# Specialized AI Hardware: TPUs, NPUs, and Neuromorphic Chips

Beyond GPUs, dedicated hardware accelerates AI and reduces power. This sub-article expands [Section 1.3 of Hardware & Software](../hardware-software.md).

## TPUs (Tensor Processing Units)
- Google's custom ASICs for tensor operations.
- Optimized for large Transformer training and inference.
- Co-designed with TensorFlow; deployed in Google Cloud clusters.

## NPUs (Neural Processing Units)
- Consumer chips (Apple Neural Engine, Huawei NPU, Qualcomm AI) accelerate on-device inference.
- Trade flexibility for efficiency; enable edge AI and privacy-preserving processing.

## FPGAs and ASICs
- FPGAs are reprogrammable, useful for specialized low-latency workloads.
- ASICs are fixed-function and maximally efficient for one workload.

## Neuromorphic Computing
- IBM TrueNorth and Intel Loihi emulate spiking neural networks.
- Event-driven, ultra-low-power sensing and control.
- Still niche; programming and accuracy remain challenges.

## Why It Matters
Specialization cuts cost and energy for both training and inference, enabling scale that general-purpose hardware cannot sustain.

## Sources
- Google. *Cloud TPU documentation.*
- Intel. *Loihi documentation.*
- IBM. *TrueNorth overview.*

*Back to [Hardware & Software](../hardware-software.md).*