Offline batch inference, which leverages the flexibility of request batching to achieve higher throughput and lower costs, is becoming more popular for latency-insensitive applications. Meanwhile, recent progress in model capability and modality makes requests more diverse in compute and memory demands, creating unique opportunities for throughput improvement by resource overlapping. However, a request schedule that maximizes resource overlapping can conflict with the schedule that maximizes prefix sharing, a widely-used performance optimization, causing sub-optimal inference throughput. We present BlendServe, a system that maximizes resource utilization of offline batch inference by combining the benefits of resource overlapping and prefix sharing using a resource-aware prefix tree. BlendServe exploits the relaxed latency requirements in offline batch inference to reorder and overlap requests with varied resource demands while ensuring high prefix sharing. We evaluate BlendServe on a variety of synthetic multi-modal workloads and show that it provides up to $1.44\times$ throughput boost compared to widely-used industry standards, vLLM and SGLang.

Disentangling attributes of various sensory signals is central to human-like perception and reasoning and a critical task for higher-order cognitive and neuro-symbolic AI systems. An elegant approach to represent this intricate factorization is via high-dimensional holographic vectors drawing on brain-inspired vector symbolic architectures. However, holographic factorization involves iterative computation with high-dimensional matrix-vector multiplications and suffers from non-convergence problems. In this paper, we present H3DFact, a heterogeneous 3D integrated in-memory compute engine capable of efficiently factorizing high-dimensional holographic representations. H3DFact exploits the computation-in-superposition capability of holographic vectors and the intrinsic stochasticity associated with memristive-based 3D compute-in-memory. Evaluated on large-scale factorization and perceptual problems, H3DFact demonstrates superior capability in factorization accuracy and operational capacity by up to five orders of magnitude, with 5.5x compute density, 1.2x energy efficiency improvements, and 5.9x less silicon footprint compared to iso-capacity 2D designs.

Electronic-photonic computing systems offer immense potential in energy-efficient artificial intelligence (AI) acceleration tasks due to the superior computing speed and efficiency of optics, especially for real-time, low-energy deep neural network (DNN) inference tasks on resource-restricted edge platforms. However, current optical neural accelerators based on foundry-available devices and conventional system architecture still encounter a performance gap compared to highly customized electronic counterparts. To bridge the performance gap due to lack of domain specialization, we present a time-multiplexed dynamic photonic tensor accelerator, dubbed TeMPO, with cross-layer device/circuit/architecture customization. At the device level, we present foundry-compatible, customized photonic devices, including a slow-light electro-optic modulator with experimental demonstration, optical splitters, and phase shifters that significantly reduce the footprint and power in input encoding and dot-product calculation. At the circuit level, partial products are hierarchically accumulated via parallel photocurrent aggregation, lightweight capacitive temporal integration, and sequential digital summation, considerably relieving the analog-to-digital conversion bottleneck. We also employ a multi-tile, multi-core architecture to maximize hardware sharing for higher efficiency. Across diverse edge AI workloads, TeMPO delivers digital-comparable task accuracy with superior quantization/noise tolerance. We achieve a 368.6 TOPS peak performance, 22.3 TOPS/W energy efficiency, and 1.2 TOPS/mm$^2$ compute density, pushing the Pareto frontier in edge AI hardware. This work signifies the power of cross-layer co-design and domain-specific customization, paving the way for future electronic-photonic accelerators with even greater performance and efficiency.

Photonic computing shows promise for transformative advancements in machine learning (ML) acceleration, offering ultra-fast speed, massive parallelism, and high energy efficiency. However, current photonic tensor core (PTC) designs based on standard optical components hinder scalability and compute density due to their large spatial footprint. To address this, we propose an ultra-compact PTC using customized programmable multi-operand multimode interference (MOMMI) devices, named M3ICRO. The programmable MOMMI leverages the intrinsic light propagation principle, providing a single-device programmable matrix unit beyond the conventional computing paradigm of one multiply-accumulate (MAC) operation per device. To overcome the optimization difficulty of customized devices that often requires time-consuming simulation, we apply ML for optics to predict the device behavior and enable a differentiable optimization flow. We thoroughly investigate the reconfigurability and matrix expressivity of our customized PTC, and introduce a novel block unfolding method to fully exploit the computing capabilities of a complex-valued PTC for near-universal real-valued linear transformations. Extensive evaluations demonstrate that M3ICRO achieves a 3.4-9.6x smaller footprint, 1.6-4.4x higher speed, 10.6-42x higher compute density, 3.7-12x higher system throughput, and superior noise robustness compared to state-of-the-art coherent PTC designs, while maintaining close-to-digital task accuracy across various ML benchmarks. Our code is open-sourced at https://github.com/JeremieMelo/M3ICRO-MOMMI.

Neural networks have historically been built layerwise from the set of functions in ${f: \mathbb{R}^n \to \mathbb{R}^m }$, i.e. with activations and weights/parameters represented by real numbers, $\mathbb{R}$. Our work considers a richer set of objects for activations and weights, and undertakes a comprehensive study of alternative algebras as number representations by studying their performance on two challenging problems: large-scale image classification using the ImageNet dataset and language modeling using the enwiki8 and WikiText-103 datasets. We denote this broader class of models as AlgebraNets. Our findings indicate that the conclusions of prior work, which explored neural networks constructed from $\mathbb{C}$ (complex numbers) and $\mathbb{H}$ (quaternions) on smaller datasets, do not always transfer to these challenging settings. However, our results demonstrate that there are alternative algebras which deliver better parameter and computational efficiency compared with $\mathbb{R}$. We consider $\mathbb{C}$, $\mathbb{H}$, $M_{2}(\mathbb{R})$ (the set of $2\times2$ real-valued matrices), $M_{2}(\mathbb{C})$, $M_{3}(\mathbb{R})$ and $M_{4}(\mathbb{R})$. Additionally, we note that multiplication in these algebras has higher compute density than real multiplication, a useful property in situations with inherently limited parameter reuse such as auto-regressive inference and sparse neural networks. We therefore investigate how to induce sparsity within AlgebraNets. We hope that our strong results on large-scale, practical benchmarks will spur further exploration of these unconventional architectures which challenge the default choice of using real numbers for neural network weights and activations.

Last fall, a bit of nerdy controversy arose around AI chip startup NovuMind when the company announced its first low-power chip for processing neural networks. The company claimed that its patented design could natively process 3-D tensor data far more efficiently than other designs that require pre-processing the 3D tensors into 2D matrices. NovuMind portrayed its advantages in low-resolution and high-resolution environments, while typical (Resnet 50) benchmarks target medium resolution algorithms common in datacenters (where NovuMind has a smaller advantage). Now, recent customer wins and trial deployments may provide the trump card in the debate. Let's take a look at the company, its first product, and some case studies that seem to bear out NovuMind's dramatic performance/watt claims.

In some cases, it could be just 0 or 1, with 1 meaning that there is a training set data point close to the location in question in the grid, 0 meaning that you are far enough away from any neighbor. To compute density estimates on each cell of the grid, draw a 3x3 window around each yellow cell, and add 1 to all locations (cells) in that 3x3 window. In our example, the number of groups (g 3: low risk, medium risk, high risk of default) will be multiplied by 2 (M / F) x 3 (young / medium / old), resulting in 18 groups, for instance "young females with medium risk of default" being one of these 18 groups. Of course, accessing a cell in the grid (represented by a 2-dim array), while extremely fast and not depending on the number of observations, still requires a tiny bit of time, but it is entirely dependent only on the size of the array, and its dimension.

Supercomputers used to be a market unto themselves. In their time, giants such as IBM, Control Data Corporation (CDC), Evans & Sutherland (E&S), Silicon Graphics and Cray Computing ruled the supercomputing market. But with Moore's Law restricted to only scaling transistor count, scale-out has taken over a conference series that had epitomized scale-up. Because the nature of supercomputing has changed so much in the past two decades, the market has expanded into a much broader high-performance computing (HPC) market. At this year's SC16 conference, SGI made news by being purchased by Hewlett Packard Enterprise (HPE) and the Cray booth looked a bit forlorn.