Tom Coughlin Coughlin Associates Ed Grochowski, Computer

Transcription

Tom Coughlin Coughlin Associates Ed Grochowski, Computer
Advances in Non-­‐
Vola0le Storage Technologies Tom Coughlin Coughlin Associates Ed Grochowski, Computer Memory/
Storage Consultant
© 2015 Coughlin Associates 1 Outline •  The Shape of Things to Come •  In Search of New Memories •  Intel/Micron’s 3D XPoint Technology •  The Memory is the Computer •  Conclusions © 2015 Coughlin Associates 2 The Shape of Things to Come © 2015 Coughlin Associates 3 IOPS Required Respondents
40%
30%
79% between 10K
and 1M IOPS
20%
10%
0%
10
100
1K
10K
100K
1M
10M
IOPS
From the 2014 How Many IOPS Do You Really Need Report,
Coughlin and Handy, http://www.tomcoughlin.com/techpapers.htm
© 2015 Coughlin Associates 4 Maximum Latency Requirement Respondents
40%
30%
36% at 10ms
latency
20%
10%
0%
<10ns
100ns
10µs
1ms
100ms
>1 sec
Latency
From the 2014 How Many IOPS Do You Really Need Report,
Coughlin and Handy, http://www.tomcoughlin.com/techpapers.htm
© 2015 Coughlin Associates 5 Latencies Separate Computer Memory from Storage Non-­‐
Uniform Memory Access Doug Voight, HP, 2015 FMS © 2015 Coughlin Associates 6 Touch rate versus response Ume indicaUng various types of uses © 2015 Coughlin Associates 7 Digital storage technologies regions overlaid on the Touch Rate/Response Time chart Memory © 2015 Coughlin Associates 8 Ent. SSDs, perf. HDDs and capacity HDDs © 2015 Coughlin Associates 9 Storage Latency with Current and Future NV Solid State NV Technologies (from NVMP talk at SNIA Winter Symposium, 2014) SoOware !
© 2015 Coughlin Associates 10 ImplicaUons of Persistent Memory •  NV memory retains its data even when power off—thus instant recovery of state before power down is now possible •  NV memory with the right SW changes can provide much be_er latencies than current systems and SW •  Non-­‐volaUle memory will save power since refreshes not needed •  Persistent memory creates new opportuniUes to share that memory between different computers or computer chips using Remote Direct Memory Access (RDMA) •  Embedded NVM technology can lead to “logic-­‐in-­‐memory architecture” for future SoC—this can lead to new distributed computer architectures © 2015 Coughlin Associates 11 Challenges of Persistent Memory •  If the memory isn’t cleared by rebooUng a system, then reboots to recover funcUonality with corrupted data won’t work •  There will need to be a special reboot and clear funcUon that erases and recovers data in memory •  Or, and this may be an even be_er soluUon, we need to built devices and soeware that are self-­‐
monitoring and self correcUng so we don’t need to reboot in order to retain funcUonality © 2015 Coughlin Associates 12 In Search of New Memories © 2015 Coughlin Associates 13 Memory and Storage Today 2015 Emerging NVM and Their Manufacture Report, Coughlin Associates © 2015 Coughlin Associates 14 !
Hard Disk Drives “Boilerplate”
© 2015 Coughlin Associates 15 HDD Areal Density Growth
Recent Developments: SMR, HAMR, He-­‐filled HDDs, MAMR with 3D Recording, © 2015 Coughlin Associates 16 Actual Quarterly Announced HDD Product AD Charts 1000
100 % CAGR
80% CAGR
60% CAGR
Gbits per square inch
40% CAGR
20% CAGR
HDD Product
100
© Coughlin Associates, 2015 10
Q1 2000 Q1 2001 Q1 2002 Q1 2003 Q1 2004 Q1 2005 Q1 2006 Q1 2007 Q1 2008 Q1 2009 Q1 2010 Q1 2011 Q1 2012 Q1 2013 Q1 2014 Q1 2015
© 2015 Coughlin Associates 17 Flash Memory !
!
© 2015 Coughlin Associates 18 NAND Technology Roadmap
J. Choe FMS 2015
© 2015 Coughlin Associates 19 Annual HDD/SSD Storage Shipments Annual Shipped Storage Capacity (Exabytes) 1,200.00 1,000.00 HDD SSD 800.00 600.00 400.00 ~10 X
200.00 0.00 2012 2013 2014 2015 © 2015 Coughlin Associates 2016 2017 2018 20 Comparisons of Flash and HDDs Mike Workman Keynote, 2015 FMS © 2015 Coughlin Associates 21 Emerging Memory Technology •  NVM will save power •  Persistent memory enables memory sharing (RDMA) •  Embedded NVM technology can lead to “logic-­‐in-­‐
memory architecture” © 2015 Coughlin Associates The emerging NVM market could exceed $2 B by 2019 2014 Emerging NVM Report and Their
Manufacture, Coughlin Associates
22 FeRAM !
FeRAM products shipping for many years for niche applicaUons, hard to scale density… © 2015 Coughlin Associates 23 Phase Change Memory (PCM) !
!
Several industry experts think that Micron/Intel’s 3D XPoint Technology is a crossbar phase change memory © 2015 Coughlin Associates 24 TYPES OF RRAM TECHNOLOGY
Sources/ Crossbar, Tezzaron Semi.
Oxide Conduction
Via Oxygen Vacancies
Metal Filament
Formation (CBRAM)
© 2015 Coughlin Associates 25 STT MRAM Replacing DRAM and SRAM Over 50 M MRAM chips shipped by Everspin !
© 2015 Coughlin Associates 26 NVM Progress (Amigo Tsutsui, Sony at the 2015 FMS)
© 2015 Coughlin Associates 27 Memory Selector Device OpUons An Chen, 2014 AVS TFUG Seminar © 2015 Coughlin Associates 28 © 2015 Coughlin Associates 29 Annual Memory Capacity Shipments (2014 Emerging NVM Report and Their Manufacture, Coughlin Associates) © 2015 Coughlin Associates 30 3D XPoint
Technology
© 2015 Coughlin Associates 31 Comparison of Memory Technologies Slide from 2015 Storage Visions Conference NV-­‐
Memory Normalized
Log Scale
Capacity 1MB = 1unit 100
101
SRAM NAND Power/GB 10mW = 1unit 102
DISK DISK DRAM 103
104
DRAM NAND 105
106
107
108
DISK SRAM Next-­‐Gen Storage NAND DRAM SRAM Cost/GB 10c = 1unit SRAM Latency 1ns = 1unit DRAM NAND 1000X Next-­‐Gen NV-­‐
Mem DISK 32 32
Next Gen Memory must fill the DRAM & Flash latency gap
© 2015 Coughlin Associates 3D XPoint Memory © 2015 Coughlin Associates 33 3D XPoint Comparisons to NAND R
Latency (ls) 10,000 Source: Storage Technology Group, Intel 200 175 SSD NAND technology offers ~100X reduc0on In latency versus HDD 150 125 100 75 50 25 0 I NVMeTM eliminates 20 μs of latency today 3D XPointTM technology reduces NVM latency offering ~10x reduc0on in latency vs NAND SSD HDD +SAS/
SATA Drive Latency ControllerLatency Soeware Latency (i.e. SAS HBA) SSD 3D Xpoint +NVMeTM Technology claims are based on comparisons of latency, density and write cycling metrics amongst memory technologies recorded on published specificaUons of in m© arket mCemory products against internal Intel specificaUons 2015 oughlin Associates 34 SSD NAND +SAS/SATA SSD NAND +NVMeTM 3D Xpoint Memory © 2015 Coughlin Associates 35 The Memory is the Computer © 2015 Coughlin Associates 36 Spintronics • 
From Tohoku University in Japan © 2015 Coughlin Associates 37 Spin-­‐Based CompuUng mLogic: All Spin Logic Device and Circuits for Future Electronics, Jimmy Zhu, et. Al, March 2015 © 2015 Coughlin Associates 38 mCell Spin Logic Unit (J. Zhu, Carnegie Mellon Univ.) © 2015 Coughlin Associates 39 STT Logic ApplicaUon (Reducing Dark Silicon) © 2015 Coughlin Associates 40 What if Spin Transfer Replaced Charge Transfer (Current)? •  Spin transfer would not generate the heat that electrical currents generate •  This allow building powerful devices in small spaces •  Spin could be used for both processing and memory/storage © 2015 Coughlin Associates 41 Conclusions •  Faster memory/storage is needed for modern applicaUons •  The storage/memory landscape has more opUons than ever—e.g. 3D XPoint •  Non-­‐volaUle memories will replace volaUle memories •  NVM and processors want to come together •  Spin-­‐based electronics are one example of a technology that could put memory and processing in one device © 2015 Coughlin Associates 42 References •  2014 How Many IOPS Do You Really Need Report, Coughlin and Handy, Coughlin Associates •  2014 and 2015 Emerging NVM Report and Their Manufacture, Coughlin Associates •  2015 Storage Visions Conference PresentaUons •  Touch Rate: A metric for analyzing storage system performance, Steven Heltzer and Tom Coughlin, 2015 •  InformaUon on all these documents are available at:h_p://www.tomcoughlin.com/techpapers.htm © 2015 Coughlin Associates 43 Thanks
© 2015 Coughlin Associates 44