Amazon Halts Development of Inferentia AI Inference Chips to Focus on Trainium Training Chips
Since announcing its entry into the AI chip market in 2018, Amazon AWS has developed Inferentia chips for inference and Trainium chips for training AI models.
Rahul Kulkarni, Director of Compute at Amazon AWS, stated, “The two product lines will converge. We will focus on Trainium to deliver both inference and training performance.”
Inferentia was designed to perform smaller computations at lower costs. However, as generative AI continues to evolve, both training and inference increasingly require greater computational capacity. The benefits of using separate chips for these tasks are diminishing.
Trainium, in contrast, offers larger memory capacity and supports a wider range of data formats. It also includes mechanisms for fast computation and communication when handling massive datasets across multiple servers.
At its annual tech conference in Las Vegas, Amazon announced the launch of the Trainium2 chip. The company also revealed plans to release the Trainium3 chip in the second half of 2025. Trainium3 will be manufactured using an advanced 3nm process, with computing performance expected to double that of Trainium2.
Amazon aims to challenge NVIDIA's dominance in the AI chip market, where NVIDIA currently holds approximately 90% of the global share. Amazon has been aggressively promoting Trainium2 to cloud customers, emphasizing its lower operational costs compared to NVIDIA products.
Apple has announced it will use Trainium2 chips for its AI development. Additionally, U.S.-based AI startup Anthropic, which has received an $8 billion investment from Amazon, is acquiring hundreds of thousands of Trainium2 chips to assist in developing its generative AI models.