43
Department of Electrical Engineering and Computer Science RISC vs. CISC Chetan Patil http://chetanpatil.info/ Few Slides Adapted From : Power Struggles: Revisiting the RISC vs. CISC Debate on Contemporary ARM and x86 Architectures (HPCA 2013)

RISC vs CISC

Embed Size (px)

DESCRIPTION

Learn how an architecture being a RISC or CISC affects and whether it's relevant or not.

Citation preview

Page 1: RISC vs CISC

Department of Electrical Engineering and Computer Science

RISC vs. CISC

Chetan Patilhttp://chetanpatil.info/

Few Slides Adapted From : Power Struggles: Revisiting the RISC vs. CISC Debate on Contemporary

ARM and x86 Architectures (HPCA 2013)

Page 2: RISC vs CISC

Department of Electrical Engineering and Computer Science

PAPERS

• D. Bhandarkar and D. W. Clark. Performance from architecture: comparing a RISC and a CISC with similar hardware organization. In ASPLOS '91. [LINK]

• Power Struggles: Revisiting the RISC vs. CISC Debate on Contemporary ARM and x86 Architectures. In HPCA '13. [LINK]

Page 3: RISC vs CISC

Department of Electrical Engineering and Computer Science

OVERVIEW

• What is RISC & CISC?

• Methods

• Evaluation & Key Findings

• Conclusion

Page 4: RISC vs CISC

Department of Electrical Engineering and Computer Science

What is CISC & RISC?CISC Approach :

MUL 2:3, 5:2

RISC Approach :

LOAD A, 2:3LOAD B, 5:2PROD A, BSTORE 2:3, A

Page 5: RISC vs CISC

Department of Electrical Engineering and Computer Science

1991 : METHOD

Page 6: RISC vs CISC

Department of Electrical Engineering and Computer Science

MIPS M/200 vs VAX 8700

Page 7: RISC vs CISC

Department of Electrical Engineering and Computer Science

SPECIFICATIONS

Page 8: RISC vs CISC

Department of Electrical Engineering and Computer Science

1991 : EVALUATION & KEY FINDINGS

Page 9: RISC vs CISC

Department of Electrical Engineering and Computer Science

METHODOLOGY

RESULT DISCUSSION

Page 10: RISC vs CISC

Department of Electrical Engineering and Computer Science

Result: Instructions & CPI

• Instruction Ratio: Ratio of MIPS instruction extentions to VAX.• RISC factor: Ratio of number of cycles per program on the VAX compared to MIPS

Page 11: RISC vs CISC

Department of Electrical Engineering and Computer Science

Result: Operation Counts

Page 12: RISC vs CISC

Department of Electrical Engineering and Computer Science

Result: Cache Behavior

• VAS cache results from hardware monitor.• MIPS cache results from simulator.

Page 13: RISC vs CISC

Department of Electrical Engineering and Computer Science

Discussion: Architectural Factors

Favoring MIPS:

• Operand specified decoding • Number of registers.• Floating-point hardware and instruction overlap.• Simple jumps and braches.• Fancy VAX instructions.• Instruction scheduling.• Translation buffers.• Branch displacement size.

Favoring VAX:• Big I-Stream constants. • Not-taken branches.

Page 14: RISC vs CISC

Department of Electrical Engineering and Computer Science

1991 : CONCLUSION

Page 15: RISC vs CISC

Department of Electrical Engineering and Computer Science

From architectural point of view: RISC as exmplified by MIPS offer a significant processor performance advantage over a VAX (CISC) of

comparable hardware organization.

Drawbacks:

• Compiler.• Number of bechmarks.• Application-level processor performance only. Not systems level.

Page 16: RISC vs CISC

Department of Electrical Engineering and Computer Science

2013 : METHOD

Page 17: RISC vs CISC

17

Department of Electrical Engineering and Computer Science

BeagleBoard ARM Cortex A8

Intel Atom N450

PandaBoard ARM Cortex A9

Intel Sandy BridgeCore i7

Linux 2.6

GCC

Page 18: RISC vs CISC

18

Department of Electrical Engineering and Computer Science

SPEC CPU2006

Desktop

CoreMarkWebKit

MobileLighttpdCLucene

Database kernels

Server

Page 19: RISC vs CISC

19

Department of Electrical Engineering and Computer Science

• Performance measurement on real hardware

• Extensive use of performance counters• Cycles, instructions, cache misses, branch misses…

• Power measurements using Wattsup meters

Page 20: RISC vs CISC

Department of Electrical Engineering and Computer Science

Page 21: RISC vs CISC

Department of Electrical Engineering and Computer Science

Page 22: RISC vs CISC

Department of Electrical Engineering and Computer Science

2013 : EVALUATION & KEY FINDINGS

Page 23: RISC vs CISC

Department of Electrical Engineering and Computer Science

METHODOLOGY

PERFORMANCE

POWER &

ENERGY

TRADE OFF

METHODOLOGY

PERFORMANCE

POWER &

ENERGY

TRADE OFF

METHODOLOGY

PERFORMANCE POWER & ENERGY TRADE OFF

Page 24: RISC vs CISC

Department of Electrical Engineering and Computer Science

Performance Analysis Flow:

• Present execution time for each benchmark.

• Normalize frequency’s impact using cycle counts.

• To understand differences in cycle count and the influence of the ISA, present the dynamic instruction count measures, measured in both macro-ops and micro-ops.

• Use instruction mix, code binary size, and average dynamic instruction length to understand ISA’s influence.

• To understand performance differences not attributable to ISA, look at detailed microarchitecture events.

• Attribute performance gaps to frequency, ISA, or ISAindependent microarchitecture features. Qualitatively reason about whether the ISA forces microarchitecture features.

Page 25: RISC vs CISC

Department of Electrical Engineering and Computer Science

Performance Analysis: Key Finding 1 Execution Time

• Large performance gaps exist across the four platforms studied, as expected, since frequency ranges from 600 MHz to 3.4 GHz and microarchitectures are very different.

Page 26: RISC vs CISC

Department of Electrical Engineering and Computer Science

Performance Analysis: Key Finding 2 Cycle-Count

• Performance gaps, when normalized to cycle counts, are less than 2:5 when comparing in-order cores to each other and out-of-order cores to each other.

Page 27: RISC vs CISC

Department of Electrical Engineering and Computer Science

Performance Analysis: Key Finding 3 Instruction Count

• Instruction and cycle counts imply CPI is less on x86 implementations: geometric mean CPI is 3.4 for A8, 2.2 for A9, 2.1 for Atom, and 0.7 for i7 across all suites. x86 ISA overheads, if any, are overcome by microarchitecture.

Page 28: RISC vs CISC

Department of Electrical Engineering and Computer Science

Performance Analysis: Key Finding 4 Instruction Format & Mix

• Combining the instruction-count and mixfindings, conclude that ISA effects are indistinguishable between x86 and ARM implementations.

Page 29: RISC vs CISC

Department of Electrical Engineering and Computer Science

Performance Analysis: Key Finding 5 Microarchitecture

Page 30: RISC vs CISC

Department of Electrical Engineering and Computer Science

Performance Analysis: Key Finding 5 Microarchitecture Contd.

•The microarchitecture has significant impact on performance. The ARM and x86 architectures have similar instruction counts. The highly accurate branch predictor and large caches, in particular, effectively allow x86 architectures to sustain high performance. x86 performance inefficiencies, if any, are not observed. The microarchitecture, not the ISA, is responsible for performance differences.

Page 31: RISC vs CISC

Department of Electrical Engineering and Computer Science

Performance Analysis: Key Finding 6 ISA Influence on Microarchitecture

• Beyond the translation to microops, pipelined implementation of an x86 ISA introduces no additional overheads over an ARM ISA for these performance levels.

Page 32: RISC vs CISC

Department of Electrical Engineering and Computer Science

Power and Energy Analysis Flow:

• Present per benchmark raw power measurements.

• To factor out the impact of technology, present technology-independent power by scaling all processors to 45nm and normalizing the frequency to 1 GHz.

• To understand the interplay between power and performance, examine raw energy.

• Qualitatively reason about the ISA influence on microarchitecture in terms of energy.

Page 33: RISC vs CISC

Department of Electrical Engineering and Computer Science

Power and Energy Analysis: Key Finding 7 Average Power

• Overall x86 implementations consume significantly more power than ARM implementations.

Page 34: RISC vs CISC

Department of Electrical Engineering and Computer Science

Power and Energy Analysis: Key Finding 8 Average Technology Independent Power

• The choice of power or performance optimized core designs impacts core power use more than ISA.

Page 35: RISC vs CISC

Department of Electrical Engineering and Computer Science

Power and Energy Analysis: Key Finding 9 Average Energy

• The choice of power or performance optimized core designs impacts core power use more than ISA.

Page 36: RISC vs CISC

Department of Electrical Engineering and Computer Science

Trade-off Analysis Flow:

• Combining the performance and power measures, compare the processor implementations using Pareto-frontiers.

• Compare measured and synthetic processor implementations using Energy-Performance Pareto-frontiers.

Page 37: RISC vs CISC

Department of Electrical Engineering and Computer Science

Trade-off Analysis: Key Finding 10 Power-Performance Trade-offs

• Regardless of ISA or energy-efficiency, high-performance processors require more power than lower performance processors. They follow well established cubic power/performance trade-offs.

Page 38: RISC vs CISC

Department of Electrical Engineering and Computer Science

Trade-off Analysis: Key Finding 11 Energy-Performance Trade-offs

• It is the microarchitecture and design methodologies that really matter.

Page 39: RISC vs CISC

Department of Electrical Engineering and Computer Science

2013: CONCLUSION

Page 40: RISC vs CISC

Department of Electrical Engineering and Computer Science

ISA being RISC or CISC does not matter for power and performance of modern

processors.

Page 41: RISC vs CISC

Department of Electrical Engineering and Computer Science

What is the ISA’s role?

• Supporting specialization•AVX crypto, Virtualization extensions•Jazelle DBX, ARM Trustzone…

• Exposing more workload-specific semantic information to the substrate•Transactional Memory support •Reliability-oriented extensions•Many more…

Page 42: RISC vs CISC

Department of Electrical Engineering and Computer Science

Questions?are guaranteed in life, answers aren't.

Page 43: RISC vs CISC

Department of Electrical Engineering and Computer Science

Thank You

http://chetanpatil.info/talks.html