As artificial intelligence continues to evolve, the demand for electricity in data centers has surged. However, current processing techniques have led to inefficiencies, forcing operators to reduce their power consumption by up to 30%.
During a keynote at Nvidia's annual GTC conference, CEO Jensen Huang highlighted the issue, stating, "Every unused watt is revenue lost." This issue has prompted the emergence of innovative solutions in the sector.
Today, Tel Aviv-based startup Niv-AI has officially launched with $12 million in seed funding, aiming to tackle the challenge of GPU power management. By employing advanced sensors to accurately monitor power usage, Niv-AI plans to develop tools that enhance efficiency in data centers.
Founded last year by CEO Tomer Timor and CTO Edward Kizis, Niv-AI is supported by notable investors including Glilot Capital and Grove Ventures. While the company has not disclosed its valuation, its mission is clear: to optimize the power dynamics within data centers.
As data centers operate thousands of GPUs simultaneously, they experience rapid fluctuations in power demand as processors shift between tasks. This variability complicates power management, often resulting in costly solutions like temporary energy storage or throttled GPU usage, both of which diminish the return on investment in high-end hardware.
Lior Handlesman, a partner at Grove Ventures and a board member at Niv, noted, "We just can't continue building data centers the way we build them now."
Niv-AI's initial strategy involves deploying sensors that monitor power usage at the millisecond level across its GPUs and those of its design partners. The objective is to analyze the power profiles of various deep learning tasks and devise strategies to maximize existing capacity.
The engineers at Niv-AI plan to leverage the collected data to train an AI model that can predict and synchronize power loads across data centers, acting as a "copilot" for engineers managing these facilities.
In the coming six to eight months, Niv-AI anticipates having operational systems in several U.S. data centers. This initiative comes at a crucial time as hyperscalers face challenges related to land use and supply chain logistics when building new data centers. The founders envision their solution as an essential "intelligence layer" that bridges data centers and the electrical grid.
Timor remarked, "The grid is actually afraid of the data center consuming too much power at a specific time." He emphasized the dual nature of the problem: enabling data centers to utilize more GPUs while fostering responsible power profiles between the data centers and the grid.