ZettaLith: zettaFLOPS of AI inference in a < 100 kW rack, exaFLOPS on the desktop
Year: 2025
Type: article
Abstract: The immense computational cost and power consumption of large AI models, particularly Transformers, represent a critical barrier to their widespread deployment and further scaling. Current hardware approaches face fundamental efficiency limits. This paper introduces ZettaLith, a novel, scalable computing architecture designed to slash the cost and power of AI inference by over 1,000× compared to current GPU-based systems. Specifically, a single flagship ZettaLith rack is projected to achieve 1.507 zettaFLOPS in 2027 - 1,047× faster, 1,490× more power efficient, and 2,325× more cost-effective than Nvidia’s current flagship GB200 NVL72 for FP4 transformer inference. The ZettaLith architecture achieves these significant gains not through speculative new science, but by jettisoning general purpose GPU applications, and via the multiplicative effect of numerous co-designed architectural innovations using established digital electronic technologies, as detailed in this paper. ZettaLith’s core architectural principles scale down efficiently to exaFLOPS desktop systems and petaFLOPS mobile chips, maintaining their roughly 1,000× advantage. Analysis is provided supporting ZettaLith's potential to fundamentally reshape the economics and accessibility of large-scale AI inference. Furthermore, ZettaLith presents a radically simpler system architecture compared to the complex hierarchy of current GPU clusters. ZettaLith is optimized exclusively for AI inference and is not applicable for AI training.
Author: Silverbrook, Kia
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6893260868
FWCI:
Cites:
Cited by:
Related to: