19
Karthik Aadithya, UC Berkeley ([email protected]) Slide 1/20 2014/03/17, Designing with Uncertainty Workshop, York MUSTARD: A CAD tool for Predicting the Impact of Random Telegraph Noise on SRAMs and DRAMs Karthik Aadithya ([email protected]) Joint work with Alper Demir, Koc University, Istanbul, Turkey Sriramkumar Venugopalan, UC Berkeley Jaijeet Roychowdhury, UC Berkeley

MUSTARD: A CAD tool for Predicting the Impact of Random ... Jaijeet Roychowdhury.pdf · What causes RTN? Capture Release Random Processes Filled traps modify number, mobility of electrons

  • Upload
    others

  • View
    2

  • Download
    0

Embed Size (px)

Citation preview

Karthik Aadithya, UC Berkeley ([email protected]) Slide 1/202014/03/17, Designing with Uncertainty Workshop, York

MUSTARD: A CAD tool for Predicting the Impact of Random Telegraph

Noise on SRAMs and DRAMs

Karthik Aadithya([email protected])

Joint work with

Alper Demir, Koc University, Istanbul, TurkeySriramkumar Venugopalan, UC Berkeley

Jaijeet Roychowdhury, UC Berkeley

Karthik Aadithya, UC Berkeley ([email protected]) Slide 2/202014/03/17, Designing with Uncertainty Workshop, York

Overview of this talk

• Why worry about RTN? (SRAMs, DRAMs)

• RTN basics

• Our contributions: RTN+circuit co-simulation• discrete Monte-Carlo ↔ nonlinear ckt. simulation

Karthik Aadithya, UC Berkeley ([email protected]) Slide 3/202014/03/17, Designing with Uncertainty Workshop, York

Importance of SRAMs and DRAMs

Core i7 die

Laptop, desktop, tablet memories

Billions of DRAM cells

Cache memory (L1, L2, L3, etc.)

Millions of SRAM cells

Karthik Aadithya, UC Berkeley ([email protected]) Slide 4/202014/03/17, Designing with Uncertainty Workshop, York

6T SRAM cell: Write Operation

● Writing a 1 to the SRAM cell● Switch BL to high, BL_bar to low● Briefly enable WL

● By the end of the clock cycle● Q should settle to 1● Q_bar should settle to 0

● Cross-coupled inverter pair will maintain this (stable) state

● Key idea: Back-to-back inverters

Karthik Aadithya, UC Berkeley ([email protected]) Slide 5/202014/03/17, Designing with Uncertainty Workshop, York

How RTN can impact SRAM write

RTN can induce dynamic SRAM write failure!(read failures also reported)

Karthik Aadithya, UC Berkeley ([email protected]) Slide 6/202014/03/17, Designing with Uncertainty Workshop, York

RTN in SRAMs: Experimental Evidence

Measured data ● Confirms temporal SRAM failures due to RTN● Quantifies RTN in terms at circuit level

Fig. Credit: Seng O. Toh, PhD thesis, UCB

Write to SRAM

Read from SRAM.Record errors.

Select VddFor each , record min, max fail bit count

Vdd

Vdd

Karthik Aadithya, UC Berkeley ([email protected]) Slide 7/202014/03/17, Designing with Uncertainty Workshop, York

Noise: An SRAM designer's viewpoint

Figure credits: Y. Tsukamoto, Renesas Electronics Corp.

RTN impact is steadily increasing!

At 22nm, RTN can drive

design margins negative!

RTN-induced SRAM failures experimentally reported

Seng et. al. (IEDM 2009), Yas et. al. (IRPS 2010)

Karthik Aadithya, UC Berkeley ([email protected]) Slide 8/202014/03/17, Designing with Uncertainty Workshop, York

What causes RTN?

Capture

Release

Random ProcessesFilled traps modify number, mobility

of electrons in inversion layer

Change in drain current Measured as RTN

Karthik Aadithya, UC Berkeley ([email protected]) Slide 9/202014/03/17, Designing with Uncertainty Workshop, York

RTN is Non-Stationary

Bias-dependence leads to non-stationary RTN

Capture

Release

V_gs

Hz

Karthik Aadithya, UC Berkeley ([email protected]) Slide 10/202014/03/17, Designing with Uncertainty Workshop, York

RTN: Bridging the Gap

?

CAD tool incorporating device level RTN models for circuit level non-stationary

RTN characterisation

Device level RTN models

Circuit level RTN measurementsMUSTARD

Karthik Aadithya, UC Berkeley ([email protected]) Slide 11/202014/03/17, Designing with Uncertainty Workshop, York

MUSTARD's RTN Model

Core idea: Simulate the bi-directionally coupled MC-DAE system using an

intelligent Monte-Carlo scheme

Markov chain(RTN) Coupled

Markov state affects DAE's q(.), f(.)

DifferentialAlgebraic Equations(Rest of the circuit)

ongoing DAE solution affects Markov propensities

DiscreteStochastic

ContinuousDeterministic

Karthik Aadithya, UC Berkeley ([email protected]) Slide 12/202014/03/17, Designing with Uncertainty Workshop, York

MUSTARD: Non-stationary Trap Simulation with Bi-Directional Coupling

Keep → update rate

… and so on

Keep → update rate

Discard → don't update rate

Karthik Aadithya, UC Berkeley ([email protected]) Slide 13/202014/03/17, Designing with Uncertainty Workshop, York

MUSTARD: Simulation Methodology

Generate candidate RTN event

Start

Simulate circuit until event

Probabilistically decide: keep (or) discard event?

Update circuit equations

Keep!

Dis

card

Discard

Keep

Exact, non-stationary statistics preserved!

Karthik Aadithya, UC Berkeley ([email protected]) Slide 14/202014/03/17, Designing with Uncertainty Workshop, York

MUSTARD: Individual SRAM cell

No RTNNo V_th variability

No RTNYes V_th variability

Yes RTNYes V_th variability

Karthik Aadithya, UC Berkeley ([email protected]) Slide 15/202014/03/17, Designing with Uncertainty Workshop, York

RTN+Variability: MUSTARD on 6T SRAM across (VDD,Vth) landscape

Without RTNWith RTN

Karthik Aadithya, UC Berkeley ([email protected]) Slide 16/202014/03/17, Designing with Uncertainty Workshop, York

MUSTARD-generated BER vs VDD plots

● asdf

Karthik Aadithya, UC Berkeley ([email protected]) Slide 17/202014/03/17, Designing with Uncertainty Workshop, York

MUSTARD: RTN Effects on DRAMs

● asdf

Karthik Aadithya, UC Berkeley ([email protected]) Slide 18/202014/03/17, Designing with Uncertainty Workshop, York

Summary and Conclusions

● RTN is a concern for SRAM scaling

● Bottom-up prediction of bit errors challenging● discrete random events + non-stationarity + “feedback”

● MUSTARD offers a simulation-based solution● strong mathematical guarantee on accuracy: applies to

• any trap configuration• any circuit (SRAM, DRAM, etc.)• any device model, any model for RTN, etc.

Karthik Aadithya, UC Berkeley ([email protected]) Slide 19/202014/03/17, Designing with Uncertainty Workshop, York

Publications

(1) Aadithya V Karthik, Alper Demir, Sriramkumar Venugopalan and Jaijeet Roychowdhury. SAMURAI: An accurate method for modelling and simulating non-stationary Random Telegraph Noise in SRAMs. In Proceedings of the Design, Automation and Test Conference in Europe, 2011.

(2) Aadithya V Karthik, Sriramkumar Venugopalan, Alper Demir and Jaijeet Roychowdhury. MUSTARD: A coupled, stochastic-deterministic, discrete-continuous technique for predicting the impact of Random Telegraph Noise on SRAMs and DRAMs. In Proceedings of the Design Automation Conference 2011.

(3) Aadithya V Karthik, Alper Demir, Sriramkumar Venugopalan, and Jaijeet Roychowdhury. Accurate Prediction of Random Telegraph Noise Effects in SRAMs and DRAMs. In the 2013 IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, Volume 32, Issue 1.