Razor Systems

advertisement
Class presentation based on ISSCC 2013- 24.5 :
A Low-power 1GHz Razor FIR
Accelerator with Time-Borrow
Tracking Pipeline and Approximate
Error Correction in 65nm CMOS
By Paul N.Whatmough, Shidhartha Das, David M,Bull
ARM, Cambridge, United Kingdom
Presented by:
1Mahnaz Rasti
1Department of Electrical and Computer Engineering, University of Tehran, Tehran, Iran
Spring 2013
1
Outline:
• Razor Systems
• ANT(Algorithmic Noise Tolerance) Circuits
• Combining Technologies
• Architecture of the Chip
2
Outline:
• Razor Systems
• ANT(Algorithmic Noise Tolerance) Circuits
• Combining Technologies
• Architecture of the Chip
2
Razor Systems:
One of the more effective and widely used
methods for power aware computing is:
DVS (Dynamic Voltage Scaling)
Voltage  (π·π‘¦π‘›π‘Žπ‘šπ‘–π‘ πΈπ‘›π‘’π‘Ÿπ‘”π‘¦)2
4
Razor Systems:
• Razor propose a new approach to DVS, based
on dynamic detection of circuit timing errors
• Key Idea:
Tune the supply voltage by monitoring the error rate
during circuit operations[1]
5
[1] (MICRO-36), December 2003 - Razor: A Low-Power Pipeline Based on Circuit-Level Timing Speculation- Dan Ernst
[1]
Razor Systems(ex. flip-flop )
• Double samples pipeline stages value
• Once with a fast clock
• Again with a time borrowing delay clock
• A metastability-tolerant comparator then validates
latch value sampled with fast clock
• In the case of a timing error, a modified pipeline
mispeculation recovery mechanism restores correct
program state.
6
[1] Ref: (MICRO-36), December 2003 - Razor: A Low-Power Pipeline Based on Circuit-Level Timing Speculation- Dan Ernst
Razor Systems:
Figure 1. Pipeline augmented with Razor latches and control lines.
Ref:(MICRO-36), December 2003 - Razor: A Low-Power Pipeline Based on Circuit-Level Timing Speculation- Dan Ernst
7
Outline:
• Razor Systems
• ANT(Algorithmic Noise Tolerance) Circuits
• Combining technologies
• Architecture of the Chip
8
ANT Circuits:
Figure 2 Ref: http://icims.csl.uiuc.edu/~vips
9
ANT Circuits:
• Modify algorithm for system level error control
Figure 3 Ref:http://icims.csl.uiuc.edu/~vips
• ANT detects such errors in system output and mitigates
their effects on system performance.
• Errors are detected by low complexity prediction scheme.
10
ANT Circuits:
Figure 4 : Error predictor in ANT Circuits-[http://icims.csl.uiuc.edu/~vips]
11
Outline:
• Razor Systems
• ANT(Algorithmic Noise Tolerance) Circuits
• Combining technologies
• Architecture of the Chip
12
Combining technologies:
Razor Systems
Low power, fault tolerant
&
High clock frequency
Low overheads
ANT •
Rely on imbalance ripple adder
and hence limited clock
frequency, increasing baseline
area and power
•
13
Outline:
• Razor Systems
• ANT(Algorithmic Noise Tolerance) Circuits
• Combining technologies
• Architecture of the Chip
14
Architecture of the Chip
Approximate Error Correction
Time Borrow Tracking
15
Figure5: FIR accelerator with Razor latches, time-borrow tracking (TBT) and approximate
error correction (AEC)- ISSCC 2013 – 24.5
Architecture of the Chip
16
Figure6: FIR accelerator with Razor latches, time-borrow tracking (TBT) and
approximate error correction (AEC)- ISSCC 2013 – 24.5
Characteristics:
• Tow distinct error correction technique
• TBT
• AEC
• A 1Ghz datapath due to elimination of ripple-carry
adders
• Energy efficiency improvement of up to 37%
17
Die:
Technology Node
TSMC CMOS 65nm LP
Dimension
530 micron * 350 micron
Pipeline Design
16-Tap FIR, 8b coefficient , 18b o/p
Total FF/RZL
1974(393 in pipeline)/ 120 (30% of
pipeline FF)
Hold Buffer Area Overhead
1.59%
Max Clock Frequency
1.008 GHz @ 1.2v with Razor
18
Figure7: Die photo- ISSCC 2013 – 24.5
References:
1. ISSCC 2013 – 24.5 – “A Low-Power 1GHz Razor FIR Accelerator with
Time-Borrow Tracking Pipeline and Approximate Error Correction in
65nm CMOS” - Paul N. Whatmough - ARM, Cambridge, United Kingdom.
2. (MICRO-36), December 2003 – “Razor: A Low-Power Pipeline Based on
Circuit-Level Timing Speculation”- Dan Ernst
3. “Algorithmic Noise Tolerance for Low Power Signal Processing in the
Deep SubMicron Area” - ECE Department University of Illinois at UrbanaChampaign
19
Download