Our Solution
- E2E DNN accelration in compact SoC with {High-Density MRAM, Efficient AMS-PIM, Custom DNN SW}
Embedded Files
MRAM-Based PIM
Best candidate for on-device DNN (store + processing)
High Density NVM for most advanced nodes
eFlash (embedded Flash)
Cannot be scaled <28nm
Low endurance and large cell size
SRAM
(Large cell size) & (scaling saturates at <28nm)
Losses data when turned off: requires external memory
AMS-PIM
Analog-Mixed-Signal (AMS) Processing Technique for high-efficiency and performance
High bandwidth, efficiency, and reliability simultaneously achieved
More energy efficient & faster than digital PIMs
More reliable, scalable, and area-efficient than analog PIMs
Simulator & SDK
SW-defined, cycle-accurate HW simulator
SDK for PIM-friendly DNN optimization
Quantization SW co-designed for PIM
Nonlinear layer SW co-designed for PIM
PTQ/QAT pipeline co-designed for PIM
Corresponding, area/energy-efficient, with-SW-co-designed HW library
Page updated
Google Sites
Report abuse