Internet company Amazon.com announced on Thursday that it is transferring some of the computing operations that power the voice assistant Alexa to chips of its own design, thus ditching NVIDIA’s chips. This is expected to make Alexa run faster while lowering Amazon’s costs.
When users of devices such as smart speakers from the Amazon Echo family ask a question to the voice assistant, the request is sent to one of the Amazon data centers to process and form a response. This response is in text format, and then it must be translated into audible speech, which will be delivered by the voice assistant.
Previously, all of these operations were done using NVIDIA chips, but now “most” of the process will be done using the Inferentia computing chip, made by Amazon itself. First announced in 2018, Amazon’s processor is specifically designed to accelerate large-scale machine learning tasks such as text-to-speech or image recognition.
Cloud computing service owners such as Amazon, Microsoft and Google have become some of the largest consumers of AI computing chips for their datacenters, leading to a boom in specialized solutions from Intel, NVIDIA and others. But large tech companies are increasingly ditching traditional semiconductor vendors to develop their own processors. In the same row, you can put Apple, which announced its first Mac computers with its own M1 processor with Arm architecture, having refused Intel.
Amazon said switching to an Infertia chip to handle some of its Alexa query processing operations resulted in a 25% decrease in latency while reducing costs by 30%.
The company also said its cloud-based facial recognition service Rekognition has also begun rolling out Inferentia chips. However, she did not specify which chips were previously used by the face recognition service, and how much of the operations were transferred to her own chips.
If you notice an error, select it with the mouse and press CTRL + ENTER.