SANTA CLARA, CA, September 22, 2026 (EZ Newswire) -- PhantaField has announced an architecture for AI processors that places layers of memory directly interlaced in the circuits that use their data, an approach intended to reduce the distance information travels within the processor, and also stack compute along with memory. The company’s Sophon architecture, introduced in June 2026, combines atomically thin semiconductor materials with vertically stacked logic and memory. The architecture addresses a data movement constraint in AI processing, where computing capacity can depend on the ability to supply processors with data at sufficient speed.
“We enable a whole new way to design the chip by 3D monolithic integration of logic and memory,” says Dr. Xuejun Sheldon Xie, PhantaField’s founder and CEO. The architecture uses two-dimensional transition-metal dichalcogenides, or 2D TMDs, as semiconductor materials. Their active layers can be only a few atoms thick, grown under low temperature, allowing additional material layers to be incorporated above existing circuitry.
PhantaField is developing deposition equipment that uses plasma and photons to grow the materials at temperatures intended to preserve circuitry beneath them. The approach is designed to support the construction of additional computing and memory layers above an existing silicon base.
PhantaField’s modeling indicates that the proposed architecture may increase the rate at which model weights are supplied to computing circuits by 100X compared with the HBM-based reference configuration used in its analysis. The findings remain theoretical at this stage and do not yet establish how the architecture would perform in hardware or how it would compare with existing memory and packaging technologies in practical applications.
Another key property is radiation hardness. 2D TMD materials are natively radiation-hard, thus increasing the lifespan in space AI data centers. The 2T0C memory architecture is more tolerant to radiation-induced bit flips than 1T1C memory used in HBM. Given the high-density thin stack of compute and no conductive substrate under their thin-film 2D TMD transistor structure, radiation-induced secondary electrons will have less chance of damaging the chip and affecting the data.
The projected performance remains subject to hardware validation. PhantaField’s Sophon white paper describes simulations and calculations while identifying further work involving device calibration, thermal analysis, and manufacturing qualification.
Hardware testing will determine whether the architecture can deliver the modeled performance under operating conditions and whether the manufacturing process can produce the structures consistently.
The manufacturing process is therefore part of the development effort alongside the chip architecture. The proposed materials would need to be deposited uniformly and integrated with existing circuitry across repeated production runs before the approach could support commercial processor manufacturing.
PhantaField has stated a roadmap targeting product availability in 2028, with prototype validation expected to precede commercialization.
The company has also identified thePantheon.ai as a project associated with the Sophon processor. The development of the processor and its data center could be incorporated into that project and other computing applications.
Media Contact
Xuejun Sheldon Xie
info@phantafield.com
