Uber Improves CPU Efficiency With Normalized Compute Units

The platform now accounts for hardware performance gaps, cutting total Kafka CPU allocation by 8 percent across its fleet.

Updated on Oct. 27, 2026 in Data Centers

Isometric editorial illustration of server rack units of varying heights, representing data center infrastructure efficiency. — Uber’s Odin platform has deployed a normalized compute unit system to standardize hardware performance, resulting in an 8 percent reduction in Kafka CPU allocation. AI Illustration. Upload story photo >

Live Poll

Do you believe advanced technical infrastructure improvements will meaningfully improve the reliability of your digital services?

Uber has implemented a new resource management method in its Odin platform that uses Normalized Compute Units (NCUs) to account for hardware performance variability. By normalizing CPU capacity across its heterogeneous fleet, the company reduced Kafka CPU allocation by approximately 8 percent.

Why it matters

Managing large-scale infrastructure requires handling significant performance gaps between hardware generations. This approach optimizes resource density and efficiency across massive, stateful distributed systems.

The Odin platform maps hardware performance to a normalized baseline of 100, with faster server types reaching factors up to 190. This normalization reduced Kafka utilization variance across hardware types from 50 percent down to 10 percent.

The players

Uber

A multinational technology company operating large-scale distributed systems and managing 5 million CPU cores.

The details

Odin acts as an orchestration layer that translates abstract compute requirements into specific physical core assignments. Because single-threaded performance can vary by a factor of two across Uber's heterogeneous infrastructure, stateful workloads often struggle with resource imbalance. Odin resolves this by assigning a performance factor to each server type and converting raw usage into normalized units. This ensures that scaling logic accounts for the actual throughput capacity of the underlying silicon.

Uber implemented the normalized compute framework as of October 2026.

The Tech Race

As data centers incorporate increasingly diverse hardware, the ability to abstract physical performance from logical demand is becoming a critical competitive advantage. Uber's approach moves beyond simple core counting to achieve higher utilization density across its massive server fleet.

This optimization specifically affects the efficiency of Kafka workloads, which form the backbone of real-time data streaming at Uber. Engineers managing large, stateful compute clusters can look to this performance normalization model as a strategy for reducing over-provisioning.

The takeaway

By implementing performance-weighted allocation, Uber effectively bridges the performance gap between disparate server generations within its 5-million-core fleet. Observers should track whether this normalization strategy becomes an industry standard for managing stateful workloads in heterogeneous cloud environments.