AI chip from AMD is behind Nvidia in speed: MosaicML

Advanced Micro Devices (AMD) has developed artificial intelligence chips that are approximately 80 percent as fast as Nvidia’s chips, according to a report by AI software firm MosaicML.
Nvidia chip businessThe report suggests that AMD has the potential to match Nvidia’s performance in the future, Reuters news report said.

Nvidia dominates the market for AI chips, which are widely used in technologies like ChatGPT and other AI services. The popularity of these services has propelled Nvidia’s value beyond $1 trillion. The high demand for Nvidia chips has resulted in a shortage, prompting tech companies to seek alternatives. AMD is seen as a strong contender in this regard.

MosaicML, an AI startup recently acquired for $1.3 billion, conducted a test comparing AI chips from AMD and Nvidia. The evaluation focused on the AMD MI250 and the Nvidia A100, both slightly older generations of their respective flagship chips but still in high demand.

The test revealed that AMD’s chip can achieve 80 percent of the performance of Nvidia’s chip, primarily due to a new version of AMD software released in late 2022 and an updated version of the open-source software PyTorch, supported by Meta Platforms, released in March 2023.

Hanlin Tang, the chief technology officer of MosaicML, stated that the company believes AMD’s upcoming software updates will enable its MI250 chip to match the performance of Nvidia’s A100. Tang also emphasized the strength of AMD’s software, which is often a weak point for many machine learning chip companies.

MosaicML utilized its tools, PyTorch and AMD software, to train a large language model without requiring any changes to its code base. Tang mentioned that if developers can find AMD’s chips at an appropriate price, they can already switch to them as they are essentially interchangeable with Nvidia chips.

MosaicML develops software that facilitates the creation of AI systems within a company’s data centers, reducing reliance on external providers like OpenAI for access to such systems. The company conducted this research to demonstrate that there are chip options available beyond Nvidia.

AMD responded to the findings, stating that they align with their strategy of supporting an open and easy-to-implement software ecosystem for AI training and inference on their hardware. AMD intends to continue collaborating with MosaicML to optimize its software further.