Low GPU Usage
Software Workflow:
The software works by first recognizing text from the audio in a video, then translating the text into a target language, synthesizing voiceover based on the target language, and finally merging the text, voiceover, and video into a new video. Heavy GPU usage is primarily during the audio-to-text transcription phase, while other phases use little or no GPU.
GPU vs CPU: Principles and Differences
Imagine training a large AI model as moving bricks.
The CPU is like an "all-around player" who can handle many tasks: calculations, logic, and management, no matter how complex. However, it has a small number of cores, at most a few dozen. No matter how fast it moves bricks, it can only move a few or, at most, a few dozen at a time, resulting in low efficiency.
On the other hand, the GPU has a frightening number of cores, easily reaching thousands or even tens of thousands. Although each core can only move one brick, the sheer number of them makes up for it! With thousands or tens of thousands of "minions" working together, the bricks are moved quickly.
The core task of AI training and inference is "matrix operations" – simply put, a large number of numbers are lined up for addition, subtraction, multiplication, and division, like a massive pile of red bricks waiting to be moved, a simple task that doesn't require much "brainpower".
The GPU's ability for "massive core parallelism" comes in handy, allowing it to process thousands or even tens of thousands of small tasks simultaneously, making it tens or even hundreds of times faster than a CPU.
What about the CPU? It is better suited for serial and complex tasks, such as playing a single-player game or writing a document. However, there are too many bricks for AI, and it can only move a few or a few dozen at a time, making it impossible to keep up with the GPU.