Transistors, 3 dimensions, Semiconductors Moore’s Law Silicon Fabs Big companies that do this Quantum chips Chips Wafers Lithography tool (ASML) Very expensive
ASIC Application
Parallel programming Tensors
Training and Inference
ASIC is like a single-purpose tool Much more efficient in running those workloads. Trade off in flexibility Startups stick to GPU
Google made TPU, Tensor Processing Unit Ironwood
Trainium
Broadcom
Companies want to use their own chips
Edge AI, instead of the cloud Running AI on your device NPU Edge AI Chip, integrated into the phone A chip with many modules that do different stuff On-device AI, manage people’s privacy, efficient for us.
FPGAs Reconfigured after they are made But have lower raw performance
TSMC Biggest chip maker US need energy