GPU

A massively parallel processor — optimized for throughput over latency.

A massively parallel processor — optimized for throughput over latency.

Thousands of simple cores executing the same instruction on different data (SIMD/SIMT)

Originally designed for rendering, now critical for ML training and inference.

Available as discrete add-in cards or integrated units sharing the SoC die with the CPU

Two main uses: (1) rendering and (2) AI inference.

NVIDIA RTX 3090/4090: discrete GPU for PC (PCB with fans, fits inside a tower)