STT MRAM for Artificial Intelligence Applications
- 1. Avalanche Technology at Semicon Taiwan 2020 STT-MRAM for Artificial Intelligence Applications The Next-Generation MRAM Company Yiming Huai EVP, Tech & Foundry Partnership Silicon Valley, California, USA
- 2. Avalanche Technology at Semicon Taiwan 2020 Introduction STT-MRAM: a Disruptive Embedded Memory Solution for AI Applications STT-MRAM Integration and MP with Foundry Partners Performance and Reliability Results at 22nm Chip Level Summary Outline 2
- 3. Avalanche Technology at Semicon Taiwan 2020 pMTJ STT-MRAM Proprietary Selector Avalanche Technology — at a Glance Next generation Magnetic Memory company creating a new multi-billion dollar market Strong partnerships 3 Embedded Stand-alone Licensed to foundries and IDMs • 108MHz(Max) QSPI 1-16 Mb Persistent SRAM (in MP) • 35ns Parallel 1-32 Mb Persistent SRAM (in MP) • 35ns Parallel 256 Mb -1Gb Persistent SRAM (2021) Manufactured at foundry partners • (22 nm) 8-64 Mb macros ( 2020) 300 mm production line Pre-eminent STT-MRAM Technology Strong IP Portfolio: 300+ granted patents WW World Class Team: MTJ, Integration, CMOS design, Testing/Reliability
- 4. Avalanche Technology at Semicon Taiwan 2020 What is STT-MRAM? High Performance, Nonvolatile, Unlimited Endurance… Memory Element: MTJ (Magnetic Tunnel Junction) Information stored by magnetic polarization (nonvolatile) instead of charge MTJ bit state “1” (high resistance) and “0” (low resistance) is written by Spin Transfer Torque with a (polarized) current across MTJ Extremely Fast (as LL Cache/DRAM) Nonvolatile (Persistent) Unlimited endurance (>1014) High Density (1T per cell) Scalable to 0x nm STT-MRAM cell : 1T+MTJ 4
- 5. Avalanche Technology at Semicon Taiwan 2020 Stand Alone Applications STT-MRAM Broad Applications STT- MRAM Embedded Applications Unified eNVM (Flash like) eFlash, eOTP, eFuse LL Cache Memory (SRAM like) L3, eDRAM Slow SRAM (New Market Applications) (AI, IoT…) One single chip for both embedded storage and working memory nvSRAM market Memory buffers Persistent DRAM DRAM* New Market Applications* Storage Class Memory *with 3D stack MRAM High speed Unlimited endurance Low power consumption Low manufacturing cost Extended Temperature (150 oC) Y. Huai, Flash Summit 2015, Santa Clara, California, August 12, 2015. 5
- 6. Avalanche Technology at Semicon Taiwan 2020 eFlash scaling is reaching its end eMRAM (Flash type) offers ‒ Scalability beyond (<) 22nm node ‒ Lower cost (~3 masks) ‒ >100x speed improvement ‒ >>1000x endurance improvement SRAM scaling slowing down with large footprint & high cost at advanced nodes eMRAM (SRAM type) offers ‒ Higher density ‒ Lower leakage power ‒ Efficient architecture STT-MRAM: Cheaper and Better STT-MRAM vs Existing Embedded Memory eSTT-MRAM eFlash SRAM eDRAM Cell Size (F2) 30-50 30-50 100-500 30-90 Additional Masks 2-3 10-20 0 4-6 6
- 7. Avalanche Technology at Semicon Taiwan 2020 Embedded Memory Requirements for Edge AI Applications 1) High speed (comparable to LLC), low standby power (vs SRAM) and low active power deliver high Tera Operations Per Second (TOPS) per watt during training and inference processes. 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 (SRAM leakage), especially for low-power edge computing. 4) Excellent scalability to advanced technology nodes (0x nm) allows ASIC design with high density (vs SRAM) for specific AI applications. 7
- 8. Avalanche Technology at Semicon Taiwan 2020 STT MRAM Provides Ideal Embedded Solution for Edge AI Smart IoT Edge processing, lower latency, better data security • Local APIs • Local data processing • Better connectivity • Intelligence and smart features • Dense memory • Low power • High speed • Low cost Needs: 8
- 9. Avalanche Technology at Semicon Taiwan 2020 Breakthrough pMTJ Performance Enabling LLC Co (2A) Ni (6A) Co (2A) Ni (6A) Co (2A) Ni (6A) Ir Amorphous CoFeB Barrier MgO (RA~5-10) Novel Free layer Multilayer (Co/Pt or Co/Ni) Thin magnetic layer W pMTJ delivered all attributes required for a disruptive NVM and can be tailored for different applications Embedded Large TMR Thermal stable MTJ Fast writing to 5-10 ns Yiming Huai, et al., Applied Physics Letters 112, 092402, 2018 Low WER (ppm) With low Vc (low power) 9 (a) (b) (c) (d)
- 10. Avalanche Technology at Semicon Taiwan 2020 Flexible/Compatible with CMOS Flexible MTJ placement between any two metal layers Lower Cost Chemical-damage-free MTJ etch with additional 2-3 masks vs. 10+ masks for eFlash – improved performance and lower cost Thermally Robust No MTJ degradation @ 400°C/150 min Excellent Data Retention Including 260°C solder reflow support for eFlash type application Magnetic Field Immunity Immune to high stray magnetic field In Collaboration with Foundry Partner UMC Mature and high yield integration process Strong Partnership for Embedded Memory MP UMC 22 nm integrated chip 10
- 11. Avalanche Technology at Semicon Taiwan 2020 eSTT-MRAM Macros/Compiler Offer MTJ UMC 22 nm MRAM 300 mm/22nm Ideal for AI applications Features eFlash Slow SRAM (New market applications) LLC CMOS D/R 22nm/1xnm CMOS Capacity 8Mbit-64Mbit Clock Speed 400MHz (maximum) Interface AMBA 3 AHB-Lite Read Speed 25ns 10ns 5ns Write Speed 100+ns 20-50ns 10ns Endurance >108 Cycles >1014 Cycles >1015 Cycles Data Retention >10y@150C >10y@105C days to months to years Reflow Support Yes - - - - 11
- 12. Avalanche Technology at Semicon Taiwan 2020 Write Shmoo at 20ns 0 0 0 0 0 0 0 0 0 0 0 0 0 0 With optimized stack, 20ns write speed can be achieved with large operation window. Write Rl “ 0” Shmoo Write Rh “1” Shmoo 0 0 0 0 0 0 0 0 0 0 0 12
- 13. Avalanche Technology at Semicon Taiwan 2020 Excellent Endurance >1014 Practically Unlimited Endurance >1014 Achieved in 22nm Macros 13
- 14. Avalanche Technology at Semicon Taiwan 2020 10 Years at 125 oC Data Retention 1ppm Data Retention vs 1/T with STT MRAM Macros 14
- 15. Avalanche Technology at Semicon Taiwan 2020 EcoSystem Collaboration to Entering MP STT-MRAM Technology Drivers Foundries & IDMs Upstream vendors: equipment, target and metrology tools Downstream: System designers and application developers 15
- 16. Avalanche Technology at Semicon Taiwan 2020 Summary STT-MRAM is a disruptive embedded memory solution for Al applications High yield MTJ BEOL process well established with Foundry Partners eMRAM is coming to SoC markets with a unique value proposition of persistence, performance, reliability and scalability Low cost and low power eFlash/eSRAM macros are ready for foundry SoC customers at 22/28 nm Acknowledgements: Authors gratefully acknowledge the support of UMC in the MRAM product development and manufacturing. 16