Evaluating the Imagine Stream Processor

1
Evaluating the Imagine Stream Processor

• Jung Ho Ahn, William J. Dally, Brucek Khailany,
  Ujval J. Kapasi,
• and Abhishek Das
• ISCA 2004

2
Motivation

• Provide efficiency of an ASIC
• Provide flexibility of a programmable processor
• Simplify special-purpose processor design
• Lower special-purpose processor design cost
• Provide better applicability
• Target media applications

3
Stream Architecture
4
Development Board
PowerPC, 150 MHz 2 x Imagine, 200 MHz FPGA
Bridge, 66 MHz 256MB of SDRAM / Imagine, 100 MHz
5
Applications
6
Mapping
7
Execution on a Single Stream
Kernel 1
SRF
Iteration 1

Input Stream


Output Stream

Iteration n



8
Execution of Multiple Kernels
Kernel 1
SRF
Stream 1

processing


Stream 2
Kernel 2


Stream 3
processing


Kernel 3
Stream 4


processing

9
Application Performance
GOPS 18
GFLOPS 60
10
Sources of Overhead
11
Stream Length Effects
12
Access Pattern Effects
13
Energy Efficiency

• Energy consumption per FLOP
• (when normalized to 0.13um 1.2V process)
• Imagine _at_ 200 MHz
• 277pJ/FLOP
• TI C67x DSP _at_ 225MHz
• 889pJ/FLOP (3.2x more)
• Intel Pentium M _at_ 1200GHz
• 3600pJ/FLOP (13x more)

14
Memory Bandwidth Requirement
15
Host Processor Bandwidth Requirement
16
Programming Model
17
Compiler OptimizationsStream Ordering
18
Compiler OptimizationsSRF Overlapping and Packing
19
Compiler OptimizationsStrip-mining
20
Compiler OptimizationsLoop Unrolling and
Software Pipelining
21
Conclusions

• Provides performance close to that of ASIC and
  flexibility via programming
• Can sustain between 16 and 60 of the peak
  arithmetic performance
• Exposed 2-level register file allows compiler to
  exploit locality
• Broader applicability
• Requires considerable programming effort
• Limited to media applications with regular
  control-flow

22
Collab Questions

• How does the performance compare to other
  processors? (Dan, Marko, Jason, Prateeksha,
  Chris)
• What is the compiler efficiency? (Mario, Liang)
• How were the design decisions motivated? (Jing,
  Marisabel)
• How does the programming model compare to that of
  GPUs? (Greg)

23
(No Transcript)
24
Kernels