breaks the spatial correlation between aggressor and victim rows by dynamically quarantining the aggressor row in a dedicated region of memory. AQUA allows for an effective row migration threshold much higher than in RRS, leading to an order of magnitude less slowdown and SRAM. As the security of AQUA is not reliant on keeping the destination row a secret, we further reduce the SRAM overheads of the indirection table by storing it in DRAM, and accessing it on-demand. We derive the size of the quarantine region required to ensure security for AQUA and show that reserving about 1% of DRAM is sufficient to mitigate Rowhammer at a threshold of 1K. Our evaluations show that AQUA incurs an average slowdown of 2% and an SRAM overhead (for mapping and migration) of only 41KB per rank at a Rowhammer threshold of 1K.

The popularity of Deep Neural Networks (DNNs) has led to many DNN accelerator architectures, which typically focus on the on-chip storage and computation costs. However, much of the energy is spent on accesses to off-chip DRAM memory. While emerging resistive memory technologies such as MRAM, PCM, and RRAM can potentially reduce this energy component, they suffer from drawbacks such as low endurance that prevent them from being a DRAM replacement in DNN applications. In this paper, we examine how DNN accelerators can be designed to overcome these limitations and how emerging memories can be used for off-chip storage. We demonstrate that through (a) careful mapping of DNN computation to the accelerator and (b) a hybrid setup (both DRAM and an emerging memory), we can reduce inference energy over a DRAM-only design by a factor ranging from 1.12× on EfficientNetB7 to 6.3× on ResNet-50, while also increasing the endurance from 2 weeks to over a decade. As the energy benefits vary dramatically across DNN models, we also develop a simple analytical heuristic solely based on DNN model parameters that predicts the suitability of a given DNN for emerging-memory-based accelerators.

Recommender models are commonly used to suggest relevant items to a user for e-commerce and online advertisement-based applications. These models use massive embedding tables to store numerical representation of items' and users' categorical variables (memory intensive) and employ neural networks (compute intensive) to generate final recommendations. Training these large-scale recommendation models is evolving to require increasing data and compute resources. The highly parallel neural networks portion of these models can benefit from GPU acceleration however, large embedding tables often cannot fit in the limited-capacity GPU device memory. Hence, this paper deep dives into the semantics of training data and obtains insights about the feature access, transfer, and usage patterns of these models. We observe that, due to the popularity of certain inputs, the accesses to the embeddings are highly skewed with a few embedding entries being accessed up to 10000X more. This paper leverages this asymmetrical access pattern to offer a framework, called FAE, and proposes a hot-embedding aware data layout for training recommender models. This layout utilizes the scarce GPU memory for storing the highly accessed embeddings, thus reduces the data transfers from CPU to GPU. At the same time, FAE engages the GPU to accelerate the executions of these hot embedding entries. Experiments on production-scale recommendation models with real datasets show that FAE reduces the overall training time by 2.3X and 1.52X in comparison to XDL CPU-only and XDL CPU-GPU execution while maintaining baseline accuracy.

In this paper, we identify a previously untapped source of compressibility in cache working sets: clusters of cachelines that are similar, but not identical, to one another. To compress the cache, we can then store the "clusteroid" of each cluster together with the (much smaller) "diffs" needed to reconstruct the rest of the cluster. To exploit this opportunity, we propose a hardware-level on-line cacheline clustering mechanism based on locality-sensitive hashing. Our method dynamically forms clusters as they appear in the data access stream and retires them as they disappear from the cache. Our evaluations show that we achieve 2.25× compression on average (and up to 9.9×) on SPEC~CPU~2017 suite and is significantly higher than prior proposals scaled to an iso-silicon budget.

The importance of caches for performance, and their high silicon area cost, have motivated hardware solutions that transparently compress the cached data to increase effective capacity without sacrificing silicon area. To this end, prior work has taken one of two approaches: either (a) deduplicating identical cache blocks across the cache to take advantage of inter-block redundancy or (b) compressing common patterns within each cache block to take advantage of intra-block redundancy.(p)(/p)In this paper, we demonstrate that leveraging only one of these redundancy types leads to a significant loss in compression opportunities for several applications: some workloads exhibit either inter-block or intra-block redundancy, while others exhibit both. We propose 2DCC (Two Dimensional Cache Compression), a simple technique that takes advantage of both types of redundancy. Across the SPEC and Parsec benchmark suites, 2DCC results in a 2.12× compression factor (geomean) compared to 1.44-1.49× for best prior techniques on an iso-silicon basis. For the cache-sensitive subset of these benchmarks run in isolation, 2DCC also achieves a 11.7% speedup (geomean).

Spin-Transfer Torque Random Access Memory (STT-RAM) is a potential alternative for SRAM-based on-chip caches. STT-RAM offers high density and low leakage power, thereby can be used to build a large capacity last-level caches (LLC). Unfortunately, the write latency of the STT-RAM is significantly longer, and its write energy is considerably higher compared to SRAM. To mitigate these concerns, researchers have proposed hybrid caches that are comprised of SRAM and STT-RAM regions. In such hybrid caches, an intelligent block placement policy is necessary to store as many write-intensive blocks in the SRAM region. This paper proposes an adaptive block placement framework with metadata embedding (ADAM) for hybrid caches. ADAM embeds metadata (i.e., write-intensity) into a cache block when it is evicted from LLC. When a cache block is brought from the main memory, metadata embedded in the block is extracted and used to determine the write-intensity of the block. Our evaluation shows that ADAM can improve performance by 26 % (on average) over a baseline block placement scheme.

Conventionally, systems have relied on technology scaling to provide smaller cells, which helps in increasing the capacity of on-chip and off-chip structures. Unfortunately, scaling technology to smaller nodes causes increased susceptibility to faults. We study the problem of efficiently tolerating transient failures using scalable Spin-Transfer Torque RAM (STTRAM) as an example. At smaller feature sizes, the energy required to flip a STTRAM cell reduces, which makes these cells more susceptible to random failures caused by thermal noise. Such failures can be tolerated by periodic scrubbing and provisioning each line with Error Correction Code (ECC). However, to tolerate the desired bit-error-rate, the cache needs ECC-6 (six bit error correction) per line, incurring impractical storage overheads. Ideally, we want to tolerate these faults without relying on multi-bit ECC. We propose SuDoku, a design that provisions each line with ECC-1 and a strong error detection code, and relies on a region-based RAID-4 to perform correction of multi-bit errors. Unfortunately, simply having such a RAID-4 based architecture is ineffective at tolerating a high-rate of transient faults and provides an MTTF in the order of only a few seconds. We describe a novel data resurrection scheme that can repair multiple faulty lines in a RAID-4 region to increase the MTTF to several hours. We propose an extension of SuDoku, which hashes a given line into two regions of RAID-4 to significantly enhance reliability and increase the MTTF to trillions of hours. Our evaluations show that SuDoku provides 874x higher reliability than ECC-6, incurs 30% less storage than ECC-6, and performs within 0.1% of an ideal fault-free baseline.

Existing and near-term quantum computers face significant reliability challenges because of high error rates caused by noise. Such machines are operated in the Noisy Intermediate Scale Quantum (NISQ) model of computing. As NISQ machines exhibit high error-rates, only programs that require a few qubits can be executed reliably. Therefore, NISQ machines tend to underutilize its resources. In this paper, we propose to improve the throughput and utilization of NISQ machines by using multi-programming and enabling the NISQ machine to concurrently execute multiple workloads.

Compression is seen as a simple technique to increase the effective cache capacity. Unfortunately, compression techniques either incur tag area overheads or restrict cache block placement to only include neighboring addresses. Ideally, we should be able to place compressed cache blocks without any restrictions or overheads.

Building trusted data-centers requires resilient memories which are protected from both adversarial attacks and errors. Unfortunately, the state-of-the-art memory security solutions incur considerable performance overheads due to accesses for security metadata like Message Authentication Codes (MACs). At the same time, commercial secure memory solutions tend to be designed oblivious to the presence of memory reliability mechanisms (such as ECC-DIMMs), that provide tolerance to memory errors. Fortunately, ECC-DIMMs possess an additional chip for providing error correction codes (ECC), that is accessed in parallel with data, which can be harnessed for security optimizations. If we can re-purpose the ECC-chip to store some metadata useful for security and reliability, it can prove beneficial to both. To this end, this paper proposes Synergy, a reliability-security co-design that improves performance of secure execution while providing strong reliability for systems with 9-chip ECC-DIMMs. Synergy uses the insight that MACs being capable of detecting data tampering are also useful for detecting memory errors. Therefore, MACs are best suited for being placed inside the ECC chip, to be accessed in parallel with each data access. By co-locating MAC and Data, Synergy is able to avoid a separate memory access for MAC and thereby reduce the overall memory traffic for secure memory systems. Furthermore, Synergy is able to tolerate 1 chip failure out of 9 chips by using a parity that is constructed over 9 chips (8 Data and 1 MAC), which is used for reconstructing the data of the failed chip. For memory intensive workloads, Synergy provides a speedup of 20% and reduces system Energy Delay Product by 31% compared to a secure memory baseline with ECC-DIMMs. At the same time, Synergy increases reliability by 185x compared to ECC-DIMMs that provide Single-Error Correction, Double-Error Detection (SECDED) capability. Synergy uses commercial ECC-DIMMs and does not incur any additional hardware overheads or reduction of security.

Memory systems are becoming bandwidth constrained and data compression is seen as a simple technique to increase their effective bandwidth. However, data compressionrequires accessing Metadata which incurs additional bandwidth overheads. Even after using a Metadata-Cache, the bandwidth overheads of Metadata can reduce the benefits of compression. This paper proposes Attaché, a framework that reduces the overheads of Metadata accesses. The Attaché framework consists of two components. The first component, called the Blended Metadata Engine (BLEM), enables data and its Metadata to be accessed together. BLEM incurs additional Metadata accesses only 0.003% times and removes almost all Metadata bandwidth overheads. The second component, called theCompression Pre-dictor(COPR), predicts if the memory block is compressed. TheCOPR predictor uses a fine-grained line-level predictor, a coarse-grained page-level predictor, and a global indicator. This enables Attaché to predict the compressibility of the memory block before sending a memory read request. We implement Attaché on a memory system that uses Sub-Ranking. On average, Attaché achieves 15.3% speedup (ideal 17%) and saves 22% energy consumption (ideal 23%) when compared to a baseline system that does not employ data compression. Attaché is completely hardware-based and uses only 368KB of SRAM.

Securing off-chip main memory is essential for protection from adversaries with physical access to systems. However, current secure-memory designs incur considerable performance overheads - a major cause being the multiple memory accesses required for traversing an integrity-tree, that provides protection against man-in-the-middle attacks or replay attacks. In this paper, we provide a scalable solution to this problem by proposing a compact integrity tree design that requires fewer memory accesses for its traversal. We enable this by proposing new storage-efficient representations for the counters used for encryption and integrity-tree in secure memories. Our Morphable Counters are more cacheable on-chip, as they provide more counters per cacheline than existing split counters. Additionally, they incur lower overheads due to counter-overflows, by dynamically switching between counter representations based on usage pattern. We show that using Morphable Counters enables a 128-ary integrity-tree, that can improve performance by 6.3% on average (up to 28.3%) and reduce system energy-delay product by 8.8% on average, compared to an aggressive baseline using split counters with a 64-ary integrity-tree. These benefits come without any additional storage or reduction in security and are derived from our compact counter representation, that reduces the integrity-tree size for a 16GB memory from 4MB in the baseline to 1MB. Compared to recently proposed VAULT, our design provides a speedup of 13.5% on average (up to 47.4%).

A quantum computer consists of quantum bits (qubits) and a control processor that acts as an interface between the programmer and the qubits. As qubits are very sensitive to noise, they rely on continuous error correction to maintain the correct state. Current proposals rely on software-managed error correction and require large instruction bandwidth, which must scale in proportion to the number of qubits. While such a design may be reasonable for small-scale quantum computers, we show that instruction bandwidth tends to become a critical bottleneck for scaling quantum computers.

This paper introduces a new DRAM design that enables fast and energy-efficient bulk data movement across subarrays in a DRAM chip. While bulk data movement is a key operation in many applications and operating systems, contemporary systems perform this movement inefficiently, by transferring data from DRAM to the processor, and then back to DRAM, across a narrow off-chip channel. The use of this narrow channel for bulk data movement results in high latency and energy consumption. Prior work proposed to avoid these high costs by exploiting the existing wide internal DRAM bandwidth for bulk data movement, but the limited connectivity of wires within DRAM allows fast data movement within only a single DRAM subarray. Each subarray is only a few megabytes in size, greatly restricting the range over which fast bulk data movement can happen within DRAM. We propose a new DRAM substrate, Low-Cost Inter-Linked Subarrays (LISA), whose goal is to enable fast and efficient data movement across a large range of memory at low cost. LISA adds low-cost connections between adjacent subarrays. By using these connections to interconnect the existing internal wires (bitlines) of adjacent subarrays, LISA enables wide-bandwidth data transfer across multiple subarrays with little (only 0.8%) DRAM area overhead. As a DRAM substrate, LISA is versatile, enabling an array of new applications. We describe and evaluate three such applications in detail: (1) fast inter-subarray bulk data copy, (2) in-DRAM caching using a DRAM architecture whose rows have heterogeneous access latencies, and (3) accelerated bitline precharging by linking multiple precharge units together. Our extensive evaluations show that each of LISA's three applications significantly improves performance and memory energy efficiency, and their combined benefit is higher than the benefit of each alone, on a variety of workloads and system configurations.

Large-granularity memory failures continue to be a critical impediment to system reliability. To make matters worse, as DRAM scales to smaller nodes, the frequency of unreliable bits in DRAM chips continues to increase. To mitigate such scaling-related failures, memory vendors are planning to equip existing DRAM chips with On-Die ECC. For maintaining compatibility with memory standards, On-Die ECC is kept invisible from the memory controller. This paper explores how to design high reliability memory systems in presence of On-Die ECC. We show that if On-Die ECC is not exposed to the memory system, having a 9- chip ECC-DIMM (implementing SECDED) provides almost no reliability benefits compared to an 8-chip non-ECC DIMM. We also show that if the error detection of On-Die ECC can be exposed to the memory controller, then Chipkill-level reliability can be achieved even with a 9-chip ECC-DIMM. To this end, we propose eXposed On-Die Error Detection (XED), which exposes the On-Die error detection information without requiring changes to the memory standards or consuming bandwidth overheads. When the On-Die ECC detects an error, XED transmits a pre-defined "catch-word" instead of the corrected data value. On receiving the catch-word, the memory controller uses the parity stored in the 9-chip of the ECC-DIMM to correct the faulty chip (similar to RAID-3). Our studies show that XED provides Chipkill-level reliability (172x higher than SECDED), while incurring negligible overheads, with a 21% lower execution time than Chipkill. We also show that XED can enable Chipkill systems to provide Double-Chipkill level reliability while avoiding the associated storage, performance, and power overheads.

As memory systems scale, maintaining their Reliability Availability and Serviceability (RAS) is becoming more complex. To make matters worse, recent studies of DRAM failures in data centers and supercomputer environments have highlighted that large-granularity failures are common in DRAM chips. Furthermore, the move toward 3D-stacked memories can make the system vulnerable to newer failure modes, such as those occurring from faults in Through-Silicon Vias (TSVs). To architect future systems and to use emerging technology, system designers will need to employ strong error correction and repair techniques. Unfortunately, evaluating the relative effectiveness of these reliability mechanisms is often difficult and is traditionally done with analytical models, which are both error prone and time-consuming to develop. To this end, this article proposes F ault S im , a fast configurable memory-reliability simulation tool for 2D and 3D-stacked memory systems. FaultSim employs Monte Carlo simulations, which are driven by real-world failure statistics. We discuss the novel algorithms and data structures used in FaultSim to accelerate the evaluation of different resilience schemes. We implement BCH-1 (SECDED) and ChipKill codes using FaultSim and validate against an analytical model. FaultSim implements BCH-1 and ChipKill codes with a deviation of only 0.032% and 8.41% from the analytical model. FaultSim can simulate 1 million Monte Carlo trials (each for a period of 7 years) of BCH-1 and ChipKill codes in only 34 seconds and 33 seconds, respectively.

Stacked memory modules are likely to be tightly integrated with the processor. It is vital that these memory modules operate reliably, as memory failure can require the replacement of the entire socket. To make matters worse, stacked memory designs are susceptible to newer failure modes (e.g., due to faulty through-silicon vias, or TSVs) that can cause large portions of memory, such as a bank, to become faulty. To avoid data loss from large-granularity failures, the memory system may use symbol-based codes that stripe the data for a cache line across several banks (or channels). Unfortunately, such data-striping reduces memory-level parallelism, causing significant slowdown and higher power consumption. This article proposes Citadel , a robust memory architecture that allows the memory system to retain each cache line within one bank. By retaining cache lines within banks, Citadel enables a high-performance and low-power memory system and also efficiently protects the stacked memory system from large-granularity failures. Citadel consists of three components; TSV-Swap , which can tolerate both faulty data-TSVs and faulty address-TSVs; Tri-Dimensional Parity (3DP), which can tolerate column failures, row failures, and bank failures; and Dynamic Dual-Granularity Sparing (DDS) , which can mitigate permanent faults by dynamically sparing faulty memory regions either at a row granularity or at a bank granularity. Our evaluations with real-world data for DRAM failures show that Citadel provides performance and power similar to maintaining the entire cache line in the same bank, and yet provides 700 × higher reliability than ChipKill-like ECC codes.

Phase Change Memory (PCM) is an emerging Non Volatile Memory (NVM) technology that has the potential to provide scalable high-density memory systems. While the non-volatility of PCM is a desirable property in order to save leakage power, it also has the undesirable effect of making PCM main memories susceptible to newer modes of security vulnerabilities, for example, accessibility to sensitive data if a PCM DIMM gets stolen. PCM memories can be made secure by encrypting the data. Unfortunately, such encryption comes with a significant overhead in terms of bits written to PCM memory, causing half of the bits in the line to change on every write, even if the actual number of bits being written to memory is small. Our studies show that a typical writeback modifies, on average, only 12% of the bits in the cacheline. Thus, encryption causes almost a 4x increase in the number of bits written to PCM memories. Such extraneous bit writes cause significant increase in write power, reduction in write endurance, and reduction in write bandwidth. To provide the benefit of secure memory in a write efficient manner this paper proposes Dual Counter Encryption (DEUCE). DEUCE is based on the observation that a typical writeback only changes a few words, so DEUCE reencrypts only the words that have changed. We show that DEUCE reduces the number of modified bits per writeback for a secure memory from 50% to 24%, which improves performance by 27% and increases lifetime by 2x.

Energy consumption is a primary consideration that determines the usability of emerging mobile computing devices such as smartphones. Refresh operations for main memory account for a significant fraction of the overall energy consumption, especially during idle periods, when processor can be switched off quickly, however, memory contents continue to get refreshed to avoid data loss. Given that mobile devices are idle most of the times, reducing refresh power in idle mode is critical to maximize the duration for which the device remains usable. The frequency of refresh operations in memory can be reduced significantly by using strong multi-bit error correction codes (ECC). Unfortunately, strong ECC codes incur high latency, which causes significant performance degradation (as high as 21%, and on average 10%). To obtain both low refresh power in idle periods and high performance in active periods, this paper proposes Morphable ECC (MECC). During idle periods, MECC keeps the memory protected with 6-bit ECC (ECC-6) and employs a refresh period of 1 second, instead of the typical refresh period of 64ms. During active operation, MECC reduces the refresh interval to 64ms, and converts memory from ECC-6 to weaker ECC (single-bit error correction) on a demand-basis, thus avoiding the high latency of ECC-6, except for the first access during the active mode. Our proposal reduces refresh operations during idle mode by 16x, memory power in idle mode by 2X, while retaining performance within 2% of a system that does not use any ECC.

Multirate refresh techniques exploit the non-uniformity in retention times of DRAM cells to reduce the DRAM refresh overheads. Such techniques rely on accurate profiling of retention times of cells, and perform faster refresh only for a few rows which have cells with low retention times. Unfortunately, retention times of some cells can change at runtime due to Variable Retention Time (VRT), which makes it impractical to reliably deploy multirate refresh. Based on experimental data from 24 DRAM chips, we develop architecture-level models for analyzing the impact of VRT. We show that simply relying on ECC DIMMs to correct VRT failures is unusable as it causes a data error once every few months. We propose AVATAR, a VRT-aware multirate refresh scheme that adaptively changes the refresh rate for different rows at runtime based on current VRT failures. AVATAR provides a time to failure in the regime of several tens of years while reducing refresh operations by 62%-72%.

Phase Change Memory (PCM) is an emerging memory technology that can enable scalable high-density main memory systems. Unfortunately, PCM has higher read latency than DRAM, resulting in lower system performance. This paper investigates architectural techniques to improve the read latency of PCM. We observe that there is a wide distribution in cell resistance in both the SET state and the RESET state, and that the read latency of PCM is designed conservatively to handle the worst case cell. If PCM sensing can be tuned to exploit the variability in cell resistance, then we can get reduced read latency. We propose two schemes to enable better-than-worst-case read latency for PCM systems. Our first proposal, Early Read, reads the data earlier than the specified time period. Our key observation that Early Read causes only unidirectional errors (SET being read as RESET) allows us to efficiently detect data errors using Berger codes. In the uncommon case that Early Read causes data error(s), we simply retry the read operation with original latency. Our evaluations show that Early Read can reduce the read latency by 25% while incurring a storage overhead of only 10 bits per 64 byte line. Our second proposal, Turbo Read, reduces the sensing time for read operations by pumping higher current, at the expense of accidentally switching the PCM cell with small probability during the read operation. We analyze Error Correction Codes (ECC) and Probabilistic Row Scrubbing (PRS) for maintaining data integrity under Turbo Read. We show that a combination of Early Read and Turbo Read can reduce the PCM read latency by 30%, improve the system performance by 21%, and reduce the Energy Delay Product (EDP) by 28%, while requiring minimal changes to the memory system.

Stacked memory modules are likely to be tightly integrated with the processor. It is vital that these memory modules operate reliably, as memory failure can require the replacement of the entire socket. To make matters worse, stacked memory designs are susceptible to newer failure modes (for example, due to faulty through-silicon vias, or TSVs) that can cause large portions of memory, such as a bank, to become faulty. To avoid data loss from large-granularity failures, the memory system may use symbol-based codes that stripe the data for a cache line across several banks (or channels). Unfortunately, such data-striping reduces memory level parallelism causing significant slowdown and higher power consumption. This paper proposes Citadel, a robust memory architecture that allows the memory system to retain each cache line within one bank, thus allowing high performance, lower power and efficiently protects the stacked memory from large-granularity failures. Citadel consists of three components, TSV-Swap, which can tolerate both faulty data-TSVs and faulty address-TSVs, Tri Dimensional Parity (3DP), which can tolerate column failures, row failures, and bank failures, and Dynamic Dual Granularity Sparing (DDS), which can mitigate permanent faults by dynamically sparing faulty memory regions either at a row granularity or at a bank granularity. Our evaluations with real-world data for DRAM failures show that Citadel provides performance and power similar to maintaining the entire cache line in the same bank, and yet provides 700x higher reliability than Chip Kill-like ECC codes.

Dynamic Random Access Memory (DRAM) cells rely on periodic refresh operations to maintain data integrity. As the capacity of DRAM memories has increased, so has the amount of time consumed in doing refresh. Refresh operations contend with read operations, which increases read latency and reduces system performance. We show that eliminating latency penalty due to refresh can improve average performance by 7.2%. However, simply doing intelligent scheduling of refresh operations is ineffective at obtaining significant performance improvement. This article provides an alternative and scalable option to reduce the latency penalty due to refresh. It exploits the property that each refresh operation in a typical DRAM device internally refreshes multiple DRAM rows in JEDEC-based distributed refresh mode. Therefore, a refresh operation has well-defined points at which it can potentially be Paused to service a pending read request. Leveraging this property, we propose Refresh Pausing , a solution that is highly effective at alleviating the contention from refresh operations. It provides an average performance improvement of 5.1% for 8Gb devices and becomes even more effective for future high-density technologies. We also show that Refresh Pausing significantly outperforms the recently proposed Elastic Refresh scheme.

DRAM cells rely on periodic refresh operations to maintain data integrity. As the capacity of DRAM memories has increased, so has the amount of time consumed in doing refresh. Refresh operations contend with read operations, which increases read latency and reduces system performance. We show that eliminating latency penalty due to refresh can improve average performance by 7.2%. However, simply doing intelligent scheduling of refresh operations is ineffective at obtaining significant performance improvement. This paper provides an alternative and scalable option to reduce the latency penalty due to refresh. It exploits the property that each refresh operation in a typical DRAM device internally refreshes multiple DRAM rows in JEDEC-based distributed refresh mode. Therefore, a refresh operation has well defined points at which it can potentially be Paused to service a pending read request. Leveraging this property, we propose Refresh Pausing, a solution that is highly effective at alleviating the contention from refresh operations. It provides an average performance improvement of 5.1% for 8Gb devices, and becomes even more effective for future high-density technologies. We also show that Refresh Pausing significantly outperforms the recently proposed Elastic Refresh scheme.

DRAM scaling has been the prime driver for increasing the capacity of main memory system over the past three decades. Unfortunately, scaling DRAM to smaller technology nodes has become challenging due to the inherent difficulty in designing smaller geometries, coupled with the problems of device variation and leakage. Future DRAM devices are likely to experience significantly high error-rates. Techniques that can tolerate errors efficiently can enable DRAM to scale to smaller technology nodes. However, existing techniques such as row/column sparing and ECC become prohibitive at high error-rates.