arXiv subjects

Michael Davies

Publications and source records attributed to Michael Davies.

10 recordsLinked to original sources

arXiv2025

Privacy-Preserving Performance Profiling of In-The-Wild GPUs

GPUs are the dominant platform for many important applications today including deep learning, accelerated computing, and scientific simulation. However, as the complexity of both applications and hardware increases, GPU chip manufacturers face a significant challenge: how to gather comprehensive performance characteristics and value profiles from GPUs deployed in real-world scenarios. Such data, encompassing the types of kernels executed and the time spent in each, is crucial for optimizing chip design and enhancing application performance. Unfortunately, despite the availability of low-level tools like NSYS and NCU, current methodologies fall short, offering data collection capabilities only on an individual user basis rather than a broader, more informative fleet-wide scale. This paper takes on the problem of realizing a system that allows planet-scale real-time GPU performance profiling of low-level hardware characteristics. The three fundamental problems we solve are: i) user experience of achieving this with no slowdown; ii) preserving user privacy, so that no 3rd party is aware of what applications any user runs; iii) efficacy in showing we are able to collect data and assign it applications even when run on 1000s of GPUs. Our results simulate a 100,000 size GPU deployment, running applications from the Torchbench suite, showing our system addresses all 3 problems.

cs.AR↗

arXiv2025

Pedagogically Motivated and Composable Open-Source RISC-V Processors for Computer Science Education

While most instruction set architectures (ISAs) are only available to use through the purchase of a restrictive commercial license, the RISC-V ISA presents a free and open-source alternative. Due to this availability, many free and open-source implementations have been developed and can be accessed on platforms such as GitHub. If an open source, easy-to-use, and robust RISC-V implementation could be obtained, it could be easily adapted for pedagogical and amateur use. In this work we accomplish three goals in relation to this outlook. First, we propose a set of criteria for evaluating the components of a RISC-V implementation's ecosystem from a pedagogical perspective. Second, we analyze a number of existing open-source RISC-V implementations to determine how many of the criteria they fulfill. We then develop a comprehensive solution that meets all of these criterion and is released open-source for other instructors to use. The framework is developed in a composable way that it's different components can be disaggregated per individual course needs. Finally, we also report on a limited study of student feedback.

cs.AR↗

arXiv2025

LIMINAL: Exploring The Frontiers of LLM Decode Performance

The rapid advancement of Large Language Models (LLMs) necessitates a deep understanding of their fundamental performance limits. This paper investigates the limits of LLM inference, focusing on hardware-imposed bottlenecks in auto-regressive decoding. We develop LIMINAL, an analytical performance model that abstracts application requirements and hardware capabilities to systematically explore performance and efficiency across a wide range of current, near-future, and hypothetical hardware. We find LIMINAL is accurate when comparing to LLMs executing on existing hardware, achieving a mean absolute error of $7.6\%$. Our analysis spans from current HBM3 memory technology used in AI accelerators like GPUs and TPUs to systems based on advanced HBM4 and advanced 3D-stacked DRAM technology. We identify five non-negotiable challenges for LLM inference hardware, establishing compute, memory capacity, bandwidth and collective communication as primary barriers to performance. These findings suggest that achieving significant performance gains beyond 10,000 tokens-per-second will require not just hardware evolution but also fundamental algorithmic advances.

cs.AR↗

arXiv2025

Kitsune: Enabling Dataflow Execution on GPUs

State of art DL models are growing in size and complexity, with many modern models also increasing in heterogeneity of behavior. GPUs are still the dominant platform for DL applications, relying on a bulk-synchronous execution model which has many drawbacks and is ill-suited for the graph structure of DL applications. Many industry and academic works attempt to overcome these by employing vertical fusion but this approach still fails to realize three untapped opportunities: (1) the fact that many resources on the GPU are idle while only one operator executes due to temporal multiplexing of the SM; (2) lower energy from more intelligent on-chip data-movement which lends to higher performance in a power-provisioned environment. (3) inability to exploit hidden or reduction dimensions as a source of parallelism to ease pressure on batch size. This paper explores relatively uncharted territory, answering the following key question: Can modest adjustments to the current GPU architecture enable efficient dataflow execution, thereby circumventing the constraints of vertical fusion without necessitating a clean-slate architecture design. We develop Kitsune -- a set of primitives that enable dataflow execution on GPUs and an end-to-end compiler based on PyTorch Dynamo. Across 5 challenge applications, Kitsune can provide 1.3$\times$-2.3$\times$ and 1.1$\times$-2.4$\times$ performance improvement as well as 41%-98% and 16%-42% off-chip traffic reduction for inference and training, respectively.

cs.AR↗

arXiv2024

LookupFFN: Making Transformers Compute-lite for CPU inference

While GPU clusters are the de facto choice for training large deep neural network (DNN) models today, several reasons including ease of workflow, security and cost have led to efforts investigating whether CPUs may be viable for inference in routine use in many sectors of the industry. But the imbalance between the compute capabilities of GPUs and CPUs is huge. Motivated by these considerations, we study a module which is a workhorse within modern DNN architectures, GEMM based Feed Forward Networks (FFNs), and assess the extent to which it can be made compute- (or FLOP-) lite. Specifically, we propose an alternative formulation (we call it LookupFFN) to GEMM based FFNs inspired by the recent studies of using Locality Sensitive Hashing (LSH) to approximate FFNs. Our formulation recasts most essential operations as a memory look-up, leveraging the trade-off between the two resources on any platform: compute and memory (since CPUs offer it in abundance). For RoBERTa language model pretraining, our formulation achieves similar performance compared to GEMM based FFNs, while dramatically reducing the required FLOP. Our development is complemented with a detailed hardware profiling of strategies that will maximize efficiency -- not just on contemporary hardware but on products that will be offered in the near/medium term future. Code is avaiable at \url{https://github.com/mlpen/LookupFFN}.

cs.LG↗

arXiv2023

Violet: Architecturally Exposed Orchestration, Movement, and Placement for Generalized Deep Learning

Deep learning and hardware for it has garnered immense academic and industry interest in the past 5 years, with many novel proposals. However, the state-of-art remains NVIDIA's TensorCore-based systems that provide top-of-line performance and coverage across a wide-spectrum of deep learning applications. In this paper, we first identify four key problems any new DL solution must solve: 1) Data orchestration, 2) Data movement, 3) Work placement and blending these to achieve 4) Coverage across different types of DL applications. With this as a guide, we propose Violet, a novel architecture with roots in multicore SIMD which balances the responsibilities for these four problems between the architecture, microarchitecture and software stack. Compared to the NVIDIA A100 GPU, we find Violet achieves geo-mean 2.4X/10.6X and 2.1X/9.5X performance/efficiency for inference and training across the MLPerf benchmark suite. We present detailed operator-level analysis of the MLPerf benchmark suite, extracting out key behaviors - with implications for architecture research beyond this paper, that underpin the speedup and efficiency. Overall, this paper motivates the importance of balance, that the break down of responsibilities must be thought through carefully in order to compete with incumbent architecture designs.

cs.AR↗

arXiv2021

Divergence Estimation in Message Passing algorithms

Many modern imaging applications can be modeled as compressed sensing linear inverse problems. When the measurement operator involved in the inverse problem is sufficiently random, denoising Scalable Message Passing (SMP) algorithms have a potential to demonstrate high efficiency in recovering compressed data. One of the key components enabling SMP to achieve fast convergence, stability and predictable dynamics is the Onsager correction that must be updated at each iteration of the algorithm. This correction involves the denoiser's divergence that is traditionally estimated via the Black-Box Monte Carlo (BB-MC) method \cite{MC-divergence}. While the BB-MC method demonstrates satisfying accuracy of estimation, it requires executing the denoiser additional times at each iteration and might lead to a substantial increase in computational cost of the SMP algorithms. In this work we develop two Large System Limit models of the Onsager correction for denoisers operating within SMP algorithms and use these models to propose two practical classes of divergence estimators that require no additional executions of the denoiser and demonstrate similar or superior correction compared to the BB-MC method.

cs.IT↗

arXiv2020

Compressed Sensing with Upscaled Vector Approximate Message Passing

The Recently proposed Vector Approximate Message Passing (VAMP) algorithm demonstrates a great reconstruction potential at solving compressed sensing related linear inverse problems. VAMP provides high per-iteration improvement, can utilize powerful denoisers like BM3D, has rigorously defined dynamics and is able to recover signals measured by highly undersampled and ill-conditioned linear operators. Yet, its applicability is limited to relatively small problem sizes due to the necessity to compute the expensive LMMSE estimator at each iteration. In this work we consider the problem of upscaling VAMP by utilizing Conjugate Gradient (CG) to approximate the intractable LMMSE estimator. We propose a rigorous method for correcting and tuning CG withing CG-VAMP to achieve a stable and efficient reconstruction. To further improve the performance of CG-VAMP, we design a warm-starting scheme for CG and develop theoretical models for the Onsager correction and the State Evolution of Warm-Started CG-VAMP (WS-CG-VAMP). Additionally, we develop robust and accurate methods for implementing the WS-CG-VAMP algorithm. The numerical experiments on large-scale image reconstruction problems demonstrate that WS-CG-VAMP requires much fewer CG iterations compared to CG-VAMP to achieve the same or superior level of reconstruction.

cs.IT↗

arXiv2019

Identifying and characterising the impact of excitability in a mathematical model of tumour-immune interactions

We study a five-compartment mathematical model originally proposed by Kuznetsov et al. (1994) to investigate the effect of nonlinear interactions between tumour and immune cells in the tumour microenvironment, whereby immune cells may induce tumour cell death, and tumour cells may inactivate immune cells. Exploiting a separation of timescales in the model, we use the method of matched asymptotics to derive a new two-dimensional, long-timescale, approximation of the full model, which differs from the quasi-steady-state approximation introduced by Kuznetsov et al. (1994), but is validated against numerical solutions of the full model. Through a phase-plane analysis, we show that our reduced model is excitable, a feature not traditionally associated with tumour-immune dynamics. Through a systematic parameter sensitivity analysis, we demonstrate that excitability generates complex bifurcating dynamics in the model. These are consistent with a variety of clinically observed phenomena, and suggest that excitability may underpin tumour-immune interactions. The model exhibits the three stages of immunoediting - elimination, equilibrium, and escape, via stable steady states with different tumour cell concentrations. Such heterogeneity in tumour cell numbers can stem from variability in initial conditions and/or model parameters that control the properties of the immune system and its response to the tumour. We identify different biophysical parameter targets that could be manipulated with immunotherapy in order to control tumour size, and we find that preferred strategies may differ between patients depending on the strength of their immune systems, as determined by patient-specific values of associated model parameters.

q-bio.TO↗

arXiv2016

Deep Action Sequence Learning for Causal Shape Transformation

Deep learning became the method of choice in recent year for solving a wide variety of predictive analytics tasks. For sequence prediction, recurrent neural networks (RNN) are often the go-to architecture for exploiting sequential information where the output is dependent on previous computation. However, the dependencies of the computation lie in the latent domain which may not be suitable for certain applications involving the prediction of a step-wise transformation sequence that is dependent on the previous computation only in the visible domain. We propose that a hybrid architecture of convolution neural networks (CNN) and stacked autoencoders (SAE) is sufficient to learn a sequence of actions that nonlinearly transforms an input shape or distribution into a target shape or distribution with the same support. While such a framework can be useful in a variety of problems such as robotic path planning, sequential decision-making in games, and identifying material processing pathways to achieve desired microstructures, the application of the framework is exemplified by the control of fluid deformations in a microfluidic channel by deliberately placing a sequence of pillars. Learning of a multistep topological transform has significant implications for rapid advances in material science and biomedical applications.

cs.LG↗