Embedded spin transfer torque memory for cellular neural network based processing unit

An embedded STT memory, comprising:

• an dielectric layer formed on a top metal landing pad of a logical circuit;
• an adhesion and topography planarization (ATP) layer disposed on the dielectric layer, the ATP layer comprising a plurality of ATP pads each having a size, wherein:
• the plurality of ATP pads are in electrical contact with the top metal landing pad and also have a first dielectric material filled inbetween, and
• the ATP layer comprises:
• a top etch stop layer comprising Ta, Nb, Ti, Zr, W, TaN or TiN, and
• a bottom layer comprising WN, CrN, TiN, TaN, or NiCrN;
• a MTJ film layer disposed on the ATP layer, the MTJ film layer comprising a plurality of MTJ elements each having a size, wherein the plurality of MTJ elements have a second dielectric material filled inbetween;
• and
• at least one bit line formed on the MTJ film layer.

An integrated circuit processor having embedded spin transfer torque memory, comprising:

• a processing unit comprising a logical circuit having a plurality of transistors, the logical circuit having a top metal landing pad;
• an dielectric layer formed on the top metal landing pad of the logical circuit;
• an adhesion and topography planarization (ATP) layer disposed on the dielectric layer, the ATP layer comprising a plurality of ATP pads each having a size, wherein the plurality of ATP pads are in electrical contact with the top metal landing pad and also have a first dielectric material filled inbetween;
• a MTJ film layer disposed on the ATP layer, the MTJ film layer comprising a plurality of MTJ elements each having a size, wherein the plurality of MTJ elements have a second dielectric material filled inbetween;
• at least one bit line formed on the MTJ film layer;
• and
• a passivation layer disposed on the bit line with the MTJ film layer;
• wherein the logic circuit of the processing unit is configured to use at least data read from the bit line to perform one or more convolution neural network computations.