Juan Gómez-Luna

Maurus Item, Juan Gómez-Luna, Yuxin Guo et al.

Processing-in-memory (PIM) promises to alleviate the data movement bottleneck in modern computing systems. However, current real-world PIM systems have the inherent disadvantage that their hardware is more constrained than in conventional processors (CPU, GPU), due to the difficulty and cost of building processing elements near or inside the memory. As a result, general-purpose PIM architectures support fairly limited instruction sets and struggle to execute complex operations such as transcendental functions and other hard-to-calculate operations (e.g., square root). These operations are particularly important for some modern workloads, e.g., activation functions in machine learning applications. In order to provide support for transcendental (and other hard-to-calculate) functions in general-purpose PIM systems, we present \emph{TransPimLib}, a library that provides CORDIC-based and LUT-based methods for trigonometric functions, hyperbolic functions, exponentiation, logarithm, square root, etc. We develop an implementation of TransPimLib for the UPMEM PIM architecture and perform a thorough evaluation of TransPimLib's methods in terms of performance and accuracy, using microbenchmarks and three full workloads (Blackscholes, Sigmoid, Softmax). We open-source all our code and datasets at~\url{https://github.com/CMU-SAFARI/transpimlib}.

Gagandeep Singh, Rakesh Nadig, Jisung Park et al.

Hybrid storage systems (HSS) use multiple different storage devices to provide high and scalable storage capacity at high performance. Recent research proposes various techniques that aim to accurately identify performance-critical data to place it in a "best-fit" storage device. Unfortunately, most of these techniques are rigid, which (1) limits their adaptivity to perform well for a wide range of workloads and storage device configurations, and (2) makes it difficult for designers to extend these techniques to different storage system configurations (e.g., with a different number or different types of storage devices) than the configuration they are designed for. We introduce Sibyl, the first technique that uses reinforcement learning for data placement in hybrid storage systems. Sibyl observes different features of the running workload as well as the storage devices to make system-aware data placement decisions. For every decision it makes, Sibyl receives a reward from the system that it uses to evaluate the long-term performance impact of its decision and continuously optimizes its data placement policy online. We implement Sibyl on real systems with various HSS configurations. Our results show that Sibyl provides 21.6%/19.9% performance improvement in a performance-oriented/cost-oriented HSS configuration compared to the best previous data placement technique. Our evaluation using an HSS configuration with three different storage devices shows that Sibyl outperforms the state-of-the-art data placement policy by 23.9%-48.2%, while significantly reducing the system architect's burden in designing a data placement mechanism that can simultaneously incorporate three storage devices. We show that Sibyl achieves 80% of the performance of an oracle policy that has complete knowledge of future access patterns while incurring a very modest storage overhead of only 124.4 KiB.

Geraldo F. Oliveira, Juan Gómez-Luna, Saugata Ghose et al.

Neural networks (NNs) are growing in importance and complexity. A neural network's performance (and energy efficiency) can be bound either by computation or memory resources. The processing-in-memory (PIM) paradigm, where computation is placed near or within memory arrays, is a viable solution to accelerate memory-bound NNs. However, PIM architectures vary in form, where different PIM approaches lead to different trade-offs. Our goal is to analyze, discuss, and contrast DRAM-based PIM architectures for NN performance and energy efficiency. To do so, we analyze three state-of-the-art PIM architectures: (1) UPMEM, which integrates processors and DRAM arrays into a single 2D chip; (2) Mensa, a 3D-stack-based PIM architecture tailored for edge devices; and (3) SIMDRAM, which uses the analog principles of DRAM to execute bit-serial operations. Our analysis reveals that PIM greatly benefits memory-bound NNs: (1) UPMEM provides 23x the performance of a high-end GPU when the GPU requires memory oversubscription for a general matrix-vector multiplication kernel; (2) Mensa improves energy efficiency and throughput by 3.0x and 3.1x over the Google Edge TPU for 24 Google edge NN models; and (3) SIMDRAM outperforms a CPU/GPU by 16.7x/1.4x for three binary NNs. We conclude that the ideal PIM architecture for NN models depends on a model's distinct attributes, due to the inherent architectural design choices.

Juan Gómez-Luna, Yuxin Guo, Sylvan Brocard et al.

Training machine learning (ML) algorithms is a computationally intensive process, which is frequently memory-bound due to repeatedly accessing large training datasets. As a result, processor-centric systems (e.g., CPU, GPU) suffer from costly data movement between memory units and processing units, which consumes large amounts of energy and execution cycles. Memory-centric computing systems, i.e., with processing-in-memory (PIM) capabilities, can alleviate this data movement bottleneck. Our goal is to understand the potential of modern general-purpose PIM architectures to accelerate ML training. To do so, we (1) implement several representative classic ML algorithms (namely, linear regression, logistic regression, decision tree, K-Means clustering) on a real-world general-purpose PIM architecture, (2) rigorously evaluate and characterize them in terms of accuracy, performance and scaling, and (3) compare to their counterpart implementations on CPU and GPU. Our evaluation on a real memory-centric computing system with more than 2500 PIM cores shows that general-purpose PIM architectures can greatly accelerate memory-bound ML workloads, when the necessary operations and datatypes are natively supported by PIM hardware. For example, our PIM implementation of decision tree is $27\times$ faster than a state-of-the-art CPU version on an 8-core Intel Xeon, and $1.34\times$ faster than a state-of-the-art GPU version on an NVIDIA A100. Our K-Means clustering on PIM is $2.8\times$ and $3.2\times$ than state-of-the-art CPU and GPU versions, respectively. To our knowledge, our work is the first one to evaluate ML training on a real-world PIM architecture. We conclude with key observations, takeaways, and recommendations that can inspire users of ML workloads, programmers of PIM architectures, and hardware designers & architects of future memory-centric computing systems.

Geraldo F. Oliveira, Amirali Boroumand, Saugata Ghose et al.

Today's computing systems require moving data back-and-forth between computing resources (e.g., CPUs, GPUs, accelerators) and off-chip main memory so that computation can take place on the data. Unfortunately, this data movement is a major bottleneck for system performance and energy consumption. One promising execution paradigm that alleviates the data movement bottleneck in modern and emerging applications is processing-in-memory (PIM), where the cost of data movement to/from main memory is reduced by placing computation capabilities close to memory. Naively employing PIM to accelerate data-intensive workloads can lead to sub-optimal performance due to the many design constraints PIM substrates impose. Therefore, many recent works co-design specialized PIM accelerators and algorithms to improve performance and reduce the energy consumption of (i) applications from various application domains; and (ii) various computing environments, including cloud systems, mobile systems, and edge devices. We showcase the benefits of co-designing algorithms and hardware in a way that efficiently takes advantage of the PIM paradigm for two modern data-intensive applications: (1) machine learning inference models for edge devices and (2) hybrid transactional/analytical processing databases for cloud systems. We follow a two-step approach in our system design. In the first step, we extensively analyze the computation and memory access patterns of each application to gain insights into its hardware/software requirements and major sources of performance and energy bottlenecks in processor-centric systems. In the second step, we leverage the insights from the first step to co-design algorithms and hardware accelerators to enable high-performance and energy-efficient data-centric architectures for each application.

Gagandeep Singh, Dionysios Diamantopoulos, Juan Gómez-Luna et al.

Machine learning has recently gained traction as a way to overcome the slow accelerator generation and implementation process on an FPGA. It can be used to build performance and resource usage models that enable fast early-stage design space exploration. First, training requires large amounts of data (features extracted from design synthesis and implementation tools), which is cost-inefficient because of the time-consuming accelerator design and implementation process. Second, a model trained for a specific environment cannot predict performance or resource usage for a new, unknown environment. In a cloud system, renting a platform for data collection to build an ML model can significantly increase the total-cost-ownership (TCO) of a system. Third, ML-based models trained using a limited number of samples are prone to overfitting. To overcome these limitations, we propose LEAPER, a transfer learning-based approach for prediction of performance and resource usage in FPGA-based systems. The key idea of LEAPER is to transfer an ML-based performance and resource usage model trained for a low-end edge environment to a new, high-end cloud environment to provide fast and accurate predictions for accelerator implementation. Experimental results show that LEAPER (1) provides, on average across six workloads and five FPGAs, 85% accuracy when we use our transferred model for prediction in a cloud environment with 5-shot learning and (2) reduces design-space exploration time for accelerator implementation on an FPGA by 10x, from days to only a few hours.

Juan Gómez-Luna, Yuxin Guo, Sylvan Brocard et al.

Training machine learning algorithms is a computationally intensive process, which is frequently memory-bound due to repeatedly accessing large training datasets. As a result, processor-centric systems (e.g., CPU, GPU) suffer from costly data movement between memory units and processing units, which consumes large amounts of energy and execution cycles. Memory-centric computing systems, i.e., computing systems with processing-in-memory (PIM) capabilities, can alleviate this data movement bottleneck. Our goal is to understand the potential of modern general-purpose PIM architectures to accelerate machine learning training. To do so, we (1) implement several representative classic machine learning algorithms (namely, linear regression, logistic regression, decision tree, K-means clustering) on a real-world general-purpose PIM architecture, (2) characterize them in terms of accuracy, performance and scaling, and (3) compare to their counterpart implementations on CPU and GPU. Our experimental evaluation on a memory-centric computing system with more than 2500 PIM cores shows that general-purpose PIM architectures can greatly accelerate memory-bound machine learning workloads, when the necessary operations and datatypes are natively supported by PIM hardware. To our knowledge, our work is the first one to evaluate training of machine learning algorithms on a real-world general-purpose PIM architecture.

Steve Rhyner, Haocong Luo, Juan Gómez-Luna et al.

Modern Machine Learning (ML) training on large-scale datasets is a very time-consuming workload. It relies on the optimization algorithm Stochastic Gradient Descent (SGD) due to its effectiveness, simplicity, and generalization performance. Processor-centric architectures (e.g., CPUs, GPUs) commonly used for modern ML training workloads based on SGD are bottlenecked by data movement between the processor and memory units due to the poor data locality in accessing large datasets. As a result, processor-centric architectures suffer from low performance and high energy consumption while executing ML training workloads. Processing-In-Memory (PIM) is a promising solution to alleviate the data movement bottleneck by placing the computation mechanisms inside or near memory. Our goal is to understand the capabilities of popular distributed SGD algorithms on real-world PIM systems to accelerate data-intensive ML training workloads. To this end, we 1) implement several representative centralized parallel SGD algorithms on the real-world UPMEM PIM system, 2) rigorously evaluate these algorithms for ML training on large-scale datasets in terms of performance, accuracy, and scalability, 3) compare to conventional CPU and GPU baselines, and 4) discuss implications for future PIM hardware and highlight the need for a shift to an algorithm-hardware codesign. Our results demonstrate three major findings: 1) The UPMEM PIM system can be a viable alternative to state-of-the-art CPUs and GPUs for many memory-bound ML training workloads, especially when operations and datatypes are natively supported by PIM hardware, 2) it is important to carefully choose the optimization algorithms that best fit PIM, and 3) the UPMEM PIM system does not scale approximately linearly with the number of nodes for many data-intensive ML training workloads. We open source all our code to facilitate future research.

Harshita Gupta, Mayank Kabra, Jaewoo Park et al.

Homomorphic encryption (HE) enables computation over encrypted data, offering strong privacy guarantees for untrusted computing environments. Practical adoption remains limited by high computational complexity, large ciphertext sizes, and substantial data movement. Processor-centric architectures (CPUs, GPUs, ASICs) hit fundamental bottlenecks on HE workloads because ciphertexts are large, data locality is low, and primitives such as relinearization and bootstrapping repeatedly access large auxiliary metadata. Processing-In-Memory (PIM) is a promising mitigation by computing near or inside memory. Prior PIM proposals for HE either do not target real-world PIM systems or cover only a narrow set of operations. We comprehensively characterize HE operations on a real-world, general-purpose PIM system. We implement a complete set of HE kernels used by emerging applications (databases, machine learning) on the UPMEM PIM system, evaluate performance and scalability, compare against CPU and GPU baselines, and discuss implications for future PIM hardware. Our results demonstrate four major findings. (1) HE-based applications expose distinct bottlenecks across execution stages: some kernels are compute-bound due to modular arithmetic, while others are memory-bound due to large ciphertexts and intermediate data. These bottlenecks are exacerbated by limited per-core compute and per-bank capacity, which force frequent data movement. (2) The dominant compute bottleneck is the lack of native 64-bit modular integer multiplication, a key HE primitive. (3) Limited per-bank memory capacity is the second major bottleneck, since HE ciphertexts and auxiliary metadata do not fit and require inter-bank movement. (4) Despite these limits, PIM can be a viable alternative to state-of-the-art CPU and GPU systems for HE when equipped with native modular multiplication and efficient inter-PIM data movement.

Kailash Gogineni, Sai Santosh Dayapule, Juan Gómez-Luna et al.

Reinforcement Learning (RL) trains agents to learn optimal behavior by maximizing reward signals from experience datasets. However, RL training often faces memory limitations, leading to execution latencies and prolonged training times. To overcome this, SwiftRL explores Processing-In-Memory (PIM) architectures to accelerate RL workloads. We achieve near-linear performance scaling by implementing RL algorithms like Tabular Q-learning and SARSA on UPMEM PIM systems and optimizing for hardware. Our experiments on OpenAI GYM environments using UPMEM hardware demonstrate superior performance compared to CPU and GPU implementations.

Yintao He, Haiyu Mao, Christina Giannoula et al.

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. 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. In this paper, we aim to accelerate LLM decoding while considering the dynamically changing characteristics of the kernels involved. We propose PAPI (PArallel Decoding with PIM), a PIM-enabled heterogeneous architecture that exploits dynamic scheduling of compute-bound or memory-bound kernels to suitable hardware units. PAPI has two key mechanisms: (1) online kernel characterization to dynamically schedule kernels to the most suitable hardware units at runtime and (2) a PIM-enabled heterogeneous computing system that harmoniously orchestrates both computation-centric processing units and hybrid PIM units with different computing capabilities. Our experimental results on three broadly-used LLMs show that PAPI achieves 1.8$\times$ and 11.1$\times$ speedups over a state-of-the-art heterogeneous LLM accelerator and a state-of-the-art PIM-only LLM accelerator, respectively.

Lois Orosa, Yaohua Wang, Mohammad Sadrosadati et al.

DRAM is the dominant main memory technology used in modern computing systems. Computing systems implement a memory controller that interfaces with DRAM via DRAM commands. DRAM executes the given commands using internal components (e.g., access transistors, sense amplifiers) that are orchestrated by DRAM internal timings, which are fixed foreach DRAM command. Unfortunately, the use of fixed internal timings limits the types of operations that DRAM can perform and hinders the implementation of new functionalities and custom mechanisms that improve DRAM reliability, performance and energy. To overcome these limitations, we propose enabling programmable DRAM internal timings for controlling in-DRAM components. To this end, we design CODIC, a new low-cost DRAM substrate that enables fine-grained control over four previously fixed internal DRAM timings that are key to many DRAM operations. We implement CODIC with only minimal changes to the DRAM chip and the DDRx interface. To demonstrate the potential of CODIC, we propose two new CODIC-based security mechanisms that outperform state-of-the-art mechanisms in several ways: (1) a new DRAM Physical Unclonable Function (PUF) that is more robust and has significantly higher throughput than state-of-the-art DRAM PUFs, and (2) the first cold boot attack prevention mechanism that does not introduce any performance or energy overheads at runtime.

Lois Orosa, Yaohua Wang, Ivan Puddu et al.

DRAM manufacturers have been prioritizing memory capacity, yield, and bandwidth for years, while trying to keep the design complexity as simple as possible. DRAM chips do not carry out any computation or other important functions, such as security. Processors implement most of the existing security mechanisms that protect the system against security threats, because 1) executing security mechanisms usually require non-trivial computational capabilities (e.g., encryption), and 2) commodity DRAM chips are not designed to perform computations or tasks other than data storage. In this work, we advocate for DRAM as a key component for providing security mechanisms to the system. To this end, we propose Dataplant, a new class of low-cost, high-performance, and reliable security primitives that can be integrated in commodity DRAM chips with minimal changes. The main idea of Dataplant is to slightly modify the internal DRAM timing signals to expose the inherent process variation found in all DRAM chips for generating unpredictable but reproducible values (e.g., keys) within DRAM. We use Dataplant to build two new security mechanisms. First, a new Dataplant-based physical unclonable function (PUF) with non-destructive read-out, low evaluation latency, robust responses, resiliency to temperature changes, and data-independent responses. Second, a new cold boot attack prevention mechanism that automatically destroys all data within DRAM on every power cycle with zero run-time energy and latency overheads. Using a combination of detailed simulations and experiments with 136 real commodity DRAM chips, we show that our Dataplant-based PUF has 1.8x higher throughput than the best state-of-the-art DRAM PUFs. We also demonstrate that our Dataplant-based cold boot attack protection mechanism is 19.5x faster and consumes 2.54x less energy when compared to existing mechanisms.