Uber embraces the cloud with customized CPUs

When Uber decided in 2022 to shift away from running its own data centers, the ridesharing and delivery company wanted a high level of control for how its workloads ran in the cloud, down to the CPU level.

Uber, which had formerly housed more than 90% of its infrastructure in its own data centers, turned to Oracle and Google in 2023 to be its cloud providers in an effort to shift nearly “everything” away from its own on-premises infrastructure.

Now, Uber is partnering with Ampere Computing to give it more control over how its workloads run on Oracle Cloud Infrastructure. Ampere, a US fabless semiconductor firm founded in 2017 by former Intel President Renée James, began supplying chips to Oracle in 2021. Now, 95% of Oracle’s services run on Ampere infrastructure, according to the semiconductor company.

Ampere launched a family of chips custom-designed for cloud computing providers in 2023, and it recently unveiled a 512-core chip with integrated AI acceleration.

Ampere’s ability to customize chips for cloud workloads appealed to Uber, because of its heavy and unique computing needs, says Kamran Zargahi, Uber’s senior director of technology strategy. Uber drivers and couriers complete more than 30 million trips per day, and its computing infrastructure must be capable of a massive number of route optimization and other predictions each second.

“Every second, 15 million AI models get executed,” Zargahi says. “Everything that Uber actually does is included in that number.”

Total chip control

Uber’s huge computing needs drove the company to seek an unusual level of control over the inner workings of its cloud computing providers. The new partnership between Uber, Ampere, and Oracle, announced during Oracle CloudWorld 2024, gives the ridesharing company the ability to co-design chips with Ampere.

“At Uber, we very much like to go deep into the tech,” Zargahi says. “We just didn’t use the cloud as a tool and just stay there comfortably. We really wanted to be able to understand the regional and the zonal shapes, the price performance, and all the specifics of the bits and bytes.”

Uber, Oracle, and Ampere are collaborating on chip design with a focus on both high performance and power efficiency in the cloud, says Jeff Wittich, Ampere’s chief product officer. Uber’s move to Ampere on Oracle Cloud Infrastructure enabled the company to reduce infrastructure costs while cutting power consumption by 30%, he adds.

“This was our strategy from the start, was to build something that was exceptionally well suited for the cloud, versus taking something that was built maybe for other types of use cases, other environments that existed 10, 20, 30 years ago, and just trying to adapt what legacy players do,” he says.

The partnership has Ampere and Uber engineers working “hand-in-hand” together on chip design, Wittich says.

“As we look forward to the future, the common team has been working closely on, what are those future optimizations we can make to further tune the micro-architecture of the CPU, even down to that level, to make sure that all of our future generations of CPUs are also incredibly well suited for Uber,” he adds.

Ampere is an ARM processor architecture licensee, and cloud service providers AWS, Google, and Microsoft have also customized ARM-based CPUs, notes Shane Rau, research vice president for computing semiconductors at IDC. It also may benefit cloud users to work with a semiconductor company such as Ampere to co-design CPUs, he says, adding that such a partnership would bring cloud customers the tools, relationships, and technology they need.

“Usually, companies that co-design a CPU with a semiconductor company have their own specific piece of IP and their own special set of workloads and customer types to support but they lack the capabilities to bring a product with that IP to the market,” Rau says.

Calculated route to the cloud

Uber’s cloud journey may sound familiar in some ways. The company, with a beta launch in mid-2010, chose to operate its own data centers because the modern cloud computing market was still in its infancy. AWS was less than a decade old, and Microsoft rebranded Azure, originally Windows Azure and about a half-decade old, just months before Uber’s beta launch.

While most of Uber’s computing workloads remained on-premises, the company used cloud services “opportunistically” where it made sense, Zargahi says. But after an extensive technology review in 2022, the company decided to move away from operating its own data centers and focus instead on its core business of moving passengers and delivering restaurant meals, groceries, and consumer goods.

Uber was constrained by its on-premises infrastructure, the company wrote in a September 2022 blog post. The company was primarily using manual operations for managing servers.

“The size of our server fleet was growing rapidly, and our tooling and teams weren’t able to keep up,” the blog post says. “The automated tooling we did have was constantly breaking down. Both the manual operations and automated tooling were frequent outage culprits.”

At the same time, the growth of Uber’s fleet required the company to expand into more data centers and availability zones. Turning up a new zone took several months and involved hundreds of engineers.

“Our server fleet consisted mostly of on-prem machines, with limited ability to take advantage of additional capacity that was available in the cloud,” the blog post says. “We had a single, fledgling cloud zone but manual operations implied that we were not really taking full advantage of the cloud.”

Part of Uber’s cloud journey has been partnering with companies like Ampere and Oracle. It makes sense for Ampere Computing to offer customized chips to cloud customers, says IDC’s Rau. Ampere’s recent focus is on making its CPU architecture more relevant and accessible, he says.

Recent partnerships with Oracle Cloud Infrastructure and Qualcomm demonstrate Ampere’s focus on becoming more available on the wider market, he adds.

Customized chips can open a new market for Ampere, he says.

“In addition to offering standard products on a merchant basis, offering custom products brings a semiconductor company’s processing architecture, development tools, and other parts of their solution stack to the customers who want to put their own secret sauce into a CPU design,” Rau adds.