Uber intends to use custom Ampere Computing Arm processors in the Oracle Cloud Infrastructure (OCI) cloud. In particular, companies will optimize chips for AI tasks. Until recently, Uber used primarily its own data centers, but in 2022 it decided to move most of its tasks to the cloud.
As the number of data centers and availability zones grew, managing Uber’s IT infrastructure became increasingly difficult. Putting a new zone into operation sometimes required months and hundreds of specialists, since servers were managed almost “manually”, and automation tools often failed. After considering various development options, Uber gradually came to cooperate with Ampere, Goolge and Oracle. And in February 2023, the company signed large seven-year cloud contracts with Google and Oracle.
According to Uber, the company’s drivers and couriers complete more than 30 million orders daily. This requires a large IT infrastructure, for example, to optimize routes, including using AI technologies – 15 million requests come to AI models every second. Uber and Ampere estimate that moving loads to OCI not only reduced infrastructure costs, but also reduced energy consumption by 30%. The companies are now working together on new chips, identifying what changes to the microarchitecture should be made to ensure that future processors are optimally suited for Uber’s tasks.
Image source: Denys Nevozhai/unsplash.com
AWS, Google Cloud and Microsoft Azure are developing their own Arm processors – Graviton, Axion and Cobalt 100, respectively. However, they do not engage in customization for a specific client, even a large one. However, IDC experts believe that cloud clients will certainly benefit from joint preparation of new semiconductor solutions with chip developers. Clients may have specific knowledge and intellectual property, but usually do not have the ability to bring the finished product to market on their own.
Uber is now migrating thousands of microservices, numerous data storage platforms, and dozens of AI models to OCI. The company has already migrated a significant portion of its serverless workloads to Ampere-based platforms. However, the matter is not limited only to Ampere – the company also actively uses instances based on AMD chips.