And if STT-MRAM is never able to achieve the required density improvements, I’ll looking to ReRam and it’s 8-10 Gb/cm² density instead. Unlike STT-MRAM, which relies on magnetic states and complex multi-layer magnetic tunnel junctions, ReRAM employs a simpler two-terminal structure that creates and dissolves nanoscale conductive filaments within an oxide layer. Where STT-MRAM requires precise magnetic field control and separate access transistors, ReRAM's operation is based on voltage-induced ion migration, enabling a more compact cell structure. This fundamental difference shows up in both the fabrication process and cell size - ReRAM's simpler structure requires fewer mask layers and can achieve cell sizes of 4-6F² compared to STT-MRAM's 12-14F², though it trades this density advantage for lower write endurance and higher variability.

In the spectrum of emerging memories, ReRAM occupies a sweet spot between the high endurance but lower density of STT-MRAM and the high density but poor endurance of PCM. Its density of 8-10 Gb/cm² approaches that of DRAM (10-12 Gb/cm²) while offering non-volatility and significantly lower power consumption. Compared to NAND Flash, ReRAM provides orders of magnitude better write speed (50-100ns vs. microseconds) and endurance (10⁹-10¹⁰ vs. 10⁵ cycles). It surpasses FeRAM's density limitations while offering better scaling potential, though it cannot match FeRAM's write endurance. Most crucially for AI applications, ReRAM's analog nature enables direct multiplication-in-memory operations that neither DRAM nor most other emerging memories can natively support.

The technology's primary limitation is endurance, with typical devices achieving only 10⁹-10¹⁰ write cycles, making it less suitable for training workloads that require frequent weight updates. Device-to-device variation and resistance drift over time pose challenges for reliable multi-level cell operation and long-term stability. Research teams are exploring novel materials and architectures to address these limitations, including incorporating new switching layers, developing sophisticated error correction techniques, and implementing adaptive programming schemes. Major semiconductor companies and research institutions are also working on improving yield at advanced nodes and developing hybrid architectures that combine ReRAM's density advantages with complementary technologies for write-intensive operations.

Worth watching:

• Weebit Nano (Israel): Develops SiOx ReRAM technology with CMOS fab-friendly materials. Notable for production partnership with SK hynix and qualification of embedded ReRAM at 28nm with Tower Semiconductor.
• Crossbar Inc (USA): Pioneers silver-based ReRAM technology using CMOS-compatible materials and a unique selector device. Notable for their 3D stackable architecture and early demonstration of MLCs.
• Intrinsic Semi (UK): Focuses on RRAM-based compute-in-memory solutions specifically optimized for AI inference. Their architecture enables direct matrix multiplication within memory arrays.