Origami: Folding Warps for Energy Efficient GPUs Mohammad - - PowerPoint PPT Presentation

SLIDE 1

Origami:
 Folding Warps for Energy Efficient GPUs

Mohammad Abdel-Majeed*, Daniel Wong†, Justin Huang‡ and Murali Annavaram* * University of Southern California

†University of California, Riverside ‡Stanford University

SLIDE 2

Outline

• GPU overview
• Motivation and related work
• Warp Folding
• Origami Scheduler
• Evaluation

2

SLIDE 3

GPGPU Overview (GTX480)


C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C

LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST

SFU SFU SFU SFU

SFU LD/ST INT Unit Operands Result Queue FP Unit

Warp Scheduler (2-level) Register File 128KB Execution Units 64KB shared Memory/L1 cache SM Instruction Cache Fetch and decode

3

19

SLIDE 4

GPGPU Overview (GTX480)


C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C

LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST LD/ST

SFU SFU SFU SFU

SFU LD/ST INT Unit Operands Result Queue FP Unit

Warp Scheduler (2-level) Register File 128KB Execution Units 64KB shared Memory/L1 cache SM Instruction Cache Fetch and decode

3

19

SLIDE 5

DRAM 0.178 RF 0.134 Pipeline 0.114 Constant 0.112 NOC 0.095 Other 0.072 MC 0.048 L2 0.045 EXE 0.201

GPGPU Power Break-Down

4 GPUWattch, ISCA 2013

SLIDE 6

DRAM 0.178 RF 0.134 Pipeline 0.114 Constant 0.112 NOC 0.095 Other 0.072 MC 0.048 L2 0.045 EXE 0.201

GPGPU Power Break-Down

4

EXE 20.1%

GPUWattch, ISCA 2013

SLIDE 7

GPU Scaling Trend

5 GPU Fermi GTX 480 Kepler GTX 680 Maxwell GTX 980 Cores (SMs) 16 8 16 Execution Units 512 1536 2048 RF size 128KB/SM 256KB/SM 256KB/SM #transistors 3 billion 3.5 billion 5.2 billion

SLIDE 8

GPU Scaling Trend

5 GPU Fermi GTX 480 Kepler GTX 680 Maxwell GTX 980 Cores (SMs) 16 8 16 Execution Units 512 1536 2048 RF size 128KB/SM 256KB/SM 256KB/SM #transistors 3 billion 3.5 billion 5.2 billion

SLIDE 9

GPU Scaling Trend

5 GPU Fermi GTX 480 Kepler GTX 680 Maxwell GTX 980 Cores (SMs) 16 8 16 Execution Units 512 1536 2048 RF size 128KB/SM 256KB/SM 256KB/SM #transistors 3 billion 3.5 billion 5.2 billion

SLIDE 10

GPU Scaling Trend

5 GPU Fermi GTX 480 Kepler GTX 680 Maxwell GTX 980 Cores (SMs) 16 8 16 Execution Units 512 1536 2048 RF size 128KB/SM 256KB/SM 256KB/SM #transistors 3 billion 3.5 billion 5.2 billion

SLIDE 11

GPU Scaling Trend

5 GPU Fermi GTX 480 Kepler GTX 680 Maxwell GTX 980 Cores (SMs) 16 8 16 Execution Units 512 1536 2048 RF size 128KB/SM 256KB/SM 256KB/SM #transistors 3 billion 3.5 billion 5.2 billion

SLIDE 12

Technology Scaling

6

• As technology scales leakage power will increase

– Accounts for 50% of the execution units power

• Power Gating can be used to reduce the leakage

power

– Need long idle periods to be effective

6 Warped Gates, MICRO 2012

SLIDE 13

Power Gating Challenges in GPGPUs

• Int. Unit idle period length distribution for hotspot

– Assume 5 idle detect, 14 BET

7

SLIDE 14

Power Gating Challenges in GPGPUs

• Int. Unit idle period length distribution for hotspot

– Assume 5 idle detect, 14 BET

Lost Opportunity

7

SLIDE 15

Power Gating Challenges in GPGPUs

• Int. Unit idle period length distribution for hotspot

– Assume 5 idle detect, 14 BET

Lost Opportunity Energy Loss or Neutral

7

SLIDE 16

Power Gating Challenges in GPGPUs

• Int. Unit idle period length distribution for hotspot

– Assume 5 idle detect, 14 BET

Lost Opportunity Energy Loss or Neutral Energy Savings

7

SLIDE 17

Power Gating Challenges in GPGPUs

• Int. Unit idle period length distribution for hotspot

– Assume 5 idle detect, 14 BET

Lost Opportunity Energy Loss or Neutral Energy Savings

7

SLIDE 18

Power Gating Challenges in GPGPUs

• Int. Unit idle period length distribution for hotspot

– Assume 5 idle detect, 14 BET

Lost Opportunity Energy Loss or Neutral Energy Savings

Need to increase idle period length

7

SLIDE 19

Warp Scheduler Effect on Power Gating

INT FP

FP INT INT FP INTO

Ready Warps

INT

8

8

SLIDE 20

Warp Scheduler Effect on Power Gating

INT FP

Ready Warps

8

Busy C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(FP)

Idle C

(FP)

8

SLIDE 21

• Idle periods


interrupted
 by instructions
 that are greedily
 scheduled

Warp Scheduler Effect on Power Gating

INT FP

Ready Warps

8

Busy C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(FP)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(INT)

C

(FP)

Idle C

(FP)

Need to coalesce warp issues 
 by resource type

8

SLIDE 22

Related Work/Warped-Gates*

• Schedule instructions based on their type
• Force power gated units to stay in power gating

state for at least the breakeven time

9

Frequency 0% 8% 15% 23% 30% Idle period length 6 13 19 25

54.3% 0.0% 45.7%

*Warped-Gates, MICRO 2013

SLIDE 23

Related Work/Warped-Gates*

• Schedule instructions based on their type
• Force power gated units to stay in power gating

state for at least the breakeven time

9

Frequency 0% 8% 15% 23% 30% Idle period length 6 13 19 25

54.3% 0.0% 45.7%

*Warped-Gates, MICRO 2013

SLIDE 24

Fine grain idleness

• Temporal idleness

– Infrequent issues to the same pipeline – Finely interspersed leading to limited power gating

• pportunities

10

SLIDE 25

Fine grain idleness

• Temporal idleness

– Infrequent issues to the same pipeline – Finely interspersed leading to limited power gating

• pportunities

10

Scheduler

INT INT INT INT FP FP FP FP

SLIDE 26

Fine grain idleness

• Temporal idleness

– Infrequent issues to the same pipeline – Finely interspersed leading to limited power gating

• pportunities

10

Scheduler

INT INT INT INT FP FP FP FP

SLIDE 27

Fine grain idleness

• Temporal idleness

– Infrequent issues to the same pipeline – Finely interspersed leading to limited power gating

• pportunities

10

Scheduler

INT INT INT INT FP FP FP FP

SP0

SLIDE 28

Fine grain idleness

• Temporal idleness

– Infrequent issues to the same pipeline – Finely interspersed leading to limited power gating

• pportunities

10

Scheduler

INT INT INT INT FP FP FP FP

SP0 Cycle X 1111 1111

SLIDE 29

Fine grain idleness

• Temporal idleness

– Infrequent issues to the same pipeline – Finely interspersed leading to limited power gating

• pportunities

10

Scheduler

INT INT INT INT FP FP FP FP

SP0 Cycle X 1111 1111 Cycle X+1 Bubble

SLIDE 30

Fine grain idleness

• Temporal idleness

– Infrequent issues to the same pipeline – Finely interspersed leading to limited power gating

• pportunities

10

Scheduler

INT INT INT INT FP FP FP FP

SP0 Cycle X 1111 1111 Cycle X+1 Cycle X+2 Bubble 1111 1111

SLIDE 31

Fine grain idleness

• Temporal idleness

– Infrequent issues to the same pipeline – Finely interspersed leading to limited power gating

• pportunities

10

Scheduler

INT INT INT INT FP FP FP FP

SP0 Cycle X 1111 1111 Cycle X+1 Bubble Cycle X+2 Bubble Cycle X+3 1111 1111

SLIDE 32

Fine grain idleness

• Temporal idleness

– Infrequent issues to the same pipeline – Finely interspersed leading to limited power gating

• pportunities

10

Scheduler

INT INT INT INT FP FP FP FP

SP0 Cycle X 1111 1111 Cycle X+1 Bubble Cycle X+2 Bubble Cycle X+3 1111 1111

SLIDE 33

Fine grain idleness

• Spatial Idleness

– Lanes have different activity

• Branch divergence
• Insufficient parallelism

11

SLIDE 34

Warp Folding

➢Improve the power gating potential by coalescing the pipeline bubbles

12

SLIDE 35

Warp Folding

Scheduler Issued Warps Ready Warps Queue Active Mask: Bubble 1 1 1 1 1 1 1 1 Active Mask: 39

SLIDE 36

Warp Folding

Sub_Warp0: Sub_Warp1: Scheduler Issued Warps Ready Warps Queue Active Mask: Bubble 1 1 1 1 1 1 1 1 Active Mask: 39

SLIDE 37

Warp Folding

Sub_Warp0: Sub_Warp1: Scheduler Issued Warps Ready Warps Queue Active Mask: Bubble 1 1 1 1 1 1 1 1 Active Mask: 39

SLIDE 38

Warp Folding

Sub_Warp0: Sub_Warp1: Scheduler Issued Warps Ready Warps Queue Active Mask: Bubble 1 1 1 1 1 1 1 1 Active Mask: 39

SLIDE 39

Warp Folding

Sub_Warp0: Sub_Warp1: Scheduler Issued Warps Ready Warps Queue 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 39

SLIDE 40

Warp Folding

Sub_Warp0: Sub_Warp1: Scheduler Issued Warps Ready Warps Queue 0 0 0 0 0 0 0 0 Sub_Warp0: Sub_Warp1: 1 1 1 1 1 1 1 1 39

SLIDE 41

Warp Folding

Sub_Warp0: Sub_Warp1: Scheduler Issued Warps Ready Warps Queue 0 0 0 0 0 0 0 0 Sub_Warp0: Sub_Warp1: 1 1 1 1 1 1 1 1 39

SLIDE 42

Warp Folding

Scheduler Issued Warps Ready Warps Queue 0 0 0 0 0 0 0 0 Sub_Warp0: Sub_Warp1: 1 1 1 1 1 1 1 1 39

SLIDE 43

Warp Folding

Scheduler Issued Warps C C C C C C C C Ready Warps Queue 0 0 0 0 0 0 0 0 Sub_Warp0: Sub_Warp1: 1 1 1 1 1 1 1 1 39

SLIDE 44

Warp Folding

Scheduler Issued Warps C C C C C C C C Ready Warps Queue 0 0 0 0 0 0 0 0 Sub_Warp0: Sub_Warp1: 1 1 1 1 1 1 1 1 39

SLIDE 45

14

Folding Granularity

Scheduler Issued Warps C C C C C C C C Active Mask: 1 1 1 1 1 1 1 1 Ready Warps Queue

SLIDE 46

14

Folding Granularity

Sub_Warp0: Sub_Warp1: Scheduler Issued Warps C C C C C C C C Active Mask: 1 1 1 1 1 1 1 1 Ready Warps Queue

SLIDE 47

1 1 1 1 0 0 0 0 0 0 0 0 1 1 1 1 14

Folding Granularity

Sub_Warp0: Sub_Warp1: Scheduler Issued Warps C C C C C C C C Active Mask: 1 1 1 1 1 1 1 1 Ready Warps Queue

SLIDE 48

1 1 1 1 0 0 0 0 0 0 0 0 1 1 1 1 14

Folding Granularity

Sub_Warp0: Sub_Warp1: Scheduler Issued Warps C C C C C C C C Active Mask: 1 1 1 1 1 1 1 1 Ready Warps Queue

SLIDE 49

1 1 1 1 0 0 0 0 0 0 0 0 1 1 1 1 14

Folding Granularity

Sub_Warp0: Sub_Warp1: Scheduler Issued Warps C C C C C C C C Active Mask: 1 1 1 1 1 1 1 1 Ready Warps Queue + Simple

• High wiring overhead
• Delay

SLIDE 50

1 1 1 1 0 0 0 0 0 0 0 0 1 1 1 1 14

Folding Granularity

Sub_Warp0: Sub_Warp1: Scheduler Issued Warps C C C C C C C C Active Mask: 1 1 1 1 1 1 1 1 Ready Warps Queue

SLIDE 51

14

Folding Granularity

Sub_Warp0: Sub_Warp1: Scheduler Issued Warps C C C C C C C C Active Mask: 1 1 1 1 1 1 1 1 Ready Warps Queue

SLIDE 52

14

Folding Granularity

Sub_Warp0: Sub_Warp1: Scheduler Issued Warps C C C C C C C C Active Mask: 1 1 1 1 1 1 1 1 Ready Warps Queue

SLIDE 53

14

Folding Granularity

Sub_Warp0: Sub_Warp1: Scheduler Issued Warps C C C C C C C C Active Mask: 1 1 1 1 1 1 1 1 Ready Warps Queue 1 1 0 0 1 1 0 0 0 0 1 1 0 0 1 1

SLIDE 54

14

Folding Granularity

Sub_Warp0: Sub_Warp1: Scheduler Issued Warps C C C C C C C C Active Mask: 1 1 1 1 1 1 1 1 Ready Warps Queue 1 1 0 0 1 1 0 0 0 0 1 1 0 0 1 1

SLIDE 55

14

Folding Granularity

Sub_Warp0: Sub_Warp1: Scheduler Issued Warps C C C C C C C C Active Mask: 1 1 1 1 1 1 1 1 Ready Warps Queue 1 1 0 0 1 1 0 0 0 0 1 1 0 0 1 1 + Simple + Low wiring overhead + Small delay +Support for lane shuffling

SLIDE 56

14

Folding Granularity

Sub_Warp0: Sub_Warp1: Scheduler Issued Warps C C C C C C C C Active Mask: 1 1 1 1 1 1 1 1 Ready Warps Queue 1 1 0 0 1 1 0 0 0 0 1 1 0 0 1 1 + Simple + Low wiring overhead + Small delay +Support for lane shuffling

SLIDE 57

Warp Folding Pipeline

15 SP Pipe C C C C C C C C 1 1 1 1 1 1 1 1 Result Collector

SLIDE 58

Warp Folding Pipeline

15 SP Pipe C C C C C C C C 1 1 1 1 1 1 1 1 Result Collector

SLIDE 59

Warp Folding Pipeline

15 SP Pipe C C C C C C C C 1 1 1 1 1 1 1 1 Result Collector

Selective Write

Shifting Logic

SP Pipe

Shifting Logic

C C C C C C C C

Re-Shifting Logic Re-Shifting Logic

Result Collector

SLIDE 60

Example

16

SLIDE 61

Example

16

1111 1111

SLIDE 62

Example

16

1111 1111 1100 1100

SLIDE 63

Example

16

1111 1111 1100 1100 0011 0011

SLIDE 64

Example

16

1111 1111 1100 1100 0011 0011

SLIDE 65

Example

16

Shifting Logic

SP Pipe

Shifting Logic 1 1 0 0 1 1 0 0 1 1 0 0 1 1 0 0

C C C C C C C C

Re-Shifting Logic Re-Shifting Logic

Result Collector

1 1 0 0 1 1 0 0

1111 1111 1100 1100 0011 0011

SLIDE 66

Example

16

Shifting Logic

SP Pipe

Shifting Logic 1 1 0 0 1 1 0 0 1 1 0 0 1 1 0 0

C C C C C C C C

Re-Shifting Logic Re-Shifting Logic

Result Collector

1 1 0 0 1 1 0 0

1111 1111 1100 1100 0011 0011

SLIDE 67

Example

16

Shifting Logic

SP Pipe

Shifting Logic 1 1 0 0 1 1 0 0 1 1 0 0 1 1 0 0

C C C C C C C C

Re-Shifting Logic Re-Shifting Logic

Result Collector

1 1 0 0 1 1 0 0 Shifting Logic

SP Pipe

Shifting Logic 0 0 1 1 0 0 1 1 1 1 0 0 1 1 0 0

C C C C C C C C

Re-Shifting Logic Re-Shifting Logic

Result Collector

0 0 1 1 0 0 1 1

1111 1111 1100 1100 0011 0011

SLIDE 68

Example

16

Shifting Logic

SP Pipe

Shifting Logic 1 1 0 0 1 1 0 0 1 1 0 0 1 1 0 0

C C C C C C C C

Re-Shifting Logic Re-Shifting Logic

Result Collector

1 1 0 0 1 1 0 0 Shifting Logic

SP Pipe

Shifting Logic 0 0 1 1 0 0 1 1 1 1 0 0 1 1 0 0

C C C C C C C C

Re-Shifting Logic Re-Shifting Logic

Result Collector

0 0 1 1 0 0 1 1

1111 1111 1100 1100 0011 0011

SLIDE 69

Example

16

Shifting Logic

SP Pipe

Shifting Logic 1 1 0 0 1 1 0 0 1 1 0 0 1 1 0 0

C C C C C C C C

Re-Shifting Logic Re-Shifting Logic

Result Collector

1 1 0 0 1 1 0 0 Shifting Logic

SP Pipe

Shifting Logic 0 0 1 1 0 0 1 1 1 1 0 0 1 1 0 0

C C C C C C C C

Re-Shifting Logic Re-Shifting Logic

Result Collector

0 0 1 1 0 0 1 1

1111 1111 1100 1100 0011 0011

SLIDE 70

Origami scheduler

➢Improve the power gating potential by coalescing warps based on:

➢Threads utilization ➢Instruction type

17

SLIDE 71

Origami scheduler

18

• Group the threads based on their active mask
• One group will have the active mask with less than 32

threads

• The other group will have the active masks with 32

active threads

SLIDE 72

Lane#: 0 1 2 3 4 5 6 7 Cycle x: 1 1 0 1 0 1 0 0 Cycle x+1: 0 1 1 1 0 1 0 0 Cycle x+2: 1 1 1 1 1 1 1 1 Cycle x+3: 0 0 1 1 0 1 0 1 Cycle x+4: 0 1 1 1 0 1 1 0 Cycle x+5: 1 1 1 1 1 1 1 1

Origami scheduler

18

• Group the threads based on their active mask
• One group will have the active mask with less than 32

threads

• The other group will have the active masks with 32

active threads

SLIDE 73

Lane#: 0 1 2 3 4 5 6 7 Cycle x: 1 1 0 1 0 1 0 0 Cycle x+1: 0 1 1 1 0 1 0 0 Cycle x+2: 0 0 1 1 0 1 0 1 Cycle x+3: 0 1 1 1 0 1 1 0 Cycle x+4: 1 1 1 1 1 1 1 1 Cycle x+5: 1 1 1 1 1 1 1 1

Equal to 32 group Less than 32 group

Origami scheduler

18

• Group the threads based on their active mask
• One group will have the active mask with less than 32

threads

• The other group will have the active masks with 32

active threads

SLIDE 74

Lane#: 0 1 2 3 4 5 6 7 Cycle x: 1 1 0 1 0 1 0 0 Cycle x+1: 0 1 1 1 0 1 0 0 Cycle x+2: 0 0 1 1 0 1 0 1 Cycle x+3: 0 1 1 1 0 1 1 0 Cycle x+4: 1 1 1 1 1 1 1 1 Cycle x+5: 1 1 1 1 1 1 1 1

Equal to 32 group Less than 32 group

Origami scheduler

18

• Group the threads based on their active mask
• One group will have the active mask with less than 32

threads

• The other group will have the active masks with 32

active threads

SLIDE 75

Lane#: 0 1 2 3 4 5 6 7 Cycle x: 1 1 0 1 0 1 0 0 Cycle x+1: 0 1 1 1 0 1 0 0 Cycle x+2: 0 0 1 1 0 1 0 1 Cycle x+3: 0 1 1 1 0 1 1 0 Cycle x+4: 1 1 1 1 1 1 1 1 Cycle x+5: 1 1 1 1 1 1 1 1

Equal to 32 group Less than 32 group

Origami scheduler

18

• Group the threads based on their active mask
• One group will have the active mask with less than 32

threads

• The other group will have the active masks with 32

active threads

SLIDE 76

Lane#: 0 1 2 3 4 5 6 7 Cycle x: 1 1 0 1 0 1 0 0 Cycle x+1: 0 1 1 1 0 1 0 0 Cycle x+2: 0 0 1 1 0 1 0 1 Cycle x+3: 0 1 1 1 0 1 1 0 Cycle x+4: 1 1 1 1 1 1 1 1 Cycle x+5: 1 1 1 1 1 1 1 1

Equal to 32 group Less than 32 group

Origami scheduler

18

• Group the threads based on their active mask
• One group will have the active mask with less than 32

threads

• The other group will have the active masks with 32

active threads

Active masks are not aligned!!!

SLIDE 77

Lane Shifting

19

• Shift the threads to the lower order SIMT lanes
• Done at the cluster level to reduce overhead

Lane#: 0 1 2 3 4 5 6 7 Cycle x: 1 1 0 1 0 1 0 0 Cycle x+1: 0 1 1 1 0 1 0 0 Cycle x+2: 0 0 1 1 0 1 0 1 Cycle x+3: 0 1 1 1 0 1 1 0 Cycle x+4: 1 1 1 1 1 1 1 1 Cycle x+5: 1 1 1 1 1 1 1 1

Equal to 32 group Less than 32 group

SLIDE 78

Lane#: 0 1 2 3 4 5 6 7 Cycle x: 1 1 1 0 1 0 0 0 Cycle x+1: 1 1 1 0 1 0 0 0 Cycle x+2: 1 1 0 0 1 1 0 0 Cycle x+3: 1 1 1 0 1 1 0 0 Cycle x+4: 1 1 1 1 1 1 1 1 Cycle x+5: 1 1 1 1 1 1 1 1

Equal to 32 group Less than 32 group

Lane Shifting

19

• Shift the threads to the lower order SIMT lanes
• Done at the cluster level to reduce overhead

SLIDE 79

Origami Scheduling

• Runtime Warp Folding Algorithm

– Folds warps long enough to guarantee savings – Adaptive Folding

• Aggressive folding for warps with lower instruction count
• Conservative folding for warps with higher instruction count
• Change folding frequency based on application utilization
• See paper for more detail!

20

Nphase = Npipelineflush + Nidledetect + Nbreakeventime

SLIDE 80

EVALUATION

21

SLIDE 81

Evaluation Methodology

• GPGPU-Sim v3.0.2

– Nvidia GTX480

• GPUWattch and McPAT 


for energy and area estimation

• Benchmarks from ISPASS, Rodinia and Parboil
• Power gating parameters

– Wakeup delay – 3 cycles – Breakeven time – 14 cycles – Idle detect – 5 cycles

22

22

SLIDE 82

Folding Ratio

23

• Folding frequency is application dependent

SLIDE 83

Folding Ratio

23

• Folding frequency is application dependent

SLIDE 84

Folding Ratio

23

• Folding frequency is application dependent

SLIDE 85

Energy Savings/INT

24

• Eliminates negative energy savings
• Origami scheduler able to amplify folding benefits
• Origami is able to save 49%

SLIDE 86

Energy Savings/INT

24

• Eliminates negative energy savings
• Origami scheduler able to amplify folding benefits
• Origami is able to save 49%

SLIDE 87

Energy Savings/FP

25

• Eliminates negative energy savings
• Origami scheduler able to amplify folding benefits
• Origami is able to save 46%

SLIDE 88

Energy Savings/FP

25

• Eliminates negative energy savings
• Origami scheduler able to amplify folding benefits
• Origami is able to save 46%

SLIDE 89

• Origami is able to reduce the performance 

• verhead significantly over Warped-Gates
• Origami scheduler has positive impact on 


performance for some workloads

Performance

26

SLIDE 90

• Origami is able to reduce the performance 

• verhead significantly over Warped-Gates
• Origami scheduler has positive impact on 


performance for some workloads

Performance

26

SLIDE 91

Conclusion

• Execution units energy efficiency is critical
• Take advantage of the spatial and temporal 


idleness to Improve the power gating potential

• Warp folding

– Adaptively fold warp to coalesce bubbles

• Origmai scheduler

– Scheduler warps based on the threads activity and type.

• Able to save 49% and 46% of the execution units 


leakage energy

• Negligible performance overhead

27

SLIDE 92

Origami:
 Folding Warps for Energy Efficient GPUs

Mohammad Abdel-Majeed*, Daniel Wong†, Justin Huang‡ and Murali Annavaram* abdelmaj@usc.edu, dwong@ece.ucr.edu annavara@usc.edu * University of Southern California

†University of California, Riverside ‡Stanford University

Questions?

SLIDE 93

THANK YOU!

29