Differentiable rendering is widely used in emerging applications that represent any 3D scene as a model trained using gradient descent from 2D images. Recent works (e.g., 3D Gaussian Splatting) use rasterization to enable rendering photo-realistic imagery at high speeds from these learned 3D models. These rasterization-based differentiable rendering methods have been demonstrated to be very promising, providing state-of-art quality for various important tasks. However, training a model to represent a scene is still time-consuming even on powerful GPUs. In this work, we observe that the gradient computation step during model training is a significant bottleneck due to the large number of atomic operations. These atomics overwhelm the atomic units in the L2 cache of GPUs, causing long stalls.

Large language models (LLMs) are widely used for natural language understanding and text generation. An LLM model relies on a time-consuming step called LLM decoding to generate output tokens. Several prior works focus on improving the performance of LLM decoding using parallelism techniques, such as batching and speculative decoding. State-of-the-art LLM decoding has both compute-bound and memory-bound kernels. Some prior works statically identify and map these different kernels to a heterogeneous architecture consisting of both processing-in-memory (PIM) units and computation-centric accelerators (e.g., GPUs). We observe that characteristics of LLM decoding kernels (e.g., whether or not a kernel is memory-bound) can change dynamically due to parameter changes to meet user and/or system demands, making (1) static kernel mapping to PIM units and computation-centric accelerators suboptimal, and (2) one-size-fits-all approach of designing PIM units inefficient due to a large degree of heterogeneity even in memory-bound kernels.

Various parallelism, such as data, tensor, and pipeline parallelism, along with memory optimizations like activation checkpointing, redundancy elimination, and offloading, have been proposed to accelerate distributed training for Large Language Models. To find the best combination of these techniques, automatic distributed training systems are proposed. However, existing systems only tune a subset of optimizations, due to the lack of overlap awareness, inability to navigate the vast search space, and ignoring the inter-microbatch imbalance, leading to sub-optimal performance. To address these shortcomings, we propose Mist, a memory, overlap, and imbalance-aware automatic distributed training system that comprehensively co-optimizes all memory footprint reduction techniques alongside parallelism. Mist is based on three key ideas: (1) fine-grained overlap-centric scheduling, orchestrating optimizations in an overlapped manner, (2) symbolic-based performance analysis that predicts runtime and memory usage using symbolic expressions for fast tuning, and (3) imbalance-aware hierarchical tuning, decoupling the process into an inter-stage imbalance and overlap aware Mixed Integer Linear Programming problem and an intra-stage Dual-Objective Constrained Optimization problem, and connecting them through Pareto frontier sampling. Our evaluation results show that Mist achieves an average of 1.28× (up to 1.73×) and 1.27× (up to 2.04×) speedup compared to state-of-the-art manual system Megatron-LM and state-of-the-art automatic system Aceso, respectively.

Atalanta is a lossless, hardware/software co-designed compression technique for the tensors of fixed-point quantized deep neural networks. Atalanta increases effective memory capacity, reduces off-die traffic, and/or helps to achieve the desired performance/energy targets while using smaller off-die memories during inference. Atalanta is architected to deliver nearly identical coding efficiency compared to Arithmetic Coding while avoiding its complexity, overhead, and bandwidth limitations. Indicatively, the Atalanta decoder and encoder units each use less than 50B of internal storage. In hardware, Atalanta is implemented as an assist over any machine learning accelerator transparently compressing/decompressing tensors just before the off-die memory controller. This work shows the performance and energy efficiency of Atalanta when implemented in a 65nm technology node. Atalanta reduces data footprint of weights and activations to 60% and 48% respectively on average over a wide set of 8-bit quantized models and complements a wide range of quantization methods. Integrated with a Tensorcore-based accelerator, Atalanta boosts the speedup and energy efficiency to 1.44× and 1.37×, respectively. Atalanta is effective at compressing the stashed activations during training for fixed-point inference.

Spike sorting is the process of parsing electrophysiological signals from neurons to identify if, when, and which particular neurons fire. Spike sorting is a particularly difficult task in computational neuroscience due to the growing scale of recording technologies and complexity in traditional spike sorting algorithms. Previous spike sorters can be divided into software-based and hardware-based solutions. Software solutions are highly accurate but operate on recordings after-the-fact, and often require utilization of high-power GPUs to process in a timely fashion, and they cannot be used in portable applications. Hardware solutions suffer in terms of accuracy due to the simplification of mechanisms for implementation's sake and process only up to 128 inputs. This work answers the question: "How much computation power and memory storage is needed to sort spikes from 1000s of channels to keep up with advances in probe technology?" We analyze the computational and memory requirements for modern software spike sorters to identify their potential bottlenecks - namely in the template memory storage. We architect Marple, a highly optimized hardware pipeline for spike sorting which incorporates a novel mechanism to reduce the template memory storage from 8 - 11x. Marple is scalable, uses a flexible vector-based back-end to perform neuron identification, and a fixed-function front-end to filter the incoming streams into areas of interest. The implementation is projected to use just 79mW in 7nm, when spike sorting 10K channels at peak activity. We further demonstrate, for the first time, a machine learning replacement for the template matching stage.

Sparse Convolution (SC) is widely used for processing 3D point clouds that are inherently sparse. Different from dense convolution, SC preserves the sparsity of the input point cloud by only allowing outputs to specific locations. To efficiently compute SC, prior SC engines first use hash tables to build a kernel map that stores the necessary General Matrix Multiplication (GEMM) operations to be executed (Map step), and then use a Gather-GEMM-Scatter process to execute these GEMM operations (GMaS step). In this work, we analyze the shortcomings of prior state-of-the-art SC engines, and propose Minuet, a novel memory-efficient SC engine tailored for modern GPUs. Minuet proposes to (i) replace the hash tables used in the Map step with a novel segmented sorting double-traversed binary search algorithm that highly utilizes the on-chip memory hierarchy of GPUs, (ii) use a lightweight scheme to autotune the tile size in the Gather and Scatter operations of the GMaS step, such that to adapt the execution to the particular characteristics of each SC layer, dataset, and GPU architecture, and (iii) employ a padding-efficient GEMM grouping approach that reduces both memory padding and kernel launching overheads. Our evaluations show that Minuet significantly outperforms prior SC engines by on average 1.74× (up to 2.22×) for end-to-end point cloud network executions. Our novel segmented sorting double-traversed binary search algorithm achieves superior speedups by 15.8× on average (up to 26.8×) over prior SC engines in the Map step. The source code of Minuet is publicly available at https://github.com/UofT-EcoSystem/Minuet.

Fine-tuning is the gateway to transferring learned knowledge in a pre-trained Large Language Model (LLM) on many downstream applications. To make LLM fine-tuning more affordable, prior works follow two paths: i) adapters freeze the pre-trained LLM weights and inject a small number of trainable weights during fine-tuning, and ii) pruners remove the less important weights in pre-trained LLMs and train the remaining sparse weights during fine-tuning. We find that the former introduces computation overheads due to the injected trainable parameters, while the latter introduces an expensive pre-processing step to identify the important weights and degrades model quality. To get the best of both worlds, we propose Sylva, a novel LLM fine-tuning procedure that provides high system performance during fine-tuning and attains state-of-the-art model quality on downstream applications. Sylva identifies the most important LLM weights via second-order information in a pre-processing step, and significantly reduces the computation and storage costs of the pre-processing step via i) a hierarchical approximation of second-order information, and ii) an online projection and rediagonalization algorithm. Sylva trains only the sparse important weights and embeds these sparse weights into the pre-trained LLM during fine-tuning to provide high system performance. We show that end-to-end fine-tuning with Sylva is, on average, 5.1 × faster than ZeRO and 1.2 × faster than LoRA, the state-of-the-art adapter approach. Sylva’s hierarchical approximation reduces the peak GPU memory in the pre-processing step by 2.3 × compared to K-FAC, the most widely used approximation to second-order information. The source code of Sylva is publicly available at https://github.com/CentML/Sylva.

Abstract Graph coloring is widely used to parallelize scientific applications by identifying subsets of independent tasks that can be executed simultaneously. Graph coloring assigns colors the vertices of a graph, such that no adjacent vertices have the same color. The number of colors used corresponds to the number of parallel steps in a real-world end-application. Therefore, the total runtime of the graph coloring kernel adds to the overall parallel overhead of the real-world end-application, whereas the number of the vertices of each color class determines the number of the independent concurrent tasks of each parallel step, thus affecting the amount of parallelism and hardware resource utilization in the execution of the real-world end-application. In this work, we propose a high-performance graph coloring algorithm, named ColorTM, that leverages Hardware Transactional Memory (HTM) to detect coloring inconsistencies between adjacent vertices. ColorTM detects and resolves coloring inconsistencies between adjacent vertices with an eager approach to minimize data access costs, and implements a speculative synchronization scheme to minimize synchronization costs and increase parallelism. We extend our proposed algorithmic design to propose a balanced graph coloring algorithm, named BalColorTM, with which all color classes include almost the same number of vertices to achieve high parallelism and resource utilization in the execution of the real-world end-applications. We evaluate ColorTM and BalColorTM using a wide variety of large real-world graphs with diverse characteristics. ColorTM and BalColorTM improve performance by 12.98 $$\times$$ × and 1.78 $$\times$$ × on average using 56 parallel threads compared to prior state-of-the-art approaches. Moreover, we study the impact of our proposed graph coloring algorithmic designs on a popular end-application, i.e., Community Detection, and demonstrate the ColorTM and BalColorTM can provide high-performance improvements in real-world end-applications across various input data given.

Near-Data-Processing (NDP) architectures present a promising way to alleviate data movement costs and can provide significant performance and energy benefits to parallel applications. Typically, NDP architectures support several NDP units, each including multiple simple cores placed close to memory. To fully leverage the benefits of NDP and achieve high performance for parallel workloads, efficient synchronization among the NDP cores of a system is necessary. However, supporting synchronization in many NDP systems is challenging because they lack shared caches and hardware cache coherence support, which are commonly used for synchronization in multicore systems, and communication across different NDP units can be expensive. This paper comprehensively examines the synchronization problem in NDP systems, and proposes SynCron, an end-to-end synchronization solution for NDP systems. SynCron adds low-cost hardware support near memory for synchronization acceleration, and avoids the need for hardware cache coherence support. SynCron has three components: 1) a specialized cache memory structure to avoid memory accesses for synchronization and minimize latency overheads, 2) a hierarchical message-passing communication protocol to minimize expensive communication across NDP units of the system, and 3) a hardware-only overflow management scheme to avoid performance degradation when hardware resources for synchronization tracking are exceeded. We evaluate SynCron using a variety of parallel workloads, covering various contention scenarios. SynCron improves performance by 1.27× on average (up to 1.78×) under high-contention scenarios, and by 1.35× on average (up to 2.29×) under low-contention real applications, compared to state-of-the-art approaches. SynCron reduces system energy consumption by 2.08× on average (up to 4.25×).

Time series analysis is a key technique for extracting and predicting events in domains as diverse as epidemiology, genomics, neuroscience, environmental sciences, economics, and more. Matrix profile, the state-of-the-art algorithm to perform time series analysis, computes the most similar subsequence for a given query subsequence within a sliced time series. Matrix profile has low arithmetic intensity, but it typically operates on large amounts of time series data. In current computing systems, this data needs to be moved between the off-chip memory units and the on-chip computation units for performing matrix profile. This causes a major performance bottleneck as data movement is extremely costly in terms of both execution time and energy. In this work, we present NATSA, the first Near-Data Processing accelerator for time series analysis. The key idea is to exploit modern 3D-stacked High Bandwidth Memory (HBM) to enable efficient and fast specialized matrix profile computation near memory, where time series data resides. NATSA provides three key benefits: 1) quickly computing the matrix profile for a wide range of applications by building specialized energy-efficient floating-point arithmetic processing units close to HBM, 2) improving the energy efficiency and execution time by reducing the need for data movement over slow and energy-hungry buses between the computation units and the memory units, and 3) analyzing time series data at scale by exploiting low-latency, high-bandwidth, and energy-efficient memory access provided by HBM. Our experimental evaluation shows that NATSA improves performance by up to 14.2× (9.9× on average) and reduces energy by up to 27.2 × (19.4 × on average), over the state-of-the-art multi-core implementation. NATSA also improves performance by 6.3 × and reduces energy by 10.2 × over a general-purpose NDP platform with 64 in-order cores.

Designing scalable concurrent priority queues for contemporary NUMA servers is challenging. Several NUMA-unaware implementations can scale up to a high number of threads exploiting the potential parallelism of the insert operations. In contrast, in deleteMin-dominated workloads, threads compete for accessing the same memory locations, i.e. the first item in the priority queue. In such cases, NUMA-aware implementations are typically used, since they reduce the coherence traffic between the nodes of a NUMA system.

Important workloads, such as machine learning and graph analytics applications, heavily involve sparse linear algebra operations. These operations use sparse matrix compression as an effective means to avoid storing zeros and performing unnecessary computation on zero elements. However, compression techniques like Compressed Sparse Row (CSR) that are widely used today introduce significant instruction overhead and expensive pointer-chasing operations to discover the positions of the non-zero elements. In this paper, we identify the discovery of the positions (i.e., indexing) of non-zero elements as a key bottleneck in sparse matrix-based workloads, which greatly reduces the benefits of compression.