Ad tech leader, Cognitiv, brings AI inference to the edge

Cognitiv powers split-second ad decisions with AI at the edge

Ad tech provider Cognitiv.ai runs its deep-learning ad placement platform in Equinix data centers, which deliver the ultra-low latency and full control over hardware necessary to optimize for AI inference at the edge. For latency-sensitive AI applications like real-time bidding, even small delays can significantly impact outcomes, making ultra-low latency a strategic requirement. 

The result is the ability to process millions of transactions per second at a significantly lower cost than an equivalent public cloud environment. By combining edge deployment with a hybrid approach to infrastructure, Cognitiv balances performance, control and cost more effectively than a purely cloud-based model.

In ad tech, every millisecond counts

Cognitiv, a leading ad tech company powered by deep learning, has used deep learning and AI to optimize digital ad placements for over a decade. “We didn’t just pivot to AI when it became cool,” says Michael Goncalves, the company’s director of engineering operations. 

In digital advertising, decisions must be made in real time as bid requests are evaluated and acted on in milliseconds, making latency a critical performance factor. To work effectively, the company’s proprietary next-generation Deep Learning Advertising Platform (DLAP) requires extremely low network latency, as well as specialized hardware capable of supporting split-millisecond computations. The platform evaluates massive volumes of bid requests and applies deep learning models in real time to determine optimal ad placements.

Optimizing infrastructure for AI

“Meeting these requirements using public cloud or traditional on-premises infrastructure isn’t really achievable,” Goncalves says, due to higher network latency rates and limited control over hardware under those setups.

Public cloud environments can introduce unpredictable latency and limit the ability to fine-tune hardware performance for specialized AI workloads.

Partnering with Equinix gives Cognitiv the low latency and full hardware control needed to drive its complex, path-breaking AI models.

By deploying inference infrastructure closer to users and data sources, Cognitiv minimizes latency and enables real-time responsiveness.

With Equinix AI-ready data centers, Cognitiv runs AI inference workloads in multiple global regions. This hybrid approach allows Cognitiv to combine the scalability of cloud with the performance and control of private infrastructure. Colocating with its customers in the same facilities allows real-time evaluation of millions of bid requests per second, maintaining end-to-end latency in the low single digits.

Operating at AI scale

In addition to supercharging the performance of Cognitiv.ai’s deep learning platform, Equinix data centers also help the company scale operations and customer reach. A globally distributed infrastructure allows Cognitiv to extend its platform closer to partners and customers, ensuring consistent performance worldwide. By colocating workloads in the same data centers used by clients and partners, Cognitiv.ai is able to offer its solution as an add-on to their own services with extremely minimal latency.

Benefits

Minimal latency

End-to-end latency rates in the low single digits. Deploying inference infrastructure close to users ensures near-instant decision-making for time-sensitive AI workloads.

Cost optimization

Significantly lower total lifetime cost than major clouds across all hardware. A hybrid infrastructure model reduces reliance on expensive public cloud resources while optimizing long-term operational costs.

Hardware performance boost

High storage performance for NoSQL databases and memory optimally tuned to services’ caching requirements. Full control over hardware enables Cognitiv to fine-tune infrastructure for its specific AI inference workloads. 

Real-time AI responsiveness

Edge-based inferencing enables the platform to process and respond to millions of events per second without latency bottlenecks.