Nvidia chips feed AI monster at DoorDash

DoorDash coordinates drivers, known as Dashers, with customers and local restaurants through its logistics engine powered by machine learning and fast Nvidia graphics processors. Coronovirus is expected to push more people to order food deliveries. (DoorDash)

DoorDash is making use of Nvidia graphics processors to accelerate its AI training, a move that should come in handy as virus quarantines skyrocket and customers turn to meal delivery.

DoorDash, started in 2013, now allows customers to order meals via a smartphone from more than 300,000 vendors in 4,000 cities in the U.S., Canada and Australia. Its app and the platform logic behind it connect food merchants with customer and delivery contractors called Dashers.

Machine learning helps a logistics engine serve up personalized restaurant tips and delivery time information to customers seeking local food businesses. The engine also sorts through options to find the best routes and delivery prices, according to an Nvidia blog posted Wednesday. The modeling is complex partly because the data is always changing.

Sponsored by Infosys

Infosys positioned as a Leader in Gartner Magic Quadrant for IT Services for Communications Service Providers, Worldwide 2020

The Gartner Magic Quadrant evaluated 12 vendors and Infosys was recognized for its completeness of vision and ability to execute.
Infosys leverages its global partner ecosystem, CSP-dedicated studio, design tools, and 5G Living Labs to boost service delivery. Innovative solutions such as the ‘Infosys Cortex2’ are driving business value for CSPs.

With that complexity, the need for processing speed is paramount. By using Nvidia GPU’s for AI training connected to the company’s modeling, DoorDash was able to increase speeds by 10 times, according to Gary Ren, machine-learning engineer at DoorDash. In addition, DoorDash has migrated from single to multiple GPUs, to increase speeds another three times.

“Faster is always better for training speeds,” Ren said in the blog. “A 10x training speed-up means we spin up cloud clusters for a tenth the time, so we get a 10x reduction in computing costs.”

DoorDash has been using a recommendation engine for about two years and has found that giving customers recommendations is becoming more important, since customers sometimes don’t know what style of food they are looking for. “Recommending the right merchants can make a difference between getting an order or the customer going elsewhere,” he said.

DoorDash has posted a series of its engineering techniques online. In one such blog DoorDash notes that it is working on complex formulations to “consider the state of the world over time. There are often cases where it is better to wait to offer a delivery instead of doing so right away…We can build these kinds of tradeoff considerations directly into the model and amplify the efficiency and quality of the system.”

Processor speed will also potentially help DoorDash with Natural Language Processing, needed when consumers talk with support agents or Dashers.

Since data is always changing, including from weather events and even a possible coronavirus pandemic, inference speeds will need to constantly improve, Ren said. “Inference speeds are good enough today, but we’ll need to plan for the future,” he said.

Ren is expected to discuss the DoorDash use of AI in an online session at Nvidia’s GPU Technology Conference, GTU Digital, starting March 25.

RELATED: Nvidia blasts off with 41% revenue boost in 4Q

Suggested Articles

Lab inside ST fab in Singapore will bring together scientists from A * STAR Institute of Microelectronics and Japan’s ULVAC

The rush to test ventilators was “like sprinting down a pier while also building the pier”

Xilinx stock gained 7% on news of the $35 billion deal while AMD dropped nearly 5%.