Title: Soft Errors in CAM and Their Impact On TLB Reliability
1Soft Errors in CAM and Their Impact On TLB
Reliability
- Feihui Li
- Swapna Dontharaju
2Introduction
- TLB - a small cache, to speed up address
translation in processors with virtual memory - Soft errors can significantly affect memories
- Because of spatial density, amount of information
stored - We are studying their impact on CAMs
- On the other hand, TLB is a major power consumer
- STRONG ARM processor, 17 of total power
consumption due to TLB though designed for low
power, and TLB is very small - We have compared Soft Error Susceptibility in CAM
designs that reduce the power consumption
3Analyzing the Basic CAM cell
- Write a 1(0)
- Pass transistors T1, T2 opened via Word line
- Bit set to 1, Bit to its inverse
- T6(T5) conducting, T5(T6) not ? storage of 1
- Search (1)
- Match line precharged high
- Search bit on Bit (1), its inverse on Bit (0)
- If 1 at Q, T8 propagates 0 (from Bit) to gate
of T9 - T9 stays nonconducting, Matchline not
discharged - Obvious Questions
- During a Write, What if there is a flip at T1
or T2 - Flip at both circuit will search instead
mismatch? - During a Search, If T7 / T8 flip will the
result be correct? - If T9 flips, Match line discharges even if no
search? - Strikes at T3, T4, T5, T6 Flip in the value
stored?
4In the TLB ..
- A flip at any of the transistors connected to the
match line can cause the match line, (even if
search successful) to discharge. - So even if address translation was available in
the TLB, the instruction could have missed the
TLB! Performance, power penalty.
5 Designs studied
- Basic CAM
- Necessary to put all bit lines to 0, to
precharge Match line high switching in bit
lines - Modified CAM
- Control line used, to precharge match line
without putting to zero the bit lines reduces
switching activity of bit lines. - Toggling Match Line CAM
- CAM alternates between active high and active
low match line output for every access reduces
switching activity in match lines.
6Charge Induction Scenarios for CMOS circuits
- Cases II I Voltage at p node go up
- Cases III IV Voltage at n node go down.
- II IV wont affect logic state of circuit,
node is already at logic value toward which
injected charge will drive it - III I affect logic value of node, may cause
incorrect operation of circuit, modeled below
Resistors ? conducting transistors, Rectangles
?drain Current sources show direction of current
flow
7Experimental Setup
- When a 1 is stored in the CAM cell, Cases III
I modeled as, - Positive pulse at Q to induce a bit flip
- Negative pulse at Q to induce a bit flip
- Similar analysis when CAM cell stores a 0
- Qcritical a measure of the current required
to flip a memory cell - Qcritical of the nodes in the cell found Q,
Q, T9. - Using Micro Magic and HSPICE, current spikes
can only be injected at nodes, (transistors
lumped by interconnect say Q, Q) rather than
each individual transistors. Not possible to
study some of the interesting flips mentioned
above, in isolation.
8Results
9Results
- Qcritical is different for each node
- Qcritical depends highly on state of the cell
precharge, read or write - Match line transistor T9 in Basic CAM, more
susceptible to bit flip than other nodes
10Analysis
A method to estimate Soft Error Rate (SER) using
Qcritical has been developed by Hazucha and
Svensson Nflux intensity of Neutron flux,
CS area of crossection of node, Qs charge
collection efficiency, Qcritical minimum charge
collected due to a particle strike that causes a
bit flip Id drain current induced by the ion,
Tf flipping time Qcritical has been used as an
index to the SER.
11Work in progress..
- We are working on the following -
- Impact of technology scaling on SER of Basic
CAM and TLB structure (180nm down to 70 nm) - For a given technology, see whether the Basic
CAM or the 6T-SRAM is more susceptible to soft
errors for a given row of the TLB, which of
them contributes more to SER - Model Power consumption of a 16 word CAM, each
word having 16 bits, using each of the above CAM
cells - Study the Qcritical Vs Power consumption
tradeoffs
12Thank You !!
13References
- 1 Thirugnanam, G. Vijay Krishnan, N. Irwin,
M.J.A novel low power CAM design, Proceedings
of the 14th Annual IEEE International ASIC/SOC
Conference, Sept. 2001, Page(s) 198 -202 - 2 T.Juan, T.Lang and J.Navarro, Reducing TLB
Power Requirements, Proceedings of the 1997
International symposium on Low power electronics
and design, 1997, pp. 196-201 - 3 V. Degalahal, N. Vijay Krishnan, M.J Irwin,
"Analyzing Soft Errors in Leakage Optimized SRAM
Design", In the Proceedings of VLSI Design
conference January 2003, New Delhi, India - 4 H Cha, J H Patel A logic level model for
alpha particle hits in CMOS circuits Proceedings
of the International Conference on Computer
Design, 1993. - 5 P. Hazucha and C. Svensson, Impact of CMOS
technology scaling on the atmospheric neutron
soft error rate, IEEE Transactions on Nuclear
Science, 2000 - 6 D.C. Burger and T.M. Austin. The SimpleScalar
tool-set, Version 2.0. Technical Report 1342,
Dept. of Computer Science, UW, June 1997 - 7 The SimpleScalar tool-set. http//www.simplesc
alar.com/