What circuits are optimized to run deep learning algorithms?

• A. TPUs
• B. GPUs
• C. FPGAs
• D. ASICs

Answer: (A)

Specialized hardware accelerators designed for neural networks deliver the high throughput needed by deep learning workloads. TPUs are Google's custom ASICs built specifically to accelerate the tensor operations at the heart of neural networks, with an architecture geared toward large-scale matrix multiplications, convolutions, and rapid data movement. This design achieves very high FLOPs per watt and fast training and inference for common deep learning models, especially when paired with the TensorFlow ecosystem. While GPUs are flexible and excellent for a wide range of tasks, TPUs are optimized for the particular patterns of computation and data reuse that deep learning requires, making them especially well-suited for running these algorithms. FPGAs offer reconfigurability and can be tuned for DL, but reaching comparable peak performance typically demands more design effort and often lower sustained throughput. ASICs cover a broad range of purpose-built chips, but TPUs represent a family of ASICs purpose-built for deep learning, which is why they are optimized for these workloads.