Computer Science · Artificial Intelligence Hardware

Consider the following points regarding Graphics Processing Units (GPUs), neural networks, and specialised AI accelerators: I. GPUs are architected to execute thousands of arithmetic operations simultaneously, which aligns with matrix and tensor-heavy workloads in deep learning. II. Neural network training fundamentally prioritises sequential control-flow efficiency over parallel data-path execution, making general-purpose CPUs inherently more suitable than GPUs for large-scale model training. III. High-end data-centre GPUs typically exhibit high power density and therefore require substantial cooling and power provisioning. IV. Tensor Processing Units (TPUs) are domain-specific accelerators designed primarily to optimise tensor algebra operations used in machine learning models. How many of the above points are correct?