Memory Design Considerations That Affect

Price and Performance

Memory Design Considerations That Affect

Price and PerformanceBill Gervasi

Technology Analyst, Transmeta

Chairman, JEDEC Memory Parametrics

bilge@transmeta.com

Posed at the Last ConferencePosed at the Last Conference

Why will DDR-I at 400 MHz data rate be a “boutique” solution?

Why will DDR-II at 400 MHz data rate be a “mainstream” solution?

AgendaAgenda

JEDEC/Industry RoadmapFactors for Market AcceptanceDifficulties in Achieving 400 MHzFactors Affecting CostWild Cards – What Can Change?

RAM EvolutionRAM Evolution

2100MB/s

2700MB/s

MainstreamMemories

DDR266

DDR333Simple,

incrementalsteps

DDR4003200MB/s DDR533

4300MB/s DDR6675400MB/s

“DDR I”

“DDR II”

Factors for Market AcceptanceFactors for Market Acceptance

Industry FocusNumber of Competing SuppliersJEDEC standardLaws of Physics

Industry FocusIndustry Focus

The JEDEC roadmap represents the industry focus for mainstream productsDDR-I tops out at 333 MHz data rateDDR-II starts at 400 MHz data rate

This DOES NOT mean that DDR-I at 400 MHz data rate will not ship in volume

It DOES mean that there will be price premiums for this speed bin

What do I mean by “Focus”?What do I mean by “Focus”?

There is serious work to hit 400 MHzVendor interoperable solutionsMix and match module configurationsSignal integrity analysis

We are counting picosecondsNo JEDEC standard yet proposed for DDR-I

at 400 MHz data rate

For Example…For Example…

How we are getting more refined in timing analysis with DDR-II…

The Charge Transfer Model for input timing measurement and derating

DDR-I Input Timing ModelDDR-I Input Timing Model

INPUT

Setup Hold

CLOCK

CLOCK

Timing derating by input signal slew rate:

1.0V/ns = base value

0.5V/ns = base value + 50ps

0.4V/ns = base value + 100ps

This got us through DDR333…

The Old Way

However…However…

This simplified model was good enough for DDR333 data rates, but leaves picoseconds of available timing lying around needed for 400+!!!

DDR266 Data Setup/Hold = 750 psDDR333 Data Setup/Hold = 600 psDDR400 Data Setup/Hold = 400 psDDR533 Data Setup/Hold = 350 ps

Can’t waste time!!!

“Focus” on Input Timing“Focus” on Input Timing

INPUT

DDR-II Charge Transfer Timing ModelAll inputs have a slew rate dependent aspect tEXT

and an independent aspect tINT

Summing tEXT + tINT gives input transition time tT

Transition time tT has min and max valuesDifferential input transitions inherently different

tINTtEXT

tT

The New Way

tEXT for Slow Slew Rate, Single EndedtEXT for Slow Slew Rate, Single Ended

VREF

VIHAC = VSAT

VIHDC

VILDC

VILAC=

VSAT

tEXT

AT = Charge to Transition

slew

At TEXT

*2

tINT

tEXT for Fast Slew Rate, Single EndedtEXT for Fast Slew Rate, Single Ended

VREF

VIHAC =VSAT

VIHDC

VILDC

VILAC =VSAT

tEXT

tSAT

ASAT = Charge to Saturation

AADD = Charge after Saturation

AT = ASAT + AADD

SAT

SATTSATEXT V

AAtt

)(

VSAT = Saturation Voltage

tINT

tEXT for Slow Slew Rate, DifferentialtEXT for Slow Slew Rate, Differential

VREF

VIHAC =VSAT

VIHDC

VILDC

VILAC =VSAT

AT = Charge to Transition

tEXTtINT

tEXT for Fast Slew Rate, DifferentialtEXT for Fast Slew Rate, Differential

VREF

VIHAC =VSAT

VIHDC

VILDC

VILAC =VSAT

AT = ASAT + AADD

tEXT tINT

“Focus” on Timing“Focus” on Timing

INPUT

Setup

CLOCK

CLOCK

DDR-II Charge Transfer Timing ModelSetup = tTmax of input - tTmin of reference

tTmax

tTmin

“Focus” on Timing“Focus” on Timing

INPUT

Hold

CLOCK

CLOCK

DDR-II Charge Transfer Timing ModelHold = tTmax of reference - tTmin of input

tTmin

tTmax

How does this help…?How does this help…?

The Charge Transfer Model gives a higher accuracy for setup and hold relationships

It also provides a way to accurately describe derating for input slew rate

These models are negotiated with all suppliers to define an industry standard

DDR-II Input Derating TablesDDR-II Input Derating Tables

2.00.5

0.5

1.0

1.0

2.0

Strobe (mV/ps avg)

Dat

a (m

V/p

s)

2.00.5

0.5

1.0

1.0

2.0

Clock (mV/ps avg)

Add

r (m

V/p

s)

2.00.5

0.5

1.0

1.0

2.0

2.00.5

0.5

1.0

1.0

2.0

HOLD

SETUP SETUP

HOLD

0

0

+

+

+

+

+

+

Strobe (mV/ps avg)

Dat

a (m

V/p

s)

Clock (mV/ps avg)A

ddr

(mV

/ps)

++

0

+

+

++

0

+

+

+

+

Derating Using Charge TransferDerating Using Charge Transfer

Accuracy from derating both signals and references Result is a two dimensional matrix relating inputs & their

references Identified inherent asymmetries in derating of setup & hold

when mixing single ended with differential signals Memory module mixes impact slew rates

The Charge Transfer model controls system cost by enabling more complex timing analysis

Charge Transfer on DDR-I?Charge Transfer on DDR-I?

This model would also help design high speed DDR-I systems

However, the work to retrofit this to DDR-I needs to be done to benefit from it

DDR-II ImprovementsDDR-II Improvements

DDR-II introduces technical improvements that reduce the cost of achieving high speedsPrefetch 4Differential data strobeI/O CalibrationLower I/O VoltageOn-Die Termination

Prefetch 4Prefetch 4

Moving to the Next LevelMoving to the Next Level

Today’s SDRAM architectures assume an inexpensive DRAM core timing

DDR I (DDR200, DDR266, and DDR333) prefetches 2 data bits: increase performance without increasing timing costs

DDR II (DDR400, DDR533, DDR667) prefetches 4 bits internally, but keeps DDR double pumped I/O

Prefetch 2 Versus 4Prefetch 2 Versus 4CK

READ

Prefetch 2

Prefetch 4

Core access time

Costs $$$

Essentially free

data

Column cycle time

Costs $$$

Prefetch Impact on CostPrefetch Impact on CostBy doubling the prefetch depth, cycle time for column

reads & writes relaxed, improving DRAM yields

DDR-I

DDR-II

Pre-fetch

2

2

2

4

4

4

266

333

400

400

533

667

7.5 ns

6 ns

5 ns

6 ns

7.5 ns

10 ns

DDR Family

Data Rate

Cycle Time

Starts to get REAL EXPENSIVE!

Comparable to DDR266 in cost

Why Not Prefetch = 8?Why Not Prefetch = 8?

DIMM width = 64 bitsPCs use 64b, servers use 128b (2 DIMMs)

64 byte prefetch okay for PC, but…128 byte prefetch for servers wastes bandwidth

DDR-II must service all applications well to insure maximum volume minimum cost

Differential Data StrobeDifferential Data Strobe

Differential Data StrobeDifferential Data Strobe

Just as DDR added differential clock to SDRDDR II adds differential data strobe to DDR I

Transition at the crosspoint of DQS and DQS

Differential Data StrobeDifferential Data Strobe

DQShigh time

VREF

DQSlow time

DQS

DQShigh time

VREF

DQSlow time

DQS

Normal balanced signal

Mismatched Rise & Fall signal

Error!

Differential Data StrobeDifferential Data Strobe

DQShigh time

VREF

DQSlow time

DQS

DQShigh time

VREF

DQSlow time

DQS

Normal balanced signal

Mismatched Rise & Fall signal

DQS

DQS

Significantly reduced symmetry error

I/O CalibrationI/O Calibration

I/O CalibrationI/O Calibration

Balance pull-up and pull-down driver strengthReduces timing errors from signal asymmetryInsures signal rise and fall times are similar

Reference

Data

DataController

DRAM

1.8V I/O Voltage1.8V I/O Voltage

1.8V Signaling1.8V Signaling2.5V

SSTL_18

1.60V

0.90V

1.43V

1.07V

1.25V

0V

0.90V1.03V

0.77V0.65V

1.15V

1.8V

VSS

VDDQ

VREF

VIHac

VIHdc

VILdc

VILacVREF

VSS

VDDQ

VIHac

VIHdc

VILdc

VILac

SSTL_2

I/O Voltage Impact on TimingI/O Voltage Impact on Timing

Assume 1mV/ps edge slew rateDDR-I = 700 mV (VILVIH) = 700 ps

DDR-II = 500 mV (VILVIH) = 500 ps

Helps meet the need for speedSignal integrity is a serious challenge at

DDR-I and 400 MHz data rate

On-Die TerminationOn-Die Termination

On-Die TerminationOn-Die Termination

Reduces system cost while improving signal integrity

Data

Controller

VTT =VDDQ 2

DRAM

Data

Controller

DRAM

VDDQ 2VDDQ 2

DDR-I

DDR-II

What Can Change?What Can Change?

Wild CardsWild Cards

100% yield of 5 ns cycle time cores (magic?)Industry gets excited about engineering

DDR-I at 400 MHzDDR-II slow transition from schedule or price

Feature creepDie penaltiesDRAM guys trying to make money for once

ConclusionsConclusions

DDR-I at 400 will ship in volume but…not likely to cross over $/bitIndustry focus is on transition to DDR-II

for 400+ MHz data rates

Thank YouThank You