Specialized Accelerator Generations: Workload Positioning and Hardware–Software Co-Design

SourcesSurvey §2.3 and §5; TPU, Neuron, MTIA and Maia corpus records.

4.25.1TPU v1 → v2/v3: Bridging Inference and Training Designs#

4.25.2TPU v4/v5e/v5p/v6e/v7: Compute, Memory, and Pod Evolution#

4.25.3TPU 8t/8i: Workload Positioning and Disclosure Boundaries in the Survey#

4.25.4AWS Inferentia/Trainium: NeuronCore and Multi-Engine Evolution#

4.25.5Trn1/Trn2: Intra-Instance Torus and Inter-UltraServer Connection#

4.25.6Trn3: Switched Fabric Evolution in the Survey#

4.25.7MTIA/Maia: Shifts Across Recommendation, Generative AI, and Inference Workloads#