GPU Interior: from architecture to trillions of computations

We often discuss modern GPU performance in abstract terms—trillions of operations, ever-increasing computational power. Less visible, however, is the underlying structure that makes it all possible. A video by Branch Education, using the GA102 chip from an NVIDIA RTX 3090 as a reference, helps explain what happens inside a modern graphics card. Its central premise is a simple yet revealing question: how can a single device perform tens of trillions of calculations per second? We learn that some high-end GPUs can handle over 36 trillion operations per second, with newer generations pushing this number even higher. The answer lies in GPU design: unlike CPUs, which handle tasks sequentially, GPUs are built for parallel processing, executing thousands of smaller operations simultaneously—an approach typically described as Single Instruction, Multiple Data (SIMD). This architecture, combined with high-bandwidth memory like GDDR6X, enables GPUs to efficiently handle data-intensive workloads, from rendering and transforming objects to increasingly complex AI tasks. The video shows that this performance stems not from a single breakthrough, but from the interplay of these elements in a tightly integrated system, reflecting the broader role GPUs now play in high-performance computing