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.

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.

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 discovery of the *lazy neuron phenomenon* (Li et al., 2022), where fewer than 10% of the feedforward networks (FFN) parameters in trained Transformers are activated per token, has spurred significant interests in *activation sparsity* for enhancing large model efficiency. While notable progress has been made in translating such sparsity to wall-time benefits across CPUs, GPUs, and TPUs, modern Transformers have moved away from the ReLU activation function crucial to this phenomenon. Existing efforts on re-introducing activation sparsity, e.g., by reverting to ReLU or applying top-k masking, often degrade model quality, increase parameter count, or complicate training. Sparse attention, the application of sparse activation to the attention mechanism, often face similar challenges. This paper introduces the Spark Transformer, a novel architecture that achieves high activation sparsity in both FFN and the attention mechanism while maintaining model quality, parameter count, and standard training procedures. Our method realizes sparsity via top-$k$ masking for explicit control over sparsity level. Crucially, we introduce *statistical top-k*, a hardware-accelerator-friendly, linear-time approximate algorithm that avoids costly sorting and mitigates significant training slowdown from standard top-k operators. Furthermore, Spark Transformer reallocates existing FFN parameters and attention key embeddings to form a low-cost predictor for identifying activated entries. This design not only mitigates quality loss from enforced sparsity, but also enhances wall-time benefit. Pretrained with the Gemma-2 recipe, Spark Transformer demonstrates competitive performance on standard benchmarks while exhibiting significant sparsity: only 8\% of FFN neurons are activated, and each token attends to a maximum of 256 tokens. This translates to a 2.5x reduction in FLOPs, leading to decoding wall-time speedups of up to 1.79x on CPU and 1.40xon GPU.

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.

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.

In response to innovations in machine learning (ML) models, production workloads changed radically and rapidly. TPU v4 is the fifth Google domain specific architecture (DSA) and its third supercomputer for such ML models. Optical circuit switches (OCSes) dynamically reconfigure its interconnect topology to improve scale, availability, utilization, modularity, deployment, security, power, and performance; users can pick a twisted 3D torus topology if desired. Much cheaper, lower power, and faster than Infiniband, OCSes and underlying optical components are <5% of system cost and <3% of system power. Each TPU v4 includes SparseCores, dataflow processors that accelerate models that rely on embeddings by 5x--7x yet use only 5% of die area and power. Deployed since 2020, TPU v4 outperforms TPU v3 by 2.1x and improves performance/Watt by 2.7x. The TPU v4 supercomputer is 4x larger at 4096 chips and thus nearly 10x faster overall, which along with OCS flexibility and availability allows a large language model to train at an average of ~60% of peak FLOPS/second. For similar sized systems, it is ~4.3x--4.5x faster than the Graphcore IPU Bow and is 1.2x--1.7x faster and uses 1.3x--1.9x less power than the Nvidia A100. TPU v4s inside the energy-optimized warehouse scale computers of Google Cloud use ~2--6x less energy and produce ~20x less CO2e than contemporary DSAs in typical on-premise data centers.

Multicore systems should support both speculative and non-speculative parallelism. Speculative parallelism is easy to use and is crucial to scale many challenging applications, while non-speculative parallelism is more efficient and allows parallel irrevocable actions (e.g., parallel I/O). Unfortunately, prior techniques are far from this goal. Hardware transactional memory (HTM) systems support speculative (transactional) and non-speculative (non-transactional) work, but lack coordination mechanisms between the two, and are limited to unordered parallelism. Prior work has extended HTMs to avoid the limitations of speculative execution, e.g., through escape actions and open-nested transactions. But these mechanisms are incompatible with systems that exploit ordered parallelism, which parallelize a broader range of applications and are easier to use. We contribute two techniques that enable seamlessly composing and coordinating speculative and non-speculative work in the context of ordered parallelism: (i) a task-based execution model that efficiently coordinates concurrent speculative and non-speculative ordered tasks, allowing them to create tasks of either kind and to operate on shared data; and (ii) a safe way for speculative tasks to invoke software-managed speculative actions that avoid hardware version management and conflict detection. These contributions improve efficiency and enable new capabilities. Across several benchmarks, they allow the system to dynamically choose whether to execute tasks speculatively or non-speculatively, avoid needless conflicts among speculative tasks, and allow speculative tasks to safely invoke irrevocable actions.

This work studies the interplay between multithreaded cores and speculative parallelism (e.g., transactional memory or thread-level speculation). These techniques are often used together, yet they have been developed independently. This disconnect causes major performance pathologies: increasing the number of threads per core adds conflicts and wasted work, and puts pressure on speculative execution resources. These pathologies often squander the benefits of multithreading.We present speculation-aware multithreading (SAM), a simple policy that addresses these pathologies. By coordinating instruction dispatch and conflict resolution priorities, SAM focuses execution resources on work that is more likely to commit, avoiding aborts and using speculation resources more efficiently.We design SAM variants for in-order and out-of-order cores. SAM is cheap to implement and makes multithreaded cores much more beneficial on speculative parallel programs. We evaluate SAM on systems with up to 64 SMT cores. With SAM, 8-threaded cores outperform single-threaded cores by 2.33x on average, while a speculation-oblivious policy yields a 1.85x speedup. SAM also reduces wasted work by 52%.

Multicore systems must exploit locality to scale, scheduling tasks to minimize data movement. While locality-aware parallelism is well studied in non-speculative systems, it has received little attention in speculative systems (e.g., HTM or TLS), which hinders their scalability. We present spatial hints, a technique that leverages program knowledge to reveal and exploit locality in speculative parallel programs. A hint is an abstract integer, given when a speculative task is created, that denotes the data that the task is likely to access. We show it is easy to modify programs to convey locality through hints. We design simple hardware techniques that allow a state-of-the-art, tiled speculative architecture to exploit hints by: (i) running tasks likely to access the same data on the same tile, (ii) serializing tasks likely to conflict, and (iii) balancing tasks across tiles in a locality-aware fashion. We also show that programs can often be restructured to make hints more effective. Together, these techniques make speculative parallelism practical on large-scale systems: at 256 cores, hints achieve near-linear scalability on nine challenging applications, improving performance over hint-oblivious scheduling by 3.3× gmean and by up to 16×. Hints also make speculation far more efficient, reducing wasted work by 6.4× and traffic by 3.5× on average.

Switches today provide a small menu of scheduling algorithms. While we can tweak scheduling parameters, we cannot modify algorithmic logic, or add a completely new algorithm, after the switch has been designed. This paper presents a design for a {\em programmable} packet scheduler, which allows scheduling algorithms---potentially algorithms that are unknown today---to be programmed into a switch without requiring hardware redesign.

Packet scheduling in switches is not programmable; operators only choose among a handful of scheduling algorithms implemented by the manufacturer. In contrast, other switch functions such as packet parsing and header processing are becoming programmable [10, 3, 6]. This paper presents a programmable packet scheduler that allows operators to program a variety of scheduling algorithms.

We present Swarm, a novel architecture that exploits ordered irregular parallelism, which is abundant but hard to mine with current software and hardware techniques. In this architecture, programs consist of short tasks with programmer-specified timestamps. Swarm executes tasks speculatively and out of order, and efficiently speculates thousands of tasks ahead of the earliest active task to uncover ordered parallelism. Swarm builds on prior TLS and HTM schemes, and contributes several new techniques that allow it to scale to large core counts and speculation windows, including a new execution model, speculation-aware hardware task management, selective aborts, and scalable ordered commits.

This article consists of a collection of slides from the author's conference presentation on the special features, system design and architectures, processing capabilities, and targeted markets for Freescale Inc.'s SCORPIO, a 36-core shared memory processor.

In the many-core era, scalable coherence and on-chip interconnects are crucial for shared memory processors. While snoopy coherence is common in small multicore systems, directory-based coherence is the de facto choice for scalability to many cores, as snoopy relies on ordered interconnects which do not scale. However, directory-based coherence does not scale beyond tens of cores due to excessive directory area overhead or inaccurate sharer tracking. Prior techniques supporting ordering on arbitrary unordered networks are impractical for full multicore chip designs.

As technology scales, SoCs are increasing in core counts, leading to the need for scalable NoCs to interconnect the multiple cores on the chip. Given aggressive SoC design targets, NoCs have to deliver low latency, high bandwidth, at low power and area overheads. In this paper, we propose Single-cycle Multi-hop Asynchronous Repeated Traversal (SMART) NoC, a NoC that reconfigures and tailors a generic mesh topology for SoC applications at runtime. The heart of our SMART NoC is a novel low-swing clockless repeated link circuit embedded within the router crossbars, that allows packets to potentially bypass all the way from source to destination core within a single clock cycle, without being latched at any intermediate router. Our clockless repeater link has been proven in silicon in 45nm SOI. Results show that at 2GHz, we can traverse 8mm within a single cycle, i.e. 8 hops with 1mm cores. We implement the SMART NoC to layout and show that SMART NoC gives 60% latency savings, and 2.2X power savings compared to a baseline mesh NoC.

The data plane is in a continuous state of flux. Every few months, researchers publish the design of a new high-performance queueing or scheduling scheme that runs inside the network fabric. Many such schemes have been queen for a day, only to be surpassed soon after as methods --- or evaluation metrics --- evolve.