Utvärdering och optimering av AI-baserad analys för basbandssystem : Effekterna av beskärning, klustring och kvantisering på neuronnätsförutsägelse av siktlinjeblockering

Millimeter-wave and sub-terahertz frequencies are emerging as key technolo- gies for future baseband networks in the telecommunications industry, as they enable higher data rates and lower latency, critical factors for both users and service providers. However, these media rely heavily on maintaining a clear line-of-sight between the emitter and the receiver to provide the desired service level. Machine learning and neural network-based models have demonstrated promising results in predicting future line-of-sight conditions between a base station and a client, particularly by leveraging radar to sense the surrounding environment. Research has been conducted on developing machine learning models for predicting future line-of-sight blockage using radar. However, no real-world implementation has yet been industrialized and deployed, as several challenges remain. Baseband equipment is subject to stringent hardware constraints to minimize energy consumption and production costs, as well as strict timing requirements that significantly influence the feasibility of executing such models on the system. This thesis specifically examines the hardware requirements associated with running a model for line-of-sight blockage prediction, how these requirements scale with the neural network optimizations: network pruning, weight clustering, and network quantization, and if it is feasible to run this model on real-world deployed hardware. To accomplish this, we tested the model with all different combinations of software optimization, on two different inference platforms based on TFlite and ONNX, on one ARM and one X86-based simulated hardware platform using the gem5 hardware simulator. Each hardware platform was designed to mirror hardware that is suitable for a real-world baseband system. This thesis examines the hardware requirements for running a model developed by Demirhan and Alkhateeb for line-of-sight blockage prediction, how these requirements scale with the three neural network optimization techniques network pruning, weight clustering, and quantization, and the feasibility of deploying this model on real-world hardware. To evaluate this, the model was tested with various combinations of software optimizations on two different inference platforms, based on TFLite and ONNX, respectively. These tests were conducted on simulated ARM- and x86-based hardware platforms using the gem5 hardware simulator. Each simulated hardware configuration was designed to reflect architectures suitable for real-world baseband systems. The results in this thesis concluded that neither unstructured pruning nor clustering affected the hardware requirements of running the model. Quantization greatly reduced the memory traffic volume and bandwidth by 44% and 62% respectively, but increased inference time by 48% on TFlite and 50% on ONNX. Structured pruning improved all hardware characteristics across the board, with a 34% to 80% decrease in the number of committed CPU instructions, a 51% to 70% decrease in memory volume, and a 25% decrease to a 45% increase in bandwidth depending on the platform, as well as providing the lowest inference time. Integer instructions, SIMD integer instructions, and SIMD floating-point instructions were the most utilized CPU instructions regardless of the platform, with quantization heavily skewing the results from SIMD floating-point to SIMD integer instructions. Lastly, we concluded that running this model for line-of-sight prediction on hardware suitable for baseband systems would be feasible, as the expected inference time on real hardware would be lower than 30ms, allowing for fast reactions to environmental changes.