Persistent Memory Productization driven by AI & ML
Danny Sabour
VP Marketing, Avalanche Technology
© 2018 SNIA Persistent Memory Summit. All Rights Reserved.
Persistent Memory Usage from Cloud to Node
CLOUD
Big data processing
Data Warehousing
EDGE
Data Visualization
Data caching
Data filtering
IoT NODE
Realtime
Data processing
2
ERP
MES
SCADA
Systems
Training
Inference
Deep
Learning
Compute
Storage
© 2018 SNIA Persistent Memory Summit. All Rights Reserved.
Node Architecture : From Data Fusion to AI
3
NV Store
Sense
V2X
Radar
Ladar
UltraS
Camera
Motion
Speed
CoProc
CoProc
CoProc
CoProc
CoProc
SensorFusion
Think
Processor
Processor
Transmit
Data Fusion at the Node
Sense
V2X
Radar
Ladar
UltraS
Camera
Motion
Speed
CoProc
CoProc
CoProc
CoProc
CoProc
SensorFusion
Think
Processor
Processor
Transmit
Inference: Local Decision Making
NV Store
ACT
+
Neural Processor
• The Node:
• Collects Data
• Aggregates data
• Sends data back to Edge
• Needs
• Persistence
• Low Power
• Smallest die size
• The Node:
• Collects Data
• Aggregates data
• Sends data back to Edge
• Make decisions per input
• Learn based on local stimuli
• Needs
• Persistence
• Low Power
• Smallest die size
© 2018 SNIA Persistent Memory Summit. All Rights Reserved.
Gen 1 SPI 8Mb
Technology
4
IoT Roadmap
Gen 2 QSPI 32Mb
Gen 3 OSPI 256Mb
Discrete
IoT Memory
Embedded
IoT Memory
28nm NVM MACRO
8-64Mb
28nm SRAM MACRO
8-64Mb
14nm SRAM MACRO
8-64Mb
28nm NVM MACRO
64-256Mb
`
© 2018 SNIA Persistent Memory Summit. All Rights Reserved.
Edge/Cloud Architecture
Processor
DRAM
• The Cloud and the Edge Architecture
are similar, but with the possible
variations in:
• GPU Cores which can be used for
Machine learning
• PCIe Cache for performance
• Communication to other domains in
the edge for inference and learning
• Needs
• Persistence
• Speed
PCH
5
CPU Core 1
CPU Core 2
CPU Core 3
Core X
DRAM
DRAM
DRAM
Storage
Processing
PCIe Cache
NVDIMM-N
AI/Machine
Learning GPU
Storage Cache
Training : Learning New Capabilities
© 2018 SNIA Persistent Memory Summit. All Rights Reserved.
Speeding up the Cloud/Edge Computing: Cache & GPU Memory
6
Medium Density Persistent RAM (MD-RAM)
Processor
DRAM
PCH
CPU Core 1
CPU Core 2
CPU Core 3
Core X
DRAM
DRAM
DRAM
Storage
Processing
PCIe Cache
AI/Machine
Learning GPU
Storage Cache
SSD
MD - AVRAM
MD - AVRAM
MD - AVRAM
MD - AVRAM
MD - AVRAM
MD - AVRAM
MD - AVRAM
MD - AVRAM
MD - AVRAM
MD - AVRAM
Controller
NAND FLASH
NAND FLASH
NAND FLASH
NAND FLASH
NAND FLASH
NAND FLASH
NAND FLASH
NAND FLASH
NAND FLASH
BusInterface
BusInterface
© 2018 SNIA Persistent Memory Summit. All Rights Reserved.
Universal Memory in the Cloud: NVDIMM
Processor
DRAM
PCH
CPU Core 1
CPU Core 2
CPU Core 3
Core X
DRAM
DRAM
DRAM
Storage
Processing
PCIe Cache
AI/Machine
Learning GPU
Storage Cache
7
NVDIMM-N: Brought Persistence to DRAM
Controller + Battery + Flash
NVDIMM-N
NVDIMM-P
NVDIMM-P: Hides Persistent Memory Endurance
Controller
NVDIMM-P: Hides Persistent Memory Endurance
Controller + Battery + Cache
Cache
DRAM
High Density Persistent RAM (HD-RAM)
AVA-DIMM: Persistent DRAM replacement
© 2018 SNIA Persistent Memory Summit. All Rights Reserved.
Gen 1 MD-AVRAM
Technology
8
High and Mid Density Persistent RAM
Mid-Density
Persistent
RAM
Regular
Small
Tiny Micro
1S1R1T1R
High-Density
Persistent
RAM
Gen 2 MD-AVRAM
DDR 1-4Gb
DDR 4-8Gb
DDR 4-8Gb
DDR 4-16Gb
Gen 3 HD-AVRAM
DDR 8-32Gb
DDR 64Gb
© 2018 SNIA Persistent Memory Summit. All Rights Reserved. 9
Where does MRAM fit into the picture
10-9 10-8 10-7 10-6 10-5
10-4 10-3
1
103
106
109
1012
1015
1018
Write Speed
Endurance(Cycles)
NAND
SRAM
RRAM
PCM
NOR
DRAM
HDD
1 Magnetic structure1 Magnetic structure
Embedded:
• High Performance
• High Temperature
E-NVME-SRAM
1 Magnetic structure
Embedded:
• High Performance
• High Temperature
Stand Alone:
• IoT
SPnvSRAM
MD-RAM
HD-RAM
1 Magnetic structure
Embedded:
• High Performance
• High Temperature
Stand Alone:
• IoT
• Med Density RAM
• High Density RAM
Avalanche Specification for different MRAMs available at: www.avalanche-technology.com/products
© 2018 SNIA Persistent Memory Summit. All Rights Reserved.
Summary
10
• MRAM products are making inroads from the cloud to the IoT Node
• It is not one product
• MRAM is unifying the Software defined world
• Compute
• Storage
• Network
• Built on a Unified CPU & Memory platform based on a unified MTJ architecture for
• Compute, Storage, Network
• Across the Cloud, Edge, Node
• Phased product introduction has begun

Persistent Memory Productization driven by AI & ML

  • 1.
    Persistent Memory Productizationdriven by AI & ML Danny Sabour VP Marketing, Avalanche Technology
  • 2.
    © 2018 SNIAPersistent Memory Summit. All Rights Reserved. Persistent Memory Usage from Cloud to Node CLOUD Big data processing Data Warehousing EDGE Data Visualization Data caching Data filtering IoT NODE Realtime Data processing 2 ERP MES SCADA Systems Training Inference Deep Learning Compute Storage
  • 3.
    © 2018 SNIAPersistent Memory Summit. All Rights Reserved. Node Architecture : From Data Fusion to AI 3 NV Store Sense V2X Radar Ladar UltraS Camera Motion Speed CoProc CoProc CoProc CoProc CoProc SensorFusion Think Processor Processor Transmit Data Fusion at the Node Sense V2X Radar Ladar UltraS Camera Motion Speed CoProc CoProc CoProc CoProc CoProc SensorFusion Think Processor Processor Transmit Inference: Local Decision Making NV Store ACT + Neural Processor • The Node: • Collects Data • Aggregates data • Sends data back to Edge • Needs • Persistence • Low Power • Smallest die size • The Node: • Collects Data • Aggregates data • Sends data back to Edge • Make decisions per input • Learn based on local stimuli • Needs • Persistence • Low Power • Smallest die size
  • 4.
    © 2018 SNIAPersistent Memory Summit. All Rights Reserved. Gen 1 SPI 8Mb Technology 4 IoT Roadmap Gen 2 QSPI 32Mb Gen 3 OSPI 256Mb Discrete IoT Memory Embedded IoT Memory 28nm NVM MACRO 8-64Mb 28nm SRAM MACRO 8-64Mb 14nm SRAM MACRO 8-64Mb 28nm NVM MACRO 64-256Mb `
  • 5.
    © 2018 SNIAPersistent Memory Summit. All Rights Reserved. Edge/Cloud Architecture Processor DRAM • The Cloud and the Edge Architecture are similar, but with the possible variations in: • GPU Cores which can be used for Machine learning • PCIe Cache for performance • Communication to other domains in the edge for inference and learning • Needs • Persistence • Speed PCH 5 CPU Core 1 CPU Core 2 CPU Core 3 Core X DRAM DRAM DRAM Storage Processing PCIe Cache NVDIMM-N AI/Machine Learning GPU Storage Cache Training : Learning New Capabilities
  • 6.
    © 2018 SNIAPersistent Memory Summit. All Rights Reserved. Speeding up the Cloud/Edge Computing: Cache & GPU Memory 6 Medium Density Persistent RAM (MD-RAM) Processor DRAM PCH CPU Core 1 CPU Core 2 CPU Core 3 Core X DRAM DRAM DRAM Storage Processing PCIe Cache AI/Machine Learning GPU Storage Cache SSD MD - AVRAM MD - AVRAM MD - AVRAM MD - AVRAM MD - AVRAM MD - AVRAM MD - AVRAM MD - AVRAM MD - AVRAM MD - AVRAM Controller NAND FLASH NAND FLASH NAND FLASH NAND FLASH NAND FLASH NAND FLASH NAND FLASH NAND FLASH NAND FLASH BusInterface BusInterface
  • 7.
    © 2018 SNIAPersistent Memory Summit. All Rights Reserved. Universal Memory in the Cloud: NVDIMM Processor DRAM PCH CPU Core 1 CPU Core 2 CPU Core 3 Core X DRAM DRAM DRAM Storage Processing PCIe Cache AI/Machine Learning GPU Storage Cache 7 NVDIMM-N: Brought Persistence to DRAM Controller + Battery + Flash NVDIMM-N NVDIMM-P NVDIMM-P: Hides Persistent Memory Endurance Controller NVDIMM-P: Hides Persistent Memory Endurance Controller + Battery + Cache Cache DRAM High Density Persistent RAM (HD-RAM) AVA-DIMM: Persistent DRAM replacement
  • 8.
    © 2018 SNIAPersistent Memory Summit. All Rights Reserved. Gen 1 MD-AVRAM Technology 8 High and Mid Density Persistent RAM Mid-Density Persistent RAM Regular Small Tiny Micro 1S1R1T1R High-Density Persistent RAM Gen 2 MD-AVRAM DDR 1-4Gb DDR 4-8Gb DDR 4-8Gb DDR 4-16Gb Gen 3 HD-AVRAM DDR 8-32Gb DDR 64Gb
  • 9.
    © 2018 SNIAPersistent Memory Summit. All Rights Reserved. 9 Where does MRAM fit into the picture 10-9 10-8 10-7 10-6 10-5 10-4 10-3 1 103 106 109 1012 1015 1018 Write Speed Endurance(Cycles) NAND SRAM RRAM PCM NOR DRAM HDD 1 Magnetic structure1 Magnetic structure Embedded: • High Performance • High Temperature E-NVME-SRAM 1 Magnetic structure Embedded: • High Performance • High Temperature Stand Alone: • IoT SPnvSRAM MD-RAM HD-RAM 1 Magnetic structure Embedded: • High Performance • High Temperature Stand Alone: • IoT • Med Density RAM • High Density RAM Avalanche Specification for different MRAMs available at: www.avalanche-technology.com/products
  • 10.
    © 2018 SNIAPersistent Memory Summit. All Rights Reserved. Summary 10 • MRAM products are making inroads from the cloud to the IoT Node • It is not one product • MRAM is unifying the Software defined world • Compute • Storage • Network • Built on a Unified CPU & Memory platform based on a unified MTJ architecture for • Compute, Storage, Network • Across the Cloud, Edge, Node • Phased product introduction has begun