Whisper Model Processing Test The NVIDIA RTX 4090 is 86 seconds slower than the Apple M3 Max

Mondo Digital Updated on 2024-01-29

IT Home reported on December 14 that developer Oliver Wehrens recently tested Apple's M1 Pro, M2 Ultra and M3 Max three Apple Silicon chips to train AI capabilities after upgrading the MLX framework, and compared NVIDIA's RTX 4090 graphics card.

Wehrens tested using OpenAI's speech recognition model, Whisper, which measures the time it takes to transcribe a 10-minute audio file.

The test results show that the M1 Pro takes 216 seconds to process audioThe NVIDIA RTX 4090 graphics card takes 186 seconds to process.

The M2 Ultra with 76 GPUs and the M3 Max with 40 GPUs have better processing performance95 seconds and 100 seconds, respectively.

In addition, Apple's Apple Silicon chip also consumes more power. The NVIDIA RTX 4090 is 242W higher in the running state than in the idle state.

The M1 Pro chip is only 38W higher than the idle state in the running state.

IT Home previously reported that the MLX framework features the following:

Familiar APIs: Python and C++ APIs have familiar frameworks such as numpy and pytorch, making it easy for experienced researchers to learn and Xi.

Easy and efficient: MLX uses combinable function transformations to optimize the performance of Apple Silicon.

Deferred computation: Prevents unnecessary computation and improves resource efficiency.

Dynamic design: Ability to adapt to input shape changes simplifies debugging and testing.

Hardware and software: MLX seamlessly leverages the CPUs and GPUs of Apple devices, ensuring that users can get the most out of their hardware.

Unified Memory Advantage: MLX leverages Apple's Unified Memory to further accelerate data movement.

Researcher-friendly: MLX is designed for researchers.

Related Pages