Gene A. Frantz, TI Senior Fellow

1
The Case for Programmability
or
How I Learned to Stop Worrying and love DSPs

• Gene A. Frantz, TI Senior Fellow
• Mike McMahan, TI Fellow
• Texas Instruments Inc.

2
Different Perspectives

• UCB and Texas Instruments sometimes have
  different views on how to approach the solution
  to a system design

• And the reason isnt simply because we are from
  Texas and you are from California

3
To the man with a hammer, every problem looks
like a nail
A custom data path would be the most efficient
solution to this problem !! Unnamed UCB
Professors
4
To the man with a hammer, every problem looks
like a nail
By gosh, this looks like a problem for a DSP !!
Unnamed TI Engineer
5
The Real Issue

• While a programmable DSP isnt ALWAYS the right
  answer, our experience is that in most cases it
  is.
• Since engineering analysis can lead to a
  different conclusion,
• The question is WHY does DSP usually win in the
  market ?

6
You Understand the Problem
Extracted from BWRC presentations
7
And the Range of Solutions
Extracted from BWRC presentations
8
But the answer isnt this simple..

• Software solutions 100 times less efficient
  (even ignoring overhead of parallel processing)
• .5-5 MIPS/mW software DSP (best case) processor
• 100-1000 MOPS/mW dedicated

Extracted from BWRC presentations
9
Agenda

• Audience
• Participation

10
Cellular Phone An example
Digital Cellular Market (Phones Shipped)
11
The Cellular Telephone

• Perhaps the highest volume consumer product in
  history
• TI estimates that 435M cell phones will ship in
  2000
• Many standards (e.g. GSM) are relatively mature
• This is the ideal market for a hardwired solution
• Yet every cell phone which is shipped is based on
  a programmable DSP----WHY ?

12
At Least 4 Reasons

• Need to take a complete systems view
• Dont look for local minimums
• Need flexibility
• The world changes
• Need to shorten time to market
• Cant afford perfect solutions
• Need to maximize and maintain value
• Programmability facilitates creation of a value
  web

13
Need a Complete System View

• Does it matter that I squeeze the last mA out of
  a function if that function accounts for only 10
  of the active power consumed in the phone
• Remember, the PA accounts for about 70 of the
  phones active power consumption
• and progress in silicon technology continues to
  reduce digital power consumption (Genes Law)

14
Change Happens !

• Standards change, product requirements and
  features evolve
• Cannot afford to spin new hardware whenever
  upgrades are required
• Examples
• Vocoder/FEC modifications
• New feature requirements

15
Need to Shorten Time to Market

• Phone replacement rates exceed 50 / yr
• Development cycles measured in months
• Each manufacturer offers dozens of phone models
  (e.g. Ericssons website lists 31 different GSM
  phones)
• Virtually every manufacturer moving to one
  programmable platform as the basis for all phone
  designs

16
Need to Maximize Value

• Value in a fixed solution declines over time
• The next solution to market will be better
• Value in a programmable solution can increase
  over time
• 3rd parties can add functionality to your
  solution and maintain competitive balance with
  competing solutions
• 3rd parties have a stake in your solution

17
Cellular Phone Block Diagram
SINGLE CHIP ANALOG BASEBAND POWER MANAGEMENT
RF Interface
Audio Interface
Op Amps
Switches
Speaker
Mic
Regulators
User Display
Keyboard
Touch Screen
SIM Card
18
Cellular Phone Baseband SOC
ROM
MCU
DSP
Gates
RAM
Analog
2000 phones on each 8 wafer _at_ .15 Leff
1M Baseband Chips per Day
19
Agenda

• Audience
• Participation

• Introduction
• The cell phone scenario

More examples Calculators Speak N
Spell Modems MP3 player Hearing aids
The Case for Programmability
20
Calculator - the first SOC

• The goal of the calculator business was a single
  chip solution
• This was achieved in the Mid 70s
• But it still had external components
• Keyboard
• Display
• As the single chip was programmable it was used
  on a whole family of calculator products with
  only the program changed

21
Agenda

• Audience
• Participation

• Introduction
• The cell phone scenario

More examples Calculators Speak N
Spell Modems MP3 player Hearing aids
The Case for Programmability
22
Speak and Spell, 1978

• First single chip DSP
• LPC-10 speech synthesizer
• Fixed function
• Introduced with largest memory devices of that
  period (128k ROM)
• Controller was TMS1000 class uC
• Same uC core as all other TI consumer products
  of that period
• Custom peripherals added to drive display and
  keyboard

23
Agenda

• Audience
• Participation

• Introduction
• The cell phone scenario

More examples Calculators Speak N
Spell Modems MP3 player Hearing aids
The Case for Programmability
24
Modem A Case Study of Programmability

• Modems born as systems made up of analog
  components
• V.21/Bell 103 (300bps)
• TIs first custom implementation of a modem
• It had a mixture of custom analog and digital on
  the same device
• V.22/Bell 212 (1200bps)
• TIs last attempt at a fixed function modem
• After multiple passes at silicon we gave up and
  took it off of the market
• With the introduction of the programmable DSP,
  the modem market entered a new way of designing
  modems.
• Allowed upgrades through software downloads
• From v.22bis (2400bps) on programmable solutions
  dominated the modem industry

25
Modem Case Study
26
Modem Case Study
100
V.32bis
Volume
150
69.95
100
200
V.32
150
69.95
200
Time
27
Modem Case Study
100
V.34
150
69.95
100
200
V.32bis
Volume
150
69.95
100
200
V.32
150
69.95
200
Time
28
Agenda

• Audience
• Participation

• Introduction
• The cell phone scenario

More examples Calculators Speak N
Spell Modems MP3 player Hearing aids
The Case for Programmability
29
MP3 player A case study of flexibility

• Music players need flexibility
• There are many encoding standards
• MP3
• AAC
• Dolby Digital
• EPAC
• ATRAC
• There are other interesting sounds and associated
  encoders

Sony Vaio Music Clip
30
Agenda

• Audience
• Participation

• Introduction
• The cell phone scenario

More examples Calculators Speak N
Spell Modems MP3 player Hearing aids
The Case for Programmability
31
Hearing Aids A Case Study of Too Many Problems
to Solve

• Hearing Aids have historically been custom analog
  solutions
• They are making a transition to digital now
• The debate should it be custom or programmable

• What the user wants
• To hear
• Long battery life
• Invisible
• Easy to use

• The problem to be solved
• Every person has a different hearing problem
• As we mature the problem changes
• Each new room environment changes the problem

The solution Programmability at low power
32
Agenda

• Audience
• Participation

• Introduction
• The cell phone scenario

More examples Calculators Speak N
Spell Modems MP3 player Hearing aids
The Case for Programmability
33
The Case for Programmability
Cost Power Dissipation Size System
Considerations Seems like a good thing to do
34
Two Decades of Integration
Typical Device capabilities
2000
2010
1980
1990

• 50
• 0.1
• 5,000
• 1,000
• 32K
• 5.00
• 0.1
• 5M
• 12"

Die size (mm) Technology (uM) MIPS MHz RAM
(bytes) Price Power (mW/MIPS) Transistors Wafer
size

• 5
• 0.02
• 50K
• 10,000
• 1M
• 0.15
• 0.001
• 50M
• 12"

• 50
• 3
• 5
• 20
• 256
• 150.00
• 250
• 50K
• 3"

• 50
• 0.8
• 40
• 80
• 2K
• 15.00
• 12.5
• 500K
• 6"

35
The Case for Programmability
Cost Power Dissipation Size System
Considerations Seems like a good thing to do
36
Power Dissipation Trends
1,000
Genes Law DSP Power
100 10 1 0.1 0.01 0.001 0.0001 0.00001
1982 1984 1986 1988 1990 1992 1994 1996 1998 2000
2002 2004 2006 2008
mW/MIPS
Year
37
The Case for Programmability
Cost Power Dissipation Size System
Considerations Seems like a good thing to do
38
The Question of Size

• Device size has become a non-issue as a result of
  process technology
• CPUs are close to or at 1mm in die size and
  shrinking
• ASIC gate density is 100K gates per mm or greater
• Memories continue to shrink
• Systems are getting more complex
• BUT
• Human Factors are not shrinking
• Hand helds and Desktops are still the same basic
  size
• Keyboards and Displays still need to be large
  enough to use
• SO
• Most products do not need the whole system on one
  chip
• The system can be broken into major sub-systems,
  e.g.
• Analog
• Digital

39
The Case for Programmability
Cost Power Dissipation Size System
Considerations Seems like a good thing to do
40
The Value of Time

• At introduction, value has three components
• Hardware
• Software
• Time to market

TTM

• At maturity, product value has only one
• component
• Hardware

S/W
Value
H/W
H/W
Time
41
The Value of Complexity
Value and Complexity are correlated to some
point of complexity
Beyond this point the value is reduced
with added complexity
Value
Complexity
42
Three Vectors of Value
Raw Performance

• Each vector of value creates new market
  opportunities
• SOC effects all three vectors
• Each end equipment requires a unique combination
  of the three

Power Dissipation
Cost
43
The Case for Programmability
Cost Power Dissipation Size System
Considerations Seems like a good thing to do
44
How to do SOC
Custom Layout ASIC Programmable Combination
45
A Different Look at Programmability
A combination of software and hardware always
gives the lowest cost system design.

• Cost can be defined as
• Financial
• Power Dissipation
• NRE
• Time to market

• Mfg cost
• Weight
• Opportunity cost
• Size

Cost
Cost
Tech trend
Combination
100 H/W (Fixed Function)
100 S/W (Programmable)
46
HW/SW Combination Still Can Be Flexible

• Use DSP for anything which can be accomplished in
  SW
• Add a dash of programmable (or reconfigurable) HW
  coprocessors where necessary

47
Conclusion

• Use programmable DSP where you can
• Use flexible HW where you must
• Dont wait for perfection
• Move quickly