an energy and bandwidth efficient ray tracing architecture

AN ENERGY AND
BANDWIDTH EFFICIENT RAY
TRACING ARCHITECTURE
Daniel Kopta, Konstantin Shkurko, Josef Spjut,
Erik Brunvand, Al Davis
The Goal
2
! 
Can we reduce ray tracing energy consumption
without hurting frame rate?
TRaX – SPMD Ray Tracing GPU
3
Cycle accurate simulator
!  Tiled “Thread Multiprocessors” (TMs)
! 
Instruction Cache
Instruction Cache
Threads
Local Store
Registers
Integer/Branch
L1 Data Cache
Floating Point Units
Where Does Energy Go?
4
Energy/Frame (Joules)
! 
Energy
estimates
from Cacti
and Synopsys
L1
0.22
L2
1.31
DRAM
2.11
0.57
0.94
1.2
~200 Watts at 30 FPS
Instruction Cache
Register File
Compute (XUs)
What can we do?
5
! 
Consider an ASIC
!  Great
! 
Configurable persistent pipelines
!  Almost
! 
energy/delay, but not flexible
as good, but more flexible
Coherent data access
!  Stay
in pipeline mode longer
!  Reduce bandwidth requirements
Macro Instruction Pipelines
6
General Purpose
Reconfigurable Pipeline
Instruction fetch
Instruction fetch
Register File
Register File
Instruction fetch
XU
XU
XU
XU
Register File
Register File
Register File
Ray - Box/Triangle Dataflow
7
TRaX TM
8
TPs
Shared XUs
TPs
Reconfigured Pipeline
9
Special Regs
Box Unit
TPs
Shared FUs
TPs
Pipeline Persistence
10
! 
Opportunities in RT
!  Box
test (traversal)
!  Triangle test (intersection)
!  Others?
! 
Streaming, Ray sorting, etc…
!  StreamRay:
Gribble, Ramani, 2008, 2009
!  Treelet decomposition: Aila & Karras, 2010
Treelets
11
! 
Sized to fit in L1 cache
= box stream
= triangle stream
1
2
3
4
STRaTA
12
! 
Streaming Treelet Ray Tracing Architecture
! 
HW support for treelet ray streams
!  Repurposes
most of L2 cache
!  L1 hit rate increases, off-chip access decreases
! 
Enables pipelines to persist longer
!  Reduce
instruction fetch and register access
Test Scenes
13
Hairball
!  R
Sibenik
Vegetation
Where Does Energy Go?
14
Energy/Frame (Joules)
L1
Treelets
L2
DRAM
Persistent pipelines
0.22
1.31
2.11
Instruction Cache
0.57
Register File
Compute (XUs)
0.94
1.2
Results (ICache + RF)
15
I$ + Register File enery (Joules)
4.7
Baseline
Treelets Only
STRaTA
4.4
3.1
2.95 3.0
2.6
2.9
2.5
2.2
Hairball
Sibenik
Vegetation
Results (L2 + DRAM energy)
16
L2 Dcache + DRAM energy (Joules)
2.18
Baseline
STRaTA
1.64
1.54
1.18
.72
.43
Hairball
Sibenik
Vegetation
Results
17
Energy/Frame (Joules)
Hairball
Sibenik
0.14
1.1
1.5
0.53
Vegetation
1.2
2.6
0.33
0.22
0.85
1.4
0.8
0.29
1.6
0.91
0.13
1.7
0.48
1.1
0.3
1.8
1
L1
L2/Stream
DRAM
Instruction Cache
Register File
Compute (XUs)
Saved
Conclusions
18
! 
Up to 38% combined energy reduction
!  (memory
hierarchy, I$, RF)
! 
No significant change in performance
! 
Simple HW/SW modifications
Thanks!
19
Thanks to:
Samuli Laine for Hairball and Vegetation,
Marko Dabrovic for Sibenik
Anonymous reviewers
Utah Hardware Ray Tracing Group