Inside a modern GPU: from architecture to billions of calculations

We often discuss the performance of modern GPUs in abstract terms—billions of operations, ever-increasing computational power. What remains less visible is the structure that makes it possible. This is where a video from Branch Education comes in, using NVIDIA's GA102 chip from the RTX 3090 as a benchmark to show what happens inside a modern graphics card. At its core is a simple but revealing question: how can a single device perform tens of trillions of calculations per second? According to the video, some high-end GPUs are capable of exceeding 36 trillion operations per second—and newer generations push that number even further. The answer lies in how GPUs are designed. Unlike CPUs, which process tasks sequentially, GPUs are built for parallel processing, enabling them to execute thousands of smaller operations simultaneously. This is typically described as a Single Instruction, Multiple Data (SIMD) approach, where the same instruction is applied to multiple data points at once. This architecture, combined with high-bandwidth memory like GDDR6X, allows GPUs to efficiently handle data-intensive workloads—from rendering and object transformations to increasingly complex artificial intelligence tasks. More than a single breakthrough, the video shows that performance results from how these elements work together within a highly integrated system, reflecting the ever-broadening role GPUs play in high-performance computing