Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
1
Dynamic Trends inNonvolatile Memory TechnologiesTom Coughlin, Coughlin Associates, www.tomcoughlin.comJim Handy, Objective Analysis, www.Objective-Analysis.com
2
Outline
• Why Now?
• Changes to the Computing Model
• Emerging Memories 101
• Who’s Producing It?
• Real Life Applications
• Outlook
• Conclusions
• References
© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
33
Why Now?
3
4
Why Emerging Persistent Memories are Necessary • Flash can’t scale with process advances
• NAND flash went 3D at 15nm
• NOR scaling stops with FinFET
• 28nm & smaller processes need something new
• SRAM scaling may stop at 14nm
• In addition, low power high density non-volatile memory is needed for embedded and data center applications
© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
5
NOR Flash Scaling Ends at 28nm
© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
6
SRAM’s In Trouble Too
© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
77
Changes to the Computing ModelEnabling the use of Emerging Memories
7
8
Key Issue: Pricing!
© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
1.E+00
1.E+01
1.E+02
1.E+03
1.E+04
1.E+05
1.E+06
1.E+07
1.E-01 1.E+00 1.E+01 1.E+02 1.E+03 1.E+04 1.E+05 1.E+06
Ban
dwid
th (M
B/s
)
Price per Gigabyte
Tape
HDD
SSD
DRAM
L3
L2
L1
101
100
102
103
104
105
106
107
101 102 103 104 105 10610010-1
3DXPoint
9
Moving from a von Neumann to aMemory-Centric Compute Model
© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
Processor
Memory
Processor
Memory
Processor
Memory
Processor
Memory
Classical von Neumann Model
10
Moving from a von Neumann to aMemory-Centric Compute Model
© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
Processor
MemoryNear Memory
Far Memory
Processor
MemoryNear Memory
Far Memory
Processor
MemoryNear Memory
Far Memory
Processor
MemoryNear Memory
Far Memory
Divide off some of the memory
11
Moving from a von Neumann to aMemory-Centric Compute Model
© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
Processor
Near Memory
Far Memory
Processor
Near Memory
Far Memory
Processor
Near Memory
Far Memory
Processor
Near Memory
Far Memory
Put this memory into a shared pool
12
Memory Interconnect Types
• CXL for “Far Memory”
• Pools heterogeneous memories
• Mixed latencies and data rates
• Gen-Z to connect storage boxes and racks
• The DDR interface will stay with us for “Near Memory”
• DDR good for smaller systems
• HBM fast but restrictive and costly
• OMI for both high speed and large capacities
© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
13
Approaches to Near Memory
© 2021 SNIA Persistent Memory + Computation Storage Summit. All Rights Reserved.
1414
Emerging Memories 101
14
15
Candidates for Persistent Memory
© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
MRAM
ReRAM FRAM
PCM
16
MRAM • Everspin
• Over 120M chips shipped
• Partnership with Global Foundries
• Used in IBM’s FlashCore modules in Storwise and FlashSystem arrays
• Renesas (Formerly IDT)• 8Mb, SPI
• Leading foundries starting to ship • GlobalFoundries
• TSMC
• Samsung
• Others
© 2021 SNIA Persistent Memory + Computation Storage Summit. All Rights Reserved.
Source: TechInsights
17
PCM (3D XPoint)
• The second-oldest emerging memory (1970)
• Intel Optane products
• NVMe shipped in 2017
• DIMMs in 2018
• Micron Abandoning 3D XPoint
• Selling Lehi fab
© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
18
ReRAM • Adesto has shipped CBRAM chips for several years
• Dialog Semiconductor acquired Adesto June 2020
• Will license CBRAM technology to GLOBALFOUNDRIES
• GF will first offer as an embedded, option on its 22FDX platform
• GF Plans to extend to other platforms.
• Cerfe Labs (Arm spin-out)• Correlated electron materials (CeRAM)
• Licensed from Symetrix.
• Others (Mitsubishi, Fujitsu, Panasonic, Winbond, Honeywell,…)
• Foundry support (GLOBALFOUNDRIES, TSMC, others)© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
19
FRAM
• The oldest emerging memory (1955)
• The highest-shipped emerging memory
• Finding new life with new materials
© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
2020
Who’s Producing It?
20
21
Major Chip Foundries OfferMRAM, FRAM, and ReRAM Options • TSMC, Samsung, GLOBALFOUNDRIES, UMC,
TI, …
• STT-MRAM, FRAM, & ReRAM today
• SOT-MRAM and other MRAM technologies later on
• Could compete against lower level cache (faster) SRAM
• Possible DRAM alternative for higher performance at lower power
© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
22
TSMC MRAM Roadmap
© 2021 SNIA Persistent Memory + Computation Storage Summit. All Rights Reserved.
2323
Real-Life ApplicationsIt’s Here Now!
23
24
MRAM IoT SoC – Ambiq Apollo
• SoC for intelligent endpoint IoT devices
• Ultra-low battery power
• Serves as both an application processor and a coprocessor
© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
25
MRAM DNN Accelerator Chip – NuMem
• Used by NASA
• 1-32 processing engines, 32-1024 ALUs per chip• Efficient for matrix multiplication, convolution, etc.
• Radiation hard, high endurance MRAM• Nonvolatility reduces energy requirements
• Numerous Applications:• Sensor fusion for super resolution
• Terrain Mapping for Depth and Terrain Classification
• Navigation systems: Object Detection & Tracking
© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
26
MRAM GPS Receiver – Sony
• CXD5605 GPS Receiver
• Used in Huawei GT 2 Smartwatch
• 8Mb Embedded MRAM
• Samsung 28 FD-SOI Process
© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
Source: TechInsights
27
Other Emerging Memory Products
• NXP MRAM MCU
• STMicro PCM MCU
• TI FRAM MCU
• Fujitsu FRAM MCUs
© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
2828
Outlook
28
29
Growth in New Memory Shipments
© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
MRAM
3DXPoint
0.01
0.10
1.00
10.00
100.00
1,000.00
10,000.00
100,000.00
1,000,000.00
10,000,000.00
2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030
Peta
byte
Shi
pmen
ts
MRAM DRAM 3D XPoint NAND
30
Summary
• NOR flash and SRAM have stopped scaling
• New non-volatile memory types will fill the void
• This will lead to new memory-centric computer architectures
• The storage/memory hierarchy will change
• There are four leading candidates: MRAM, PCM, ReRAM, and FRAM
• Leading foundries already support these new memories
• New memories are in use today
• MRAM and PCM revenues should exceed $36B by 2030
© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
31
New Report:Emerging Memories Find Their Direction
© 2021 SNIA Persistent Memory + Computational Storage Summit. All Rights Reserved.
http://www.tomcoughlin.com/techpapers.htmhttps://Objective-Analysis.com/reports/#Emerging
32
References
• Emerging Memories Find Their Direction, Coughlin Associates and Objective Analysis, https://tomcoughlin.com/tech-papers/
• The Future of Low-Latency Memory, White Paper, April 2021
• Computer Express Link 2.0 Specification: Memory Pooling, CXL BrightTalk, March 23, 2021
© 2021 SNIA Persistent Memory + Computation Storage Summit. All Rights Reserved.