Amazon said on Thursday that it transferred part of the calculations of its Alexa voice assistant to its own custom-designed chip, aimed at making work faster and cheaper, while removing it from the chip provided by Nvidia.
When users of devices such as Amazon’s Echo series of smart speakers ask a voice assistant a question, the query will be sent to one of Amazon’s data centers for multiple processing steps. When Amazon’s computer sends out an answer, the answer will be in text format and must be translated into a voice that can be heard by the voice assistant.
Amazon used to use Nvidia chips to handle this calculation, but now “most” will use its own Inferentia computing chips. The Amazon chip was first released in 2018 and has been custom-designed to speed up a large number of machine learning tasks, such as translating text into speech or recognizing images.
Cloud computing customers such as Amazon, Microsoft, and Alpahbet’s Google have become the largest buyers of computing chips, driving the boom in data center sales of companies such as Intel and Nvidia.
However, major technology companies increasingly abandon traditional chip suppliers to design their own chips. Apple on Tuesday introduced the first Mac computer with its own central processing unit instead of using Intel chips.
Amazon said that shifting to the Infertia chip due to part of its Alexa work resulted in a 25% increase in latency, which is a measure of speed, while reducing costs by 30%.
Amazon also said that its cloud-based facial recognition service Rekognition has begun to use its own Inferentia chip. However, the company did not disclose which chips have been used in facial recognition services, or how much work has been transferred to its own chips.
Since the service has been used by law enforcement agencies, it has been reviewed by civil rights organizations. After George Floyd was killed, Amazon suspended police use for a year in June.
© Thomson Reuters 2020
This is the best TV under Rs. 25,000? We discussed this on the weekly technical podcast Orbital, you can subscribe via Apple Podcast, Google Podcast or RSS, download the episode, or click the play button below.