The increasing size of deep neural networks (DNNs) necessitates effective model compression to reduce their computational and memory footprints. Sparsity and quantization are two prominent compression methods that have been shown to reduce DNNs' computational and memory footprints significantly while preserving model accuracy. However, how these two methods interact when combined together remains a key question for developers, as many tacitly assume that they are orthogonal, meaning that their combined use does not introduce additional errors beyond those introduced by each method independently. In this paper, we provide the first mathematical proof that sparsity and quantization are non-orthogonal. We corroborate these results with experiments spanning a range of large language models, including the OPT and LLaMA model families (with 125M to 8B parameters), and vision models like ViT and ResNet. We show that the order in which we apply these methods matters because applying quantization before sparsity may disrupt the relative importance of tensor elements, which may inadvertently remove significant elements from a tensor. More importantly, we show that even if applied in the correct order, the compounded errors from sparsity and quantization can significantly harm accuracy. Our findings extend to the efficient deployment of large models in resource-constrained compute platforms to reduce serving cost, offering insights into best practices for applying these compression methods to maximize hardware resource efficiency without compromising accuracy.

Pruning eliminates unnecessary parameters in neural networks; it offers a promising solution to the growing computational demands of large language models (LLMs). While many focus on post-training pruning, sparse pre-training--which combines pruning and pre-training into a single phase--provides a simpler alternative. In this work, we present the first systematic exploration of optimal sparse pre-training configurations for LLMs through an examination of 80 unique pruning schedules across different sparsity levels and training durations. We find that initiating pruning at 25\% of total training compute and concluding at 75\% achieves near-optimal final evaluation loss. These findings provide valuable insights for efficient and effective sparse pre-training of LLMs. Furthermore, we propose a new scaling law that modifies the Chinchilla scaling law to use the average parameter count over pre-training. Through empirical and theoretical validation, we demonstrate that this modified scaling law accurately models evaluation loss for both sparsely and densely pre-trained LLMs, unifying scaling laws across pre-training paradigms. Our findings indicate that while sparse pre-training achieves the same final model quality as dense pre-training for equivalent compute budgets, it provides substantial benefits through reduced model size, enabling significant potential computational savings during inference.

We propose SLoPe, a Double-Pruned **S**parse Plus **L**azy L**o**w-rank Adapter **P**r**e**training method for LLMs that improves the accuracy of sparse LLMs while accelerating their pretraining and inference and reducing their memory footprint. Sparse pretraining of LLMs reduces the accuracy of the model, to overcome this, prior work uses dense models during fine-tuning. SLoPe improves the accuracy of sparsely pretrained models by adding low-rank adapters in the final 1% iterations of pretraining without adding significant overheads to the model pretraining and inference. In addition, SLoPe uses a double-pruned backward pass formulation that prunes the transposed weight matrix using N:M sparsity structures to enable an accelerated sparse backward pass. SLoPe accelerates the training and inference of models with billions of parameters up to 1.25× and 1.54× respectively (OPT-33B and OPT-66B) while reducing their memory usage by up to 0.63× and 0.61× for training and inference respectively.

Decoding with autoregressive language models traditionally occurs sequentially, generating one token after another. Recent attempts to introduce parallelism require a pre-determined structure in the generated content to implement parallel generation, such as by pattern-matching on bullet points. In this work, we present a new technique to automate parallel generation by dynamically exploiting the semantic independence of generation outputs to implement asynchronous decoding. We introduce an annotation language Pasta-Lang for language models to initiate asynchronous decoding at inference time. We also develop an accompanying Pasta-Lang interpreter that performs on-the-fly asynchronous decoding, effectively implementing parallel generation and speeding up inference. We present an instruction-finetuning dataset with Pasta-Lang-annotated responses for teaching LLMs to annotate semantic independence with Pasta-Lang as well as the methodology for creating the dataset. Our evaluation shows using the interpreter with a Pasta-Lang-equipped model achieves significant speedup while maintaining the same generation quality.

Fine-tuning LLMs is both computationally and memory-intensive. While parameter-efficient fine-tuning methods, such as QLoRA and DoRA, reduce the number of trainable parameters and lower memory usage, they do not decrease computational cost. In some cases, they may even slow down fine-tuning. In this paper, we introduce SparseLoRA, a method that accelerates LLM fine-tuning through contextual sparsity. We propose a lightweight, training-free SVD sparsity estimator that dynamically selects a sparse subset of weights for loss and gradient computation. Also, we systematically analyze and address sensitivity across layers, tokens, and training steps. Our experimental results show that SparseLoRA reduces computational cost by up to $2.2\times$ and a measured speedup of up to $1.6\times$ while maintaining accuracy across various downstream tasks, including commonsense and arithmetic reasoning, code generation, and instruction following.

Conventional model compression techniques for LLMs address high memory consumption and slow inference challenges but typically require computationally expensive retraining to preserve accuracy. In contrast, one-shot compression methods eliminate retraining cost, but struggle to achieve accuracy comparable to dense models. This paper presents SLIM, a new one-shot compression framework that holistically integrates hardware-friendly quantization, sparsity, and low-rank approximation into a unified process. First, we formulate the quantization process using a probabilistic approach (SLIM-Quant) that enables us to apply uniform quantization. Then, we use an existing one-shot pruning method to apply semi-structured sparsity on top of the quantized weights. Finally, to compensate for the introduced aggregated quantization and sparsity error, we use a novel saliency function with unique invertible and additive features that enables us to mathematically compute the value of low-rank adapters. SLIM improves model accuracy by up to 5.66% (LLaMA-2-7B) for 2:4 sparsity with 4-bit weight quantization, outperforming prior methods. Models compressed with SLIM achieve up to 4.3× and 3.8× on Nvidia RTX3060 and A100 GPUs, respectively. Additionally, they achieve up to 0.23× end-to-end memory reduction in comparison to their dense counterparts. We also propose an optional PEFT recipe that further improves accuracy by up to 1.66% (LLaMA-2-13B) compared to SLIM without fine-tuning.

Retrieval-augmented generation (RAG) is emerging as a popular approach for reliable LLM serving.However, efficient RAG serving remains an open challenge due to the rapid emergence of many RAG variants and the substantial differences in workload characteristics across them.This paper makes three fundamental contributions to advancing RAG serving.First, we introduce RAGSchema, a structured abstraction that captures the wide range of RAG algorithms, serving as a foundation for performance optimization.Second, we analyze several representative RAG workloads with distinct RAGSchema, revealing significant performance variability across these workloads.Third, to address this variability and meet diverse performance requirements, we propose RAGO (Retrieval-Augmented Generation Optimizer), a system optimization framework for efficient RAG serving.RAGO achieves up to a 2ˆincrease in QPS per chip and a 55% reduction in time-to-first-token latency compared to RAG systems built on LLM-system extensions.

The limited memory capacity of single GPUs constrains large language model (LLM) inference, necessitating cost-prohibitive multi-GPU deployments or frequent performance-limiting CPU-GPU transfers over slow PCIe.In this work, we first benchmark recent Intel CPUs with Advanced Matrix Extensions (AMX), including 4th generation (Sapphire Rapids) and 6th generation (Granite Rapids) Xeon Scalable Processors, demonstrating matrix multiplication throughput of 20 TFLOPS and 40 TFLOPS, respectivelycomparable to some recent GPUs.These findings unlock more extensive computation offloading to CPUs, reducing CPU-GPU transfers and alleviating throughput bottlenecks compared to priorgeneration CPUs.Building on these insights, we design LIA, a single-GPU LLM inference acceleration framework leveraging cooperative AMX-enabled CPU-GPU computation and CXL offloading.LIA systematically offloads computation to CPUs, optimizing both latency and throughput.The framework also introduces a memoryoffloading policy that seamlessly integrates affordable CXL memory with DDR memory to enhance performance in throughput-driven tasks.On Saphhire Rapids (Granite Rapids) systems with a single H100 GPU, LIA achieves up to 5.1× (19×) lower latency and 3.7× (5.1×) higher throughput compared to the latest single-GPU offloading framework.Furthermore, LIA deploying CXL offloading yields an additional 1.5× throughput improvement over LIA using only DDR memory with a 1.8× increase in maximum batch size (900→1.6K).

Cycle-level simulators such as gem5 are widely used in microarchitecture design, but they are prohibitively slow for large-scale design space explorations.We present Concorde, a new methodology for learning fast and accurate performance models of microarchitectures.Unlike existing simulators and learning approaches that emulate each instruction, Concorde predicts the behavior of a program based on compact performance distributions that capture the impact of different microarchitectural components.It derives these performance distributions using simple analytical models that estimate bounds on performance induced by each microarchitectural component, providing a simple yet rich representation of a program's performance characteristics across a large space of microarchitectural parameters.Experiments show that Concorde is more than five orders of magnitude faster than a reference cycle-level simulator, with about 2% average Cycles-Per-Instruction (CPI) prediction error across a range of SPEC, open-source, and proprietary benchmarks.This enables rapid design-space exploration and performance sensitivity analyses that are currently infeasible, e.g., in about an hour, we conducted a first-of-its-kind fine-grained performance attribution to different microarchitectural components across a diverse set of programs, requiring nearly 150 million CPI evaluations.

With the ever increasing prevalence of neural networks and the upheaval from the language models, it is time to rethink neural acceleration. Up to this point, the broader research community, including ourselves, has disproportionately focused on GEneral Matrix Multiplication (GEMM) operations. The supporting argument was that the large majority of the neural operations are GEMM. This argument guided the research in Neural Processing Units (NPUs) for the last decade. However, scant attention was paid to non-GEMM operations and they are rather overlooked. As deep learning evolved and progressed, these operations have grown in diversity and also large variety of structural patterns have emerged that interweave them with the GEMM operations. However, conventional NPU designs have taken rather simplistic approaches by supporting these operations through either a number of dedicated blocks or fall back to general-purpose processors.

While (I) serverless computing is emerging as a popular form of cloud execution, datacenters are going through major changes: (II) storage dissaggregation in the system infrastructure level and (III) integration of domain-specific accelerators in the hardware level. Each of these three trends individually provide significant benefits; however, when combined the benefits diminish. On the convergence of these trends, the paper makes the observation that for serverless functions, the overhead of accessing dissaggregated storage overshadows the gains from accelerators. Therefore, to benefit from all these trends in conjunction, we propose In-Storage Domain-Specific Acceleration for Serverless Computing (dubbed DSCS-Serverless1). The idea contributes a server-less model that utilizes a programmable accelerator embedded within computational storage to unlock the potential of acceleration in disaggregated datacenters. Our results with eight applications show that integrating a comparatively small accelerator within the storage (DSCS-Serverless) that fits within the storage's power constraints (25 Watts), significantly outperforms a traditional disaggregated system that utilizes NVIDIA RTX 2080 Ti GPU (250 Watts). Further, the work highlights that disaggregation, serverless model, and the limited power budget for computation in storage device require a different design than the conventional practices of integrating microprocessors and FPGAs. This insight is in contrast with current practices of designing computational storage devices that are yet to address the challenges associated with the shifts in datacenters. In comparison with two such conventional designs that use ARM cores or a Xilinx FPGA, DSCS-Serverless provides 3.7× and 1.7× end-to-end application speedup, 4.3× and 1.9× energy reduction, and 3.2× and 2.3× better cost efficiency, respectively.

End-to-end automatic speech recognition (ASR) models have seen revolutionary quality gains with the recent development of large-scale universal speech models (USM). However, deploying these massive USMs is extremely expensive due to the enormous memory usage and computational cost. Therefore, model compression is an important research topic to fit USM-based ASR under budget in real-world scenarios. In this study, we propose a USM fine-tuning approach for ASR, with a low-bit quantization and N:M structured sparsity aware paradigm on the model weights, reducing the model complexity from parameter precision and matrix topology perspectives. We conducted extensive experiments with a 2-billion parameter USM on a large-scale voice search dataset to evaluate our proposed method. A series of ablation studies validate the effectiveness of up to int4 quantization and 2:4 sparsity. However, a single compression technique fails to recover the performance well under extreme setups including int2 quantization and 1:4 sparsity. By contrast, our proposed method can compress the model to have 9.4% of the size, at the cost of only 7.3% relative word error rate (WER) regressions. We also provided in-depth analyses on the results and discussions on the limitations and potential solutions, which would be valuable for future studies.

With the decline of Moore's law, optimizing program performance has become a major focus of software research. However, high-level optimizations such as API and algorithm changes remain elusive due to the difficulty of understanding the semantics of code. Simultaneously, pretrained large language models (LLMs) have demonstrated strong capabilities at solving a wide range of programming tasks. To that end, we introduce a framework for adapting LLMs to high-level program optimization. First, we curate a dataset of performance-improving edits made by human programmers of over 77,000 competitive C++ programming submission pairs, accompanied by extensive unit tests. A major challenge is the significant variability of measuring performance on commodity hardware, which can lead to spurious "improvements." To isolate and reliably evaluate the impact of program optimizations, we design an environment based on the gem5 full system simulator, the de facto simulator used in academia and industry. Next, we propose a broad range of adaptation strategies for code optimization; for prompting, these include retrieval-based few-shot prompting and chain-of-thought, and for finetuning, these include performance-conditioned generation and synthetic data augmentation based on self-play. A combination of these techniques achieves a mean speedup of 6.86$\times$ with eight generations, higher than average optimizations from individual programmers (3.66$\times$). Using our model's fastest generations, we set a new upper limit on the fastest speedup possible for our dataset at 9.64$\times$ compared to using the fastest human submissions available (9.56$\times$).

Autoregressive Large Language Models (LLMs) have achieved impressive performance in language tasks but face two significant bottlenecks: (1) quadratic complexity in the attention module as the number of tokens increases, and (2) limited efficiency due to the sequential processing nature of autoregressive LLMs during generation. While linear attention and speculative decoding offer potential solutions, their applicability and synergistic potential for enhancing autoregressive LLMs remain uncertain. We conduct the first comprehensive study on the efficacy of existing linear attention methods for autoregressive LLMs, integrating them with speculative decoding. We introduce an augmentation technique for linear attention that ensures compatibility with speculative decoding, enabling more efficient training and serving of LLMs. Extensive experiments and ablation studies involving seven existing linear attention models and five encoder/decoder-based LLMs consistently validate the effectiveness of our augmented linearized LLMs. Notably, our approach achieves up to a 6.67 reduction in perplexity on the LLaMA model and up to a 2$\times$ speedup during generation compared to prior linear attention methods. Codes and models are available at https://github.com/GATECH-EIC/Linearized-LLM.

Deep neural network (DNN) video analytics is crucial for autonomous systems such as self-driving vehicles, unmanned aerial vehicles (UAVs), and security robots. However, real-world deployment faces challenges due to their limited computational resources and battery power. To tackle these challenges, continuous learning exploits a lightweight “student” model at deployment (inference), leverages a larger “teacher” model for labeling sampled data (labeling), and continuously retrains the student model to adapt to changing scenarios (retraining). This paper highlights the limitations in state-of-theart continuous learning systems: (1) they focus on computations for retraining, while overlooking the compute needs for inference and labeling, (2) they rely on power-hungry GPUs, unsuitable for battery-operated autonomous systems, and (3) they are located on a remote centralized server, intended for multi-tenant scenarios, again unsuitable for autonomous systems due to privacy, network availability, and latency concerns. We propose a hardwarealgorithm co-designed solution for continuous learning, DACAPO, that enables autonomous systems to perform concurrent executions of inference, labeling, and retraining in a performant and energy-efficient manner. DACapo comprises (1) a spatiallypartitionable and precision-flexible accelerator enabling parallel execution of kernels on sub-accelerators at their respective precisions, and (2) a spatiotemporal resource allocation algorithm that strategically navigates the resource-accuracy tradeoff space, facilitating optimal decisions for resource allocation to achieve maximal accuracy. Our evaluation shows that DACAPO achieves $\mathbf{6. 5 \%}$ and $\mathbf{5. 5 \%}$ higher accuracy than a state-of-theart GPU-based continuous learning systems, Ekya and EOMU, respectively, while consuming $254 \times$ less power.

Automatic translation of programming languages has garnered renewed interest, driven by recent advancements in large language models (LLMs). Encoder-decoder transformer models, in particular, have shown promise in translating between different programming languages. However, translating between a language and its high-performance computing (HPC) extension remains underexplored due to inherent challenges like complex parallel semantics understanding. In this paper, we introduce CodeRosetta, an encoder-decoder transformer model explicitly designed for translating between programming languages and also their HPC extensions. CodeRosetta is evaluated on C++ to CUDA and Fortran to C++ translation. It employs a customized learning-based framework with tailored pretraining and training objectives that enable it to effectively capture code semantics and parallel structural nuances, allowing for bidirectional code translation. Our results show that CodeRosetta outperforms state-of-the-art baselines in C++ to CUDA translation by 2.9 BLEU and 1.72 CodeBLUE points while improving compilation accuracy by 6.05%. Compared to general closed-source LLMs, our proposed bidirectional learning-based method improves C++ to CUDA translation by 22.08 BLEU and 14.39 CodeBLUE with 2.75% higher compilation accuracy. Finally, CodeRosetta exhibits proficiency in Fortran to parallel C++ translation, marking it, to our knowledge, as the first encoder-decoder model for such a complex translation task, improving CodeBLEU at least by 4.63 points compared to closed-source LLMs and Open Code LLM.

Large language models (LLMs) have shown impressive performance on language tasks but face challenges when deployed on resource-constrained devices due to their extensive parameters and reliance on dense multiplications, resulting in high memory demands and latency bottlenecks. Shift-and-add reparameterization offers a promising solution by replacing costly multiplications with hardware-friendly primitives in both the attention and multi-layer perceptron (MLP) layers of an LLM. However, current reparameterization techniques require training from scratch or full parameter fine-tuning to restore accuracy, which is resource-intensive for LLMs. To address this, we propose accelerating pretrained LLMs through post-training shift-and-add reparameterization, creating efficient multiplication-free models, dubbed ShiftAddLLM. Specifically, we quantize each weight matrix into binary matrices paired with group-wise scaling factors. The associated multiplications are reparameterized into (1) shifts between activations and scaling factors and (2) queries and adds according to the binary matrices. To reduce accuracy loss, we present a multi-objective optimization method to minimize both weight and output activation reparameterization errors. Additionally, based on varying sensitivity across layers to reparameterization, we develop an automated bit allocation strategy to further reduce memory usage and latency. Experiments on five LLM families and eight tasks consistently validate the effectiveness of ShiftAddLLM, achieving average perplexity reductions of 5.6 and 22.7 points at comparable or lower latency compared to the most competitive quantized LLMs at 3- and 2-bit precision, respectively, and more than 80% memory and energy reductions over the original LLMs. Codes and models are available at https://github.com/GATECH-EIC/ShiftAddLLM.

Attention mechanisms, primarily designed to capture pairwise correlations between words, have become the backbone of machine learning, expanding beyond natural language processing into other domains. This growth in adaptation comes at the cost of prohibitively large memory requirements and computational complexity, especially at higher number of input elements. This limitation is due to inherently limited data reuse opportunities and quadratic growth in memory footprints, leading to severe memory-boundedness and limited scalability of input elements. This work addresses these challenges by devising a tailored dataflow optimization, called FLAT, for attention mechanisms without altering their functionality. This dataflow processes costly attention operations through a unique fusion mechanism, transforming the memory footprint quadratic growth to merely a linear one. To realize the full potential of this bespoke mechanism, we propose a tiling approach to enhance the data reuse across attention operations. Our method both mitigates the off-chip bandwidth bottleneck as well as reduces the on-chip memory requirement. FLAT delivers 1.94x (1.76x) speedup and 49% and (42%) of energy savings compared to the state-of-the-art Edge (Cloud) accelerators with no customized dataflow optimization. When on-chip resources are scarce (20 KB-200 KB), FLAT yields, on average, 1.5x end-to-end latency reduction across a diverse range of conventional attention-based models with input sequence lengths ranging from 512-token to 64K-token. Our evaluations demonstrate that state-of-the-art DNN dataflow applied to attention operations reach the efficiency limit for inputs above 512 elements. In contrast, FLAT unblocks transformer models for inputs with up to 64K elements.

Machine learning driven computer architecture tools and methods have the potential to drastically shape the future of computer architecture. The question is: how can we lay the foundation to effectively usher in this era? In this post, we delve into the transformative impact of machine learning (ML) on the research landscape, emphasizing the importance of understanding both its potential and pitfalls to fully support ML-assisted computer architecture research. By exploring these advancements, our aim is to highlight the opportunities that lie ahead and outline the collective steps that we, as a community, can take towards realizing the era of "Architecture 2.0."

Recent innovations on hardware (e.g. Nvidia A100) have motivated learning N:M structured sparsity masks from scratch for fast model inference. However, state-of-the-art learning recipes in this regime (e.g. SR-STE) are proposed for non-adaptive optimizers like momentum SGD, while incurring non-trivial accuracy drop for Adam-trained models like attention-based LLMs. In this paper, we first demonstrate such gap origins from poorly estimated second moment (i.e. variance) in Adam states given by the masked weights. We conjecture that learning N:M masks with Adam should take the critical regime of variance estimation into account. In light of this, we propose STEP, an Adam-aware recipe that learns N:M masks with two phases: first, STEP calculates a reliable variance estimate (precondition phase) and subsequently, the variance remains fixed and is used as a precondition to learn N:M masks (mask-learning phase). STEP automatically identifies the switching point of two phases by dynamically sampling variance changes over the training trajectory and testing the sample concentration. Empirically, we evaluate STEP and other baselines such as ASP and SR-STE on multiple tasks including CIFAR classification, machine translation and LLM fine-tuning (BERT-Base, GPT-2). We show STEP mitigates the accuracy drop of baseline recipes and is robust to aggressive structured sparsity ratios.

Machine learning (ML) has become a prevalent approach to tame the complexity of design space exploration for domain-specific architectures. While appealing, using ML for design space exploration poses several challenges. First, it is not straightforward to identify the most suitable algorithm from an ever-increasing pool of ML methods. Second, assessing the trade-offs between performance and sample efficiency across these methods is inconclusive. Finally, the lack of a holistic framework for fair, reproducible, and objective comparison across these methods hinders the progress of adopting ML-aided architecture design space exploration and impedes creating repeatable artifacts. To mitigate these challenges, we introduce ArchGym, an open-source gymnasium and easy-to-extend framework that connects a diverse range of search algorithms to architecture simulators. To demonstrate its utility, we evaluate ArchGym across multiple vanilla and domain-specific search algorithms in the design of a custom memory controller, deep neural network accelerators, and a custom SoC for AR/VR workloads, collectively encompassing over 21K experiments. The results suggest that with an unlimited number of samples, ML algorithms are equally favorable to meet the user-defined target specification if its hyperparameters are tuned thoroughly; no one solution is necessarily better than another (e.g., reinforcement learning vs. Bayesian methods). We coin the term "hyperparameter lottery" to describe the relatively probable chance for a search algorithm to find an optimal design provided meticulously selected hyperparameters. Additionally, the ease of data collection and aggregation in ArchGym facilitates research in ML-aided architecture design space exploration. As a case study, we show this advantage by developing a proxy cost model with an RMSE of 0.61% that offers a 2,000-fold reduction in simulation time. Code and data for ArchGym is available at https://bit.ly/ArchGym.

Modern heterogeneous CPUs incorporate hardware accelerators to enable domain-specialized execution and achieve improved efficiency. A well-known class among them, spatial accelerators, are designed with reconfigurability to accelerate a wide range of compute-heavy and data-parallel applications. Unlike CPU cores, however, they tend to require specialized compilers and software stacks, libraries, or languages to operate and cannot be utilized with ease by all applications. As a result, the accelerator's large pool of compute and memory resources sit wastefully idle when it is not explicitly programmed. Our goal is to dismantle this CPU-accelerator barrier by monitoring CPU threads for acceleration opportunities during execution and, if viable, dynamically reconfigure the accelerator to allow transparent offloading. We develop MESA (Microarchitecture Extensions for Spatial Architecture Generation), a hardware block on the CPU that translates machine code to build an accelerator configuration specialized for the running program. While such a dynamic translation/reconfiguration approach is challenging, it has a key advantage over ahead-of-time compilers: access to runtime information, revealing not only dynamic dependencies but also performance characteristics. MESA maintains a real-time performance model of the program mapped on the accelerator in the form of a spatial dataflow graph with nodes weighted by operation latency and edges weighted by data transfer latency. Features of this dataflow graph are continuously updated with runtime information captured by performance counters, allowing a feedback loop of optimization, reconfiguration, and acceleration. This performance model allows MESA to identify the accelerator's critical paths and pinpoint its bottlenecks, upon which we implement in hardware a data-driven instruction mapping algorithm that locally minimizes latency. Backed by a synthesized RTL implementation, we evaluate the feasibility of our microarchitectural solution with different accelerator configurations. Across the Rodinia benchmarks, results demonstrate an average 1.3× speedup in performance and 1.8× gain in energy efficiency against a multicore CPU baseline.

Like humans, large language models (LLMs) do not always generate the best output on their first try. Motivated by how humans refine their written text, we introduce Self-Refine, an approach for improving initial outputs from LLMs through iterative feedback and refinement. The main idea is to generate an initial output using an LLMs; then, the same LLMs provides *feedback* for its output and uses it to *refine* itself, iteratively. Self-Refine does not require any supervised training data, additional training, or reinforcement learning, and instead uses a single LLM as the generator, refiner and the feedback provider. We evaluate Self-Refine across 7 diverse tasks, ranging from dialog response generation to mathematical reasoning, using state-of-the-art (GPT-3.5, ChatGPT, and GPT-4) LLMs. Across all evaluated tasks, outputs generated with Self-Refine are preferred by humans and automatic metrics over those generated with the same LLM using conventional one-step generation, improving by $\sim$20\% absolute on average in task performance. Our work demonstrates that even state-of-the-art LLMs like GPT-4 can be further improved at test-time using our simple, standalone approach.

Self-attention is a key enabler of state-of-art accuracy for various transformer-based Natural Language Processing models. This attention mechanism calculates a correlation score for each word with respect to the other words in a sentence. Commonly, only a small subset of words highly correlates with the word under attention, which is only determined at runtime. As such, a significant amount of computation is inconsequential due to low attention scores and can potentially be pruned. The main challenge is finding the threshold for the scores below which subsequent computation will be inconsequential. Although such a threshold is discrete, this paper formulates its search through a soft differentiable regularizer integrated into the loss function of the training. This formulation piggy backs on the back-propagation training to analytically co-optimize the threshold and the weights simultaneously, striking a formally optimal balance between accuracy and computation pruning. To best utilize this mathematical innovation, we devise a bit-serial architecture, dubbed LeOPArd, for transformer language models with bit-level early termination microarchitectural mechanism. We evaluate our design across 43 back-end tasks for MemN2N, BERT, ALBERT, GPT-2, and Vision transformer models. Post-layout results show that, on average, LeOPArd yields 1.9×and 3.9×speedup and energy reduction, respectively, while keeping the average accuracy virtually intact (< 0.2% degradation).

As its core computation, a self-attention mechanism gauges pairwise correlations across the entire input sequence. Despite favorable performance, calculating pairwise correlations is prohibitively costly. While recent work has shown the benefits of runtime pruning of elements with low attention scores, the quadratic complexity of self-attention mechanisms and their on-chip memory capacity demands are overlooked. This work addresses these constraints by architecting an accelerator, called SPRINT1, which leverages the inherent parallelism of ReRAM crossbar arrays to compute attention scores in an approximate manner. Our design prunes the low attention scores using a lightweight analog thresholding circuitry within ReRAM, enabling SPRINT to fetch only a small subset of relevant data to on-chip memory. To mitigate potential negative repercussions for model accuracy, SPRINT re-computes the attention scores for the few fetched data in digital. The combined in-memory pruning and on-chip recompute of the relevant attention scores enables SPRINT to transform quadratic complexity to a merely linear one. In addition, we identify and leverage a dynamic spatial locality between the adjacent attention operations even after pruning, which eliminates costly yet redundant data fetches. We evaluate our proposed technique on a wide range of state-of-the-art transformer models. On average, SPRINT yields 7.5× speedup and 19.6× energy reduction when total l6KB on-chip memory is used, while virtually on par with iso-accuracy of the baseline models (on average 0.36% degradation).

Albeit low-power, mixed-signal circuitry suffers from significant overhead of Analog to Digital (A/D) conversion, limited range for information encoding, and susceptibility to noise. This paper aims to address these challenges by offering and leveraging the following mathematical insight regarding vector dot-product---the basic operator in Deep Neural Networks (DNNs). This operator can be reformulated as a wide regrouping of spatially parallel low-bitwidth calculations that are interleaved across the bit partitions of multiple elements of the vectors. As such, the computational building block of our accelerator becomes a wide bit-interleaved analog vector unit comprising a collection of low-bitwidth multiply-accumulate modules that operate in the analog domain and share a single A/D converter(ADC). This bit-partitioning results in a lower-resolution ADC while the wide regrouping alleviates the need for A/D conversion per operation, amortizing its cost across multiple bit-partitions of the vector elements. Moreover, the low-bitwidth modules require smaller encoding range and also provide larger margins for noise mitigation. We also utilize the switched-capacitor design for our bit-level reformulation of DNN operations. The proposed switched-capacitor circuitry performs the regrouped multiplications in the charge domain and accumulates the results of the group in its capacitors over multiple cycles. The capacitive accumulation combined with wide bit-partitioned regrouping reduces the rate of A/D conversions, further improving the overall efficiency of the design.

Historically, continuous improvements in general-purpose processors have fueled the economic success and growth of the IT industry. However, the diminishing benefits from transistor scaling and conventional optimization techniques necessitates moving beyond common practices. Approximate computing is one such unconventional technique that has shown promise in pushing the boundaries of general-purpose processing. This paper sets out to employ approximation for processors that are commonly used in cyber-physical domains and may become building blocks of Internet of Things. To this end, we propose AxMemo to exploit the computation redundancy that stems from data similarity in the inputs of code blocks. Such input behavior is prevalent in cyber-physical systems as they deal with real-world data that naturally harbors redundancy. Therefore, in contrast to existing memoization techniques that replace costly floating-point arithmetic operations with limited number of inputs, AxMemo focuses on memoizing blocks of code with potentially many inputs. As such, AxMemo aims to replace long sequences of instructions with a few hash and lookup operations. By reducing the number of dynamic instructions, AxMemo alleviates the von Neumann and execution overheads of passing instructions through the processor pipeline altogether. The challenge AxMemo facing is to provide low-cost hashing mechanisms that can generate rather unique signature for each multi-input combination. To address this challenge, we develop a novel use of Cyclic Redundancy Checking (CRC) to hash the inputs. To increase lookup table hit rate, AxMemo employs a two-level memoization lookup, which utilizes small dedicated SRAM and spare storage in the last level cache. These solutions enable AxMemo to efficiently memoize relatively large code regions with variable input sizes and types using the same underlying hardware. Our experiment shows that AxMemo offers 2.64× speedup and 2.58 × energy reduction with mere 0.2% of quality loss averaged across ten benchmarks. These benefits come with an area overhead of just 2.1%.

GPUs are bottlenecked by the off-chip communication bandwidth and its energy cost; hence near-data acceleration is particularly attractive for GPUs. Integrating the accelerators within DRAM can mitigate these bottlenecks and additionally expose them to the higher internal bandwidth of DRAM. However, such an integration is challenging, as it requires low-overhead accelerators while supporting a diverse set of applications. To enable the integration, this work leverages the approximability of GPU applications and utilizes the neural transformation, which converts diverse regions of code mainly to Multiply-Accumulate (MAC). Furthermore, to preserve the SIMT execution model of GPUs, we also propose a novel approximate MAC unit with a significantly smaller area overhead. As such, this work introduces AxRam---a novel DRAM architecture---that integrates several approximate MAC units. AxRam offers this integration without increasing the memory column pitch or modifying the internal architecture of the DRAM banks. Our results with 10 GPGPU benchmarks show that, on average, AxRam provides 2.6× speedup and 13.3× energy reduction over a baseline GPU with no acceleration. These benefits are achieved while reducing the overall DRAM system power by 26% with an area cost of merely 2.1%.

Generative Adversarial Networks (GANs) are a frontier in deep learning. GANs consist of two models: generative and discriminative. While the discriminative model uses the conventional convolution, the generative model depends on a fundamentally different operator, called transposed convolution. This operator initially inserts a large number of zeros in its input and then slides a window over this expanded input. This zero-insertion step leads to a large number of ineffectual operations and creates distinct patterns of computation across the sliding windows. The ineffectual operations along with the variation in computation patterns lead to significant resource underutilization when using conventional convolution hardware. To alleviate these sources of inefficiency, this paper devises FlexiGAN, an end-to-end solution, that generates an optimized synthesizable FPGA accelerator from a high-level GAN specification. FlexiGAN is coupled with a novel template architecture that aims to harness the benefits of both MIMD and SIMD execution models to avoid ineffectual operations. To this end, the proposed architecture separates data retrieval and data processing units at the finest granularity of each compute engine. Leveraging this separation enables the architecture to use a succinct set of operations to cope with the irregularities of transposed convolution. At the same time, it significantly reduces the on-chip memory usage, which is generally limited in FPGAs. We evaluate our end-to-end solution by generating FPGA accelerators for a variety of GANs. These generated accelerators provide 2.4× higher performance than an optimized conventional convolution design. In addition, FlexiGAN, on average, yields 2.8× (up to 3.7×) improvements in Performance-per-Watt over a Titan X GPU.

Deep Convolutional Neural Networks (CNNs) perform billions of operations for classifying a single input. To reduce these computations, this paper offers a solution that leverages a combination of runtime information and the algorithmic structure of CNNs. Specifically, in numerous modern CNNs, the outputs of compute-heavy convolution operations are fed to activation units that output zero if their input is negative. By exploiting this unique algorithmic property, we propose a predictive early activation technique, dubbed SnaPEA. This technique cuts the computation of convolution operations short if it determines that the output will be negative. SnaPEA can operate in two distinct modes, exact and predictive. In the exact mode, with no loss in classification accuracy, SnaPEA statically re-orders the weights based on their signs and periodically performs a single-bit sign check on the partial sum. Once the partial sum drops below zero, the rest of computations can simply be ignored, since the output value will be zero in any case. In the predictive mode, which trades the classification accuracy for larger savings, SnaPEA speculatively cuts the computation short even earlier than the exact mode. To control the accuracy, we develop a multi-variable optimization algorithm that thresholds the degree of speculation. As such, the proposed algorithm exposes a knob to gracefully navigate the trade-offs between the classification accuracy and computation reduction. Compared to a state-of-the-art CNN accelerator, SnaPEA in the exact mode, yields, on average, 28% speedup and 16% energy reduction in various modern CNNs without affecting their classification accuracy. With 3% loss in classification accuracy, on average, 67.8% of the convolutional layers can operate in the predictive mode. The average speedup and energy saving of these layers are 2.02x and 1.89x, respectively. The benefits grow to a maximum of 3.59x speedup and 3.14x energy reduction. Compared to static pruning approaches, which are complimentary to the dynamic approach of SnaPEA, our proposed technique offers up to 63% speedup and 49% energy reduction across the convolution layers with no loss in classification accuracy.

Generative Adversarial Networks (GANs) are one of the most recent deep learning models that generate synthetic data from limited genuine datasets. GANs are on the frontier as further extension of deep learning into many domains (e.g., medicine, robotics, content synthesis) requires massive sets of labeled data that is generally either unavailable or prohibitively costly to collect. Although GANs are gaining prominence in various fields, there are no accelerators for these new models. In fact, GANs leverage a new operator, called transposed convolution, that exposes unique challenges for hardware acceleration. This operator first inserts zeros within the multidimensional input, then convolves a kernel over this expanded array to add information to the embedded zeros. Even though there is a convolution stage in this operator, the inserted zeros lead to underutilization of the compute resources when a conventional convolution accelerator is employed. We propose the GANAX architecture to alleviate the sources of inefficiency associated with the acceleration of GANs using conventional convolution accelerators, making the first GAN accelerator design possible. We propose a reorganization of the output computations to allocate compute rows with similar patterns of zeros to adjacent processing engines, which also avoids inconsequential multiply-adds on the zeros. This compulsory adjacency reclaims data reuse across these neighboring processing engines, which had otherwise diminished due to the inserted zeros. The reordering breaks the full SIMD execution model, which is prominent in convolution accelerators. Therefore, we propose a unified MIMD-SIMD design for GANAX that leverages repeated patterns in the computation to create distinct microprograms that execute concurrently in SIMD mode. The interleaving of MIMD and SIMD modes is performed at the granularity of single microprogrammed operation. To amortize the cost of MIMD execution, we propose a decoupling of data access from data processing in GANAX. This decoupling leads to a new design that breaks each processing engine to an access micro-engine and an execute micro-engine. The proposed architecture extends the concept of access-execute architectures to the finest granularity of computation for each individual operand. Evaluations with six GAN models shows, on average, 3.6x speedup and 3.1x energy savings over Eyeriss without compromising the efficiency of conventional convolution accelerators. These benefits come with a mere ≈7.8% area increase. These results suggest that GANAX is an effective initial step that paves the way for accelerating the next generation of deep neural models.

A growing number of commercial and enterprise systems increasingly rely on compute-intensive Machine Learning (ML) algorithms. While the demand for these compute-intensive applications is growing, the performance benefits from general-purpose platforms are diminishing. Field Programmable Gate Arrays (FPGAs) provide a promising path forward to accommodate the needs of machine learning algorithms and represent an intermediate point between the efficiency of ASICs and the programmability of general-purpose processors. However, acceleration with FPGAs still requires long development cycles and extensive expertise in hardware design. To tackle this challenge, instead of designing an accelerator for a machine learning algorithm, we present TABLA, a framework that generates accelerators for a class of machine learning algorithms. The key is to identify the commonalities across a wide range of machine learning algorithms and utilize this commonality to provide a high-level abstraction for programmers. TABLA leverages the insight that many learning algorithms can be expressed as a stochastic optimization problem. Therefore, learning becomes solving an optimization problem using stochastic gradient descent that minimizes an objective function over the training data. The gradient descent solver is fixed while the objective function changes for different learning algorithms. TABLA provides a template-based framework to accelerate this class of learning algorithms. Therefore, a developer can specify the learning task by only expressing the gradient of the objective function using our high-level language. Tabla then automatically generates the synthesizable implementation of the accelerator for FPGA realization using a set of hand-optimized templates. We use Tabla to generate accelerators for ten different learning tasks targeted at a Xilinx Zynq FPGA platform. We rigorously compare the benefits of FPGA acceleration to multi-core CPUs (ARM Cortex A15 and Xeon E3) and many-core GPUs (Tegra K1, GTX 650 Ti, and Tesla K40) using real hardware measurements. TABLA-generated accelerators provide 19.4x and 2.9x average speedup over the ARM and Xeon processors, respectively. These accelerators provide 17.57x, 20.2x, and 33.4x higher Performance-per-Watt in comparison to Tegra, GTX 650 Ti and Tesla, respectively. These benefits are achieved while the programmers write less than 50 lines of code.

Conventionally, an approximate accelerator replaces every invocation of a frequently executed region of code without considering the final quality degradation. However, there is a vast decision space in which each invocation can either be delegated to the accelerator -- improving performance and efficiency -- or run on the precise core -- maintaining quality. In this paper we introduce MITHRA, a co-designed hardware-software solution, that navigates these tradeoffs to deliver high performance and efficiency while lowering the final quality loss. MITHRA seeks to identify whether each individual accelerator invocation will lead to an undesirable quality loss and, if so, directs the processor to run the original precise code. This identification is cast as a binary classification task that requires a cohesive co-design of hardware and software. The hardware component performs the classification at runtime and exposes a knob to the software mechanism to control quality tradeoffs. The software tunes this knob by solving a statistical optimization problem that maximizes benefits from approximation while providing statistical guarantees that final quality level will be met with high confidence. The software uses this knob to tune and train the hardware classifiers. We devise two distinct hardware classifiers, one table-based and one neural network based. To understand the efficacy of these mechanisms, we compare them with an ideal, but infeasible design, the oracle. Results show that, with 95% confidence the table-based design can restrict the final output quality loss to 5% for 90% of unseen input sets while providing 2.5× speedup and 2.6× energy efficiency. The neural design shows similar speedup however, improves the efficiency by 13%. Compared to the table-based design, the oracle improves speedup by 26% and efficiency by 36%. These results show that MITHRA performs within a close range of the oracle and can effectively navigate the quality tradeoffs in approximate acceleration.

This article aims to tackle two fundamental memory bottlenecks: limited off-chip bandwidth (bandwidth wall) and long access latency (memory wall). To achieve this goal, our approach exploits the inherent error resilience of a wide range of applications. We introduce an approximation technique, called Rollback-Free Value Prediction (RFVP). When certain safe-to-approximate load operations miss in the cache, RFVP predicts the requested values. However, RFVP does not check for or recover from load-value mispredictions, hence, avoiding the high cost of pipeline flushes and re-executions. RFVP mitigates the memory wall by enabling the execution to continue without stalling for long-latency memory accesses. To mitigate the bandwidth wall, RFVP drops a fraction of load requests that miss in the cache after predicting their values. Dropping requests reduces memory bandwidth contention by removing them from the system. The drop rate is a knob to control the trade-off between performance/energy efficiency and output quality. Our extensive evaluations show that RFVP, when used in GPUs, yields significant performance improvement and energy reduction for a wide range of quality-loss levels. We also evaluate RFVP’s latency benefits for a single core CPU. The results show performance improvement and energy reduction for a wide variety of applications with less than 1% loss in quality.

This work presents a framework which detects online and at operand level of granularity all the vectors which excite a set of diagnosed failures in combinational modules. The failures may be of various types and may change over time. We propose to utilize this ability to detect failures at operand level of granularity to improve yield, by not discarding those chips containing failing and redundant computational units as long as they are not failing at the same time. The main challenge in realization of such a framework is the ability for on-chip storage of all the (test) vectors which excite the set of diagnosed failures. A major contribution of this work is to significantly minimize the number of stored test cubes by inserting only a few but carefully-selected "false alarm" vectors. As a result, a computational unit may be mis-diagnosed as failing for a given operand however we show such cases are rare and the chip may continue to be used.

Graphics Processing Units (GPUs) can accelerate diverse classes of applications, such as recognition, gaming, data analytics, weather prediction, and multimedia. Many of these applications are amenable to approximate execution. This application characteristic provides an opportunity to improve GPU performance and efficiency. Among approximation techniques, neural accelerators have been shown to provide significant performance and efficiency gains when augmenting CPU processors. However, the integration of neural accelerators within a GPU processor has remained unexplored. GPUs are, in a sense, many-core accelerators that exploit large degrees of data-level parallelism in the applications through the SIMT execution model. This paper aims to harmoniously bring neural and GPU accelerators together without hindering SIMT execution or adding excessive hardware overhead. We introduce a low overhead neurally accelerated architecture for GPUs, called NGPU, that enables scalable integration of neural accelerators for large number of GPU cores. This work also devises a mechanism that controls the tradeoff between the quality of results and the benefits from neural acceleration. Compared to the baseline GPU architecture, cycle-accurate simulation results for NGPU show a 2.4× average speedup and a 2.8× average energy reduction within 10% quality loss margin across a diverse set of benchmarks. The proposed quality control mechanism retains a 1.9× average speedup and a 2.1× energy reduction while reducing the degradation in the quality of results to 2.5%. These benefits are achieved by less than 1% area overhead.

This paper demonstrates how to utilize the inherent error resilience of a wide range of applications to mitigate the memory wall -- the discrepancy between core and memory speed. We define a new microarchitecturally-triggered approximation technique called rollback-free value prediction. This technique predicts the value of safe-to-approximate loads when they miss in the cache without tracking mispredictions or requiring costly recovery from misspeculations. This technique mitigates the memory wall by allowing the core to continue computation without stalling for long-latency memory accesses. Our detailed study of the quality trade-offs shows that with a modern out-of-order processor, average 8% (up to 19%) performance improvement is possible with 0.8% (up to 1.8%) average quality loss on an approximable subset of SPEC CPU 2000/2006.

As improvements in per-transistor speed and energy efficiency diminish, radical departures from conventional approaches are becoming critical to improving the performance and energy efficiency of general-purpose processors. We propose a solution—from circuit to compiler—that enables general-purpose use of limited-precision, analog hardware to accelerate “approximable” code—code that can tolerate imprecise execution. We utilize an algorithmic transformation that automatically converts approximable regions of code from a von Neumann model to an “analog” neural model. We outline the challenges of taking an analog approach, including restricted-range value encoding, limited precision in computation, circuit inaccuracies, noise, and constraints on supported topologies. We address these limitations with a combination of circuit techniques, a hardware/software interface, neural-network training techniques, and compiler support. Analog neural acceleration provides whole application speedup of 3.7× and energy savings of 6.3× with quality loss less than 10% for all except one benchmark. These results show that using limited-precision analog circuits for code acceleration, through a neural approach, is both feasible and beneficial over a range of approximation-tolerant, emerging applications including financial analysis, signal processing, robotics, 3D gaming, compression, and image processing.

Shrinking feature sizes, reduced voltages, and higher transistor count of nano-scale silicon chips challenge designers in terms of performance, power consumption, and reliability. This paper investigates the effect of simultaneous use of dynamic voltage and frequency scaling (DVFS) and body biasing (BB) on power consumption, reliability, and performance. An analytical model of reliability as a function of body bias voltage, supply voltage, and frequency is proposed. We derive a three dimensional optimization problem by exploiting proposed reliability model in conjunction with power consumption and performance model. The resulting problem is solved using widely-used geometric optimization to identify optimal supply voltage and body bias voltage and then is validated using accurate simulation. Afterwards, it is demonstrated how joint energy-performance-reliability space optimization method can be used in an adaptive reliability-aware power management systems. Finally, we show that combined soft error aware BB and DVFS is capable of improving power consumption about 30% in comparison to reliability-aware DVFS only for the same level of reliability and performance constraints.

Advances in silicon technology and shrinking the feature size to nanometer scale make unreliability of nano devices the most important concern of fault-tolerant designs. Soft error analysis has been greatly aided by the concept of architectural vulnerability factor (AVF) and architecturally correct execution (ACE). In this work, we exploit the techniques of AVF analysis to introduce the instruction-level vulnerability metric for software reliability analysis. The proposed metric can be used to make judgments about the reliability of different programs on different processors with regard to architectural and compiler guidelines for improving the processor reliability.