Both chips dedicate ~half of the TRIP AI computation accelerators to SRAM, so effective memory bandwidth for calculations in SRAM cache is about an order of magnitude greater than DRAM bandwidth