Approximate Storage in Solid-State Memories Adrian Sampson - - PowerPoint PPT Presentation

▶
approximate storage in solid state memories
SMART_READER_LITE
LIVE PREVIEW

Approximate Storage in Solid-State Memories Adrian Sampson - - PowerPoint PPT Presentation

Jan 26, 2023 288 likes •710 views

Approximate Storage in Solid-State Memories Adrian Sampson University of Washington Jacob Nelson Karin Strauss Microsoft Research & UW University of Washington Luis Ceze sa pa MICRO 2013 Compiler Runtime Vector GPU CPU Processor

slide-1
SLIDE 1

sa pa

MICRO 2013 Adrian Sampson Jacob Nelson Karin Strauss Luis Ceze

Approximate Storage in Solid-State Memories

University of Washington Microsoft Research & UW University of Washington

slide-2
SLIDE 2

Vector Processor

GPU CPU

Compiler Runtime

Accelerator

slide-3
SLIDE 3
slide-4
SLIDE 4

Compute

Display Disk Network

I/O Storage

Memory

slide-5
SLIDE 5

Compute

Disk

I/O Storage

Memory

slide-6
SLIDE 6
slide-7
SLIDE 7
slide-8
SLIDE 8
slide-9
SLIDE 9

0% 20% 40% 60% 80% 100% 1.6 1.8 2 2.2 2.4 2.6 2.8 3

  • utput quality loss

average write steps jmeint

% 2 % 4 % 6 % 8 % 1 % 2 . 2 2 . 4 2 . 6 2 . 8 3 3 . 2 3 . 4

  • u

t p u t q u a l i t y l

  • s

s w r i t e s

× 1

7

y t r a c e r z x i n g f f t j m e i n t m c s m m s

  • r

Main-memory applications using failed blocks.

s e n s

  • r

l

  • g

b y f h a s

I m p a c t

correcting per

slide-10
SLIDE 10

Themes in approximate computing

approx precise

Interleaving: Programs are both approximate & precise

LO HI

x ± y

Error mitigation: Exploit the hardware to minimize error

slide-11
SLIDE 11

Phase-change memory (PCM)

Surpass DRAM’s scaling limits

+

Non-volatile Faster than flash “Almost” as fast as DRAM

:)

slide-12
SLIDE 12

Phase-change memory (PCM)

+

:(

Write speed & energy Cells wear out

  • ver time
slide-13
SLIDE 13

Phase-change memory (PCM) :(

Write speed & energy Cells wear out

  • ver time

Multi-level cells are denser
 but need more time and energy. Cells wear out over time and can no longer be used.

slide-14
SLIDE 14

Phase-change memory (PCM) :

Multi-level cells are denser
 but need more time and energy
 to protect against errors. Cells wear out over time and can no longer be used for precise data storage.

(

slide-15
SLIDE 15

Phase-change memory (PCM) :

(

Fast Dense

slide-16
SLIDE 16

Phase-change memory (PCM) :

(

Fast Dense Accurate

slide-17
SLIDE 17

Approximate storage in PCM

Trade off accuracy for performance in multi-level cell accesses. Use worn-out memory for approximate data instead of throwing it away.

slide-18
SLIDE 18

Approximate storage in PCM

Trade off accuracy for performance in multi-level cell accesses. Use worn-out memory for approximate data instead of throwing it away.

1 2

slide-19
SLIDE 19

Approximate storage in PCM

Trade off accuracy for performance in multi-level cell accesses. Use approximate throwing it away.

1 2

slide-20
SLIDE 20

Single-level cells

high low analog value digital value 1

slide-21
SLIDE 21

Multi-level cells

high low analog value digital value 00 11 01 10

slide-22
SLIDE 22

Writing to multi-level cells

high low analog value digital value 00 11 01 10

probability

slide-23
SLIDE 23

Writing to multi-level cells,
 approximately

high low analog value digital value 00 11 01 10

probability

slide-24
SLIDE 24

Speed Density Accuracy

slide-25
SLIDE 25

10

Iterative writes

high low 00 11 01 time

target range

slide-26
SLIDE 26

10

Iterative writes, approximately

high low 00 11 01 time

target range

slide-27
SLIDE 27

10

Iterative writes, approximately

high low 00 11 01 time

target range

slide-28
SLIDE 28

wider target range fewer iterations to converge faster writes (or better density at the same speed)

slide-29
SLIDE 29

Encoding to minimize error in approximate MLC

1 cell, 4 bits unreliable reliable

LO HI

x ± y

slide-30
SLIDE 30

lots of errors 4 cells, 16 bits

Encoding to minimize error in approximate MLC

LO HI

x ± y

slide-31
SLIDE 31

lots of errors 4 cells, 16 bits

Encoding to minimize error in approximate MLC

LO HI

x ± y

slide-32
SLIDE 32

Write speedup for approximate MLC

Writes are 1.7× faster on average with quality loss under 10%

0.5 1 1.5 2 2.5

mc smm sor fft lu zxing jmeint raytracer pa nn ml image mean

main-memory benchmarks persistent data

best write speedup

slide-33
SLIDE 33

Approximate storage in PCM

Trade off performance in accesses. Use worn-out memory for approximate data instead of throwing it away.

slide-34
SLIDE 34

Failed cells are a fact of life

0 1 1 0 1 0 0 1 1 1 0 1 0 0 1 0 1 1 1 0 0 0 1 1 0 0 1 0 1 1 0 0 1 0 0 0 1

a good block

slide-35
SLIDE 35

Failed cells are a fact of life

0 1 1 0 1 0 0 1 1 1 0 1 0 0 1 0 1 1 1 0 0 0 1 1 0 0 1 0 1 1 0 0 1 0 0 0 1

a (tragically) failed block

slide-36
SLIDE 36

Traditional error correction

0 1 1 0 1 0 0 1 1 1 0 1 0 0 1 0 1 1 1 0 0 0 1 1 0 0 1 0 1 1 0 0 1 0 0 0 1

corrected data block correction bits

slide-37
SLIDE 37

Correction resources are exhaustible

0 1 1 0 1 0 0 1 1 1 0 1 0 0 1 0 1 1 1 0 0 0 1 1 0 0 1 0 1 1 0 0 1 0 0 0 1

uncorrectable (bad) block correction bits a p p r

  • x

i m a t e

slide-38
SLIDE 38

Prioritized error correction

0 1 1 0 1 0 0 1 1 1 0 1 0 0 1 0 1 1 1 0 0 0 1 1 0 0 1 0 1 1 0 0 1 0 0 0 1

uncorrectable (bad) block correction bits a p p r

  • x

i m a t e

LO HI

x ± y

error exposed where it does the least harm

slide-39
SLIDE 39

Lifetime extension with block recycling

Lifetime extended by 23% on average

  • r from about 5.2 to 6.5 years

0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

mc smm sor fft lu zxing jmeint raytracer pa nn ml image mean

main-memory benchmarks persistent data

normalized lifetime (writes)

slide-40
SLIDE 40

Compute

Display Disk Network

I/O Storage

Memory

slide-41
SLIDE 41

Compute

Display Disk Network

I/O Storage

Memory