Ambient Scientific unveils DigAn analogue in-memory compute for edge AI

The Silicon Valley startup introduced DigAn, a patented analogue in-memory compute architecture that embeds analogue MAC units inside SRAM, eliminating the 70–75% energy penalty of moving data between memory and processor. Its first production chip, the GPX10, targets battery-powered edge AI and claims a 100-fold power, performance and area gain over conventional 32-bit microcontrollers while staying programmable via TensorFlow and PyTorch.

• DigAn embeds analogue MAC units directly inside SRAM to remove data movement energy tax
• GPX10 claims 100x better power, performance and area than 32-bit MCUs
• Chip is fully programmable through standard TensorFlow and PyTorch workflows
• Roadmap spans single-core GPX1 to GPX64 and a simulated server-class chip