TetraMem tapes out world's first 12nm analog in-memory-computing AI chip

Silicon Valley analog in-memory-computing company TetraMem has announced the successful tapeout of MLX300. The company describes it as the world's first 12nm analog in-memory-computing (A-IMC) system-on-chip based on multi-level RRAM technology, developed using Synopsys cloud-enabled electronic design automation tools.

Computing inside memory

Conventional AI chips spend time and energy moving data back and forth between processors and memory. Analog in-memory computing lets the memory cells themselves take part in computation, cutting down that data movement. TetraMem says MLX300 is a versatile single-chip accelerator built on that architecture, for audio, vision and small language models.

According to the company's specifications, MLX300 holds up to 64 million INT4 parameters on-chip, with no external DRAM needed for model weights, and targets energy efficiency of more than 50 to 80 TOPS/W. That efficiency figure is a design target, not a measured result from a shipping product.

Second tapeout this year

MLX300 is TetraMem's second full-mask tapeout in 2026, following the MLX200 and MLX200M on a commercial 22nm process. The company says both designs used Synopsys EDA tools delivered through Synopsys Cloud FlexEDA, letting engineering teams in the United States and Singapore work together; design-environment bring-up fell from weeks to minutes, and register-transfer-level verification simulation ran about 3 times faster than on-premises methods.

TetraMem CEO and co-founder Glenn Ge said the architecture converges memory and compute, calling it a fundamentally new approach to AI computing. Synopsys said the work reflects the kind of innovation needed for the next wave of AI performance and efficiency. A tapeout is the milestone of sending a design to fabrication; the company gave no mass-production timetable in this announcement.