Survey on emsenn.net

NPU LLM Inference

Sun, 08 Mar 2026 00:00:00 +0000

An assessment of running large language models on neural processing units, with particular attention to the Qualcomm Snapdragon X Elite (45 TOPS), as of early 2026.

The fundamental tension

NPUs were designed for convolutional neural network (CNN) inference: image classification, object detection, segmentation. These workloads have regular, predictable compute patterns that map cleanly onto fixed-point arithmetic and static data flow. Transformer-based language models are architecturally different: the attention mechanism involves softmax and exponentiation that create extreme dynamic ranges, and autoregressive token generation is inherently sequential. This mismatch is the root of most current limitations.

On-Device Inference Frameworks

Sun, 08 Mar 2026 00:00:00 +0000

A survey of software frameworks for running machine learning models locally on consumer hardware, focusing on Windows and NPU support as of early 2026.

ONNX Runtime

Microsoft’s ONNX Runtime is the primary framework for NPU inference on Windows. Its QNN execution provider integrates Qualcomm’s AI Engine Direct SDK, routing ONNX model operations to the Hexagon NPU. Since version 1.18.0, prebuilt packages include the QNN dependencies. Microsoft distributes updates to the QNN execution provider through Windows Update. This is the path jointly recommended by Microsoft and Qualcomm.

Structured-work paradigms: a cross-disciplinary survey

Sat, 07 Mar 2026 00:00:00 +0000

This text surveys structured-work methodologies across three disciplines — business/management, government/military, and scientific/academic research — to identify patterns that inform the design of agent workflows in the agential semioverse repository.

The survey was conducted via web research in March 2026. Each domain was researched independently to avoid cross-contamination of framing. Citations are provided per finding.

Cross-cutting patterns

Five patterns recur across all three domains. These are the findings most relevant to the emsemioverse’s workflow design.