The rise of artificial intelligence (AI) has been making a huge impact on our daily lives, especially in the fields of recommendation systems, image recognition, natural language processing, and autonomous driving. As the amount of input data, weight parameters and intermediate data in the machine learning process grows exponentially, memory becomes a critical bottleneck, which requires a high density, low power and high speed non-volatile memory (NVM) solution. Among emerging NVM technologies, spin-transfer torque magnetoresistive random access memory (STT-MRAM) based on perpendicular magnetic tunnel junctions (pMTJ) shows distinct advantages:

  1. High speed (comparable to DRAM and LLC), low standby power (vs eSRAM) and low active power deliver high tera operation per second (TOPS) per watt during the training and inference process.
  2. Practically unlimited endurance allows large amount of data to be intensively processed.
  3. High data retention allows weight parameters of neural network to be directly stored in NVM without extra power consumption (DRAM refreshing or SRAM leakage), especially for edge computing.
  4. Excellent scalability to advanced technology nodes (beyond 7 nm) allows ASIC design with high density (vs eSRAM) for specific AI applications.

We present advanced MRAM chip data at 2x nm technology nodes on 300 mm wafers (manufactured at our foundry partner UMC) with high switching efficiency and thermally stable pMTJ (400 °C for 3 hours), chemical-damage-free MTJ etch process, and scalable BEOL integration scheme. To achieve fast switching speed, MTJ stack is optimized for lower switching voltage at short pulse width. We further demonstrate high speed/low power performance down to 10 ns with 32Mb fully functional macro, along with 125°C/10 years data retention and >1012 endurance. BER and reliability improvement is achieved with MTJ stack and integration process optimization, enabling high manufacturing yield. MTJ property can be further tailored to meet specific application needs with no/minimal CMOS design change.

Our advanced STT-MRAM technology will enable a broad range of AI applications, which are currently limited by existing memory performance and computing architecture. Furthermore, such high performance and high density STT-MRAM will certainly stimulate further breakthroughs in AI technologies, such as in-memory computing.