flash’s role in big data, - hot chips · pdf fileflash’s role in big data, past...

OBJECTIVE

ANALYSIS

Flash’s Role in Big Data,

Past Present, and Future Jim Handy

Tutorial: Fast Storage for Big Data

Hot Chips Conference

August 25, 2013

Memorial Auditorium

Stanford University

OBJECTIVE ANALYSIS – www.OBJECTIVE-ANALYSIS.com

OBJECTIVE ANALYSIS Semiconductor Market Research

• Market consulting/research firm

– Market analysis, strategies, white papers

• Highly-respected lead analysts

– Jim Handy: Memories & SSDs

– Lane Mason: Memory chips

– Tom Starnes: Processors

• Industry experience & 25+ years in field

• Reports, Competitive Analysis, Consulting


Haven’t We Seen You Before?

• HotChips 2010: The Inevitable Rise

of NVM in Computing


Agenda

• The problem

• SSDs as a solution

• The role of future memories

• Tomorrow’s computing architecture


Big Data

• Data is exploding. By 2020:

– Annual data production will be 35 ZB (CSC)

– 50 billion “things” on the Internet (Cisco)

• Systems aren’t keeping pace

– CPUs moving along nicely

– HDD/DRAM/Flash capacities still growing

– Interfaces are not moving fast enough

• Result: Break the system into smaller

chunks


The DRAM/HDD Speed Gap

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

Price per Gigabyte

Ba

nd

wid

th (

MB

/s)

From: Solid State Drives in the Enterprise

100

101

102

103

104

105

106

107

10-1 100 101 102 103 104 105 106

Result:

Speed gap

is widening

Semiconductor

memories becoming

faster and cheaper

Disks becoming

cheaper, but no

faster


“Sharding” • Big Data? Subdivide the problem!

Sharded System

Multiple

Servers


Agenda

• The problem





Add Flash, Cut Server Count

Sharded System

Multiple

Servers

Single Server with SSD

SSD SSD


Other Flash Benefits

• Less power & cooling

• Improved reliability

– The fewer things there are, the fewer will fail

• Floor space reduction

• Lower licensing fees

• Faster error recovery

– RAID rebuilds

– Backup restore

– Snapshots


Agenda

• The problem





How to Maintain this Inertia?

$1

$10

$100

$1,000

$10,000

1Q00

1Q01

1Q02

1Q03

1Q04

1Q05

1Q06

1Q07

1Q08

1Q09

1Q10

1Q11

1Q12

Avera

ge P

rice p

er

Gig

ab

yte

DRAM

From: Hybrid Drives: How, Why, & When?


NAND’s Scaling Limit

• NAND will reach a limit

– Too few electrons per gate

– Needs constant shrinks for cost reductions

– 4-bit/cell hard to make

• This may be the maximum possible

• Other technologies will scale past NAND

– PCM, MRAM, RRAM, FRAM….

• Not yet clear which will win


An NVM Timeline

• 2013: 1Xnm planar cell

– Requires Hi-k gate oxide

• 2015: 1Ynm planar cell

– May be the last planar cell

• 2017: 3D NAND in volume

• 2019: 3D – “The Next Generation”

• 2021: Final (?) 3D generation

• 2023: NAND yields to new technology


How Alternatives Will Emerge

0

1

10

100

180nm 65nm 23nm 8nm

Process Node

Rela

tive C

ost

per

Bit Flash

New Tech


New Memories are Better

NAND

• Serial read

• Erase before write

• Block erase/page write

• Slow write

• Inherent bit errors

• Wear

New Memory

• Random read

• Overwrite

• Byte write

• Fast write

• Lower error rates

• Low/no wear

Opens pathway to “Storage Class Memory”


Impact of New Memories


Agenda

• The problem





NVM Won’t Cross HDD $/GB

$0.01

$0.10

$1.00

$10.00

$100.00

$1,000.00

2004 2005 2006 2007 2008 2009 2010 2011 2012

Pri

ce p

er

Gig

ab

yte

HDD Pricing Courtesy of PriceG2

20X the Price!


An Evolving Computing Framework

CPU

DRAM

HDD

CPU

DRAM

HDD

CPU

DRAM

HDD

SSD

NVM

Yesterday Today Tomorrow


NAND Fits in Computers Today

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

Price per Gigabyte

Ba

nd

wid

th (

MB

/s)


100

101

102

103

104

105

106

107

10-1 100 101 102 103 104 105 106


Flash as Memory

Hard Way

• It’s nonvolatile

– Write new code

– Create new topologies

• Worry about

coherency

• Everything is new

Easy Way

• It’s cheaper than

DRAM

– It’s faster than HDD

• Who cares about

volatility?

• Handle coherency like

you do with DRAM

– Use existing code


Flash More Economical than DRAM

From: How PC NAND

Will Undermine DRAM


Enterprise SSD Forecast

0

1

2

3

4

Mil

lio

ns o

f U

nit

s

2009 2010 2011 2012 2013 2014 2015 2016

$3.5 Billion in 2016

59% Unit CAGR

43% Revenue CAGR



Summary

• Flash belongs in all computers today

– So does HDD

• Flash vs. DRAM, not flash vs. HDD

• New NVMs will require new computing

architectures

– “Flash as memory” can be used today

OBJECTIVE

ANALYSIS

Thank You!

Jim Handy

flash’s role in big data, - hot chips · pdf fileflash’s role in big data, past...

Documents