Transcript Slide 1

Power and a New Class of
Future FPGA Architectures
Sinan Kaptanoglu
Actel, Fellow
A Rough Outline
 A historical look at FPGAs
 A new market driver?
 Future Technologies
 Future Design Methods
 Future Architectures
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
2
A Historical Look at FPGAs
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
3
Infancy (1985-1990)
 FPGAs were invented in the 1985-1987 time
frame.
● Other programmable logic (PALs and PLDs) predate
FPGAs by several years.
 Infancy period is marked by:
● Immature architectures and SW.
● Too small, too slow, and too expensive for anything
other than glue logic.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
4
Childhood (1990-1994)
 Very exciting and inventive years.
● Many startups, many new architectures.
● PLDs grew into CPLDs and temporarily challenged
the FPGAs.
● FPGAs grew bigger, faster and cheaper.
● Software became much better.
● Synthesis replaced schematic capture.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
5
But…
 By and large, FPGAs were still limited to
doing glue logic.
● For most applications, FPGAs were still not big
enough, fast enough, or easy enough to design with.
 The IC industry was driven by the PCs, in
which FPGAs played no role at all.
 In 1993, it looked like the entrenched ASICs
would not budge at all.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
6
Then something new came…
 Something that profoundly changed the
FPGA landscape.
● It was NOT a fantastic new FPGA architecture…
● It was NOT a great and totally unexpected
breakthrough in technology.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
7
It was a new market for FPGAs
 “The Internet Age” had begun.
● The rapidly growing datacom/telecom markets
valued design flexibility and quick time-to-market
more than cost or performance.
● Suddenly, everybody needed bigger and faster
routers and base infrastructure.
● Development cycles for datacom were shrunk to
mere months.
● The volumes were low (not in millions).
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
8
Datacom/telecom and FPGAs
 FPGAs and datacom/telecom ideally suited
each other, and they mutually drove each
other’s boom:
● FPGAs provided the needed flexibility and made
new datacom/telecom products possible in short
development cycles.
● Datacom/telecom rapidly increased the FPGA
revenues, starting a boom.
● New applications ignited a rapid increase in the
density and performance of FPGAs.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
9
The young adult years (1995-2001)


The datacom/telecom markets grew very rapidly, and
their share of the programmable market grew with it.
● Communication’s share was less than 10% in 1994.
● It peaked out at 58% by 2000.
New architectures were developed and refined to
better serve this market.
● FPGA vendors had a great ally: Moore’s law and the ever
shrinking process technology!


12/12/2007, IC-FPT,
Kitakyushu, Japan
This allowed FPGAs to keep up with the needs of datacom by
doubling the density every two years.
FPGAs became the next generation process drivers in the
process foundries.
Sinan Kaptanoglu, Actel Corp.
10
Changing landscape…
 Many startups died in this period.
 FPGAs which were not targeted to datacom
fell behind in revenue:
● Anti-fuses were not reprogrammable and not well
suited to datacom. This pushed Actel into other less
lucrative markets.
● Altera was flying high early, but it was in trouble by
late 1990s, because CPLDs could not keep up with
the increasing density of the FPGAs and the needs
of the newer datacom applications.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
11
Then came the bust in 2001…

The communications industry felt the bust
harder than any other sector.
 Countless datacom/telecom start-ups went
out of business.
● Demand for new datacom/telecom equipment nosedived.
● In parallel, FPGA revenues declined 40% within a
single year.
 And the middle ages started.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
12
In the middle ages…
 Xilinx kept designing FPGAs for the
datacom/telecom market first and foremost.
● Virtex was followed by: Virtex-2,4,5…

But they also developed lower cost derivatives for other
markets.
● Revenues declined initially, but they gradually
recovered.
● An era of slower growth began.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
13
In the middle ages…
 Altera made a successful transition from
CPLDs to FPGAs starting with Stratix and its
derivatives.
● Ironically, Stratix was aimed at the data-com market
of the boom years.

Other markets were addressed with derivative lower cost
products.
● Like Xilinx’s, Altera’s revenues also declined initially,
but gradually recovered.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
14
In the middle ages…
 Actel successfully transitioned itself from
making only anti-fuse FPGAs to flash based
reprogrammable FPGAs.
● These new devices were aimed at the low cost
markets as well as markets that put a premium on
design security.
● Revenues declined initially, but recovered quickly.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
15
Middle ages (2002-2008)


FPGAs have settled down to a pattern of healthy but
slower growth, with small ups and downs.
● Market leading FPGAs continue being designed for datacom,
but their low cost derivatives target all other markets
combined.
FPGAs no longer appear to be a very vibrant industry
with a lot of rapid innovation.
● That’s the price one pays for maturity?
● Take comfort: the ASICs are worse off.

12/12/2007, IC-FPT,
Kitakyushu, Japan
Dataquest estimates that ASICs shrunk by 4% in 2007.
Sinan Kaptanoglu, Actel Corp.
16
Which future awaits FPGAs?
 More of the same FPGAs?
● Similar architectures for everybody?
● Less innovation, more marketing?
 Good continued growth, but at a slower
pace…
OR
 Another boom: but what might start it?
 If history repeats, it will be a rapidly
growing new market for FPGAs!
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
17
A new market driver?
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
18
A new rapidly growing market
 We are witnessing a dramatic increase in
portable devices.
● These are usually battery operated.
 Not a temporary blip:
● More new devices and types of devices are coming
to market continuously.
● Expected to accelerate in the 2010s.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
19
Spectrum of portable devices
Consumer
Volume
Industrial, Medical
Military, Automotive
Features
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
20
Portable market predictions
 Gartner-Dataquest predicts 1.6B smart
phones by 2010.
 Juniper Research predicts 56M high end MP3
players by 2010.
 Industrial/Medical/Automotive sectors are
growing more rapidly than any other FPGA
market.
● Portable devices within these sectors are growing
the fastest.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
21
Gartner-Dataquest estimates
2002 Market - $2.3B
Mil/Aero
6.9%
2010 Market - $6.8B
Auto
0.6%
Mil/Aero
9.3%
Consumer
7.6%
Industrial
15.3%
Comm
53.8%
Consumer
14.6%
Auto
6.2%
Comm
37.4%
Data Proc
15.7%
New adopters are mostly
from the consumer and
the industrial sectors
12/12/2007, IC-FPT,
Kitakyushu, Japan
Industrial
24.2%
Sinan Kaptanoglu, Actel Corp.
Data Proc
8.2%
22
What’s the big deal about the portable
market?
 Possibly there is no big deal:
● Perhaps these are just yet another type of devices
which may use FPGAs sometimes.
● But we could have said the same thing about the
datacom market back in 1993.
 What similarities are there between the
datacom market in 1993 and the portable
device market in 2007?
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
23
Similarities are…


In 1993, datacom had very short design cycles and
preferred FPGAs.
● But it could not use FPGAs frequently because the FPGAs
often failed to meet minimum performance and density
requirements.
In 2007, portable device market has short design
cycles and prefer FPGAs.
● But it cannot use FPGAs frequently because the FPGAs often
fail to meet the minimum static power requirements.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
24
This market is different…
 Because:
● Most portables are battery powered.
● Portable devices spend most of their time awake,
idle, but doing very little.


Varies from 50% to 99.99% of the life.
This requires very low static power usage.
 Compare that to datacom:
● An FPGA in a router is active all the time.
● Idle is for repairs and updates only.
● A good FPGA for datacom is unlikely to be well
suited for portables and vice versa.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
25
Don’t the low-cost derivatives address
this market?
 The low cost derivatives of market leading
FPGAs indeed save area and power both.
 But the static power of these devices are
still two to three orders of magnitude too
large for most battery powered applications.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
26
The minimum requirements:
 The FPGAs for portable devices need a very
low power “Idle” state.
● Measured in tens of mW.
 Idle state does not mean “switched off” or
even deep “stand-by”:
● Idle means very light activity.
● Idle means that all FPGA flip-flops retain their state
without being saved to and restored from a memory.
● Switching from idle to active must be very quick in a
few clock cycles, not thousands.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
27
Can this be done?
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
28
An optimist’s view…
 If we design new FPGA architectures custom
tailored for the unique needs of the Portable
device market, we give ourselves the best
chance to witness a second boom in FPGAs.
● These new FPGAs will not replace today’s products.
They will co-exist with them to serve different
markets.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
29
A pessimist’s view…
 Most ICs don’t experience any boom. FPGAs
were lucky to have had one!
 We are unlikely to meet the idle power
requirements of the portable market very
soon. Therefore:
● Accept maturity, and be happy with the slower but
steady growth.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
30
OK, let’s be optimists for a
while…
How do we design FPGAs for
the portable market?
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
31
First learn from ASICs…
 Use all the circuit and process tricks that the
low-power ASICs have employed:
● For the lowest leakage power, start by using a
designated “low power” process from your favorite
FAB.
● Use high-k dielectrics for the gate, if available.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
32
Learn from ASICs:
 Use a multiple-VT process.
● Use highest VT devices wherever you can without
adversely affecting performance.
● For example, all configuration bits in an SRAM based
FPGA should be built out of highest VT transistors.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
33
More tricks from ASICs…
 Consider using a multi-Vdd process.
● There has been a wealth of research on multi-VT and
multi-Vdd circuit design for power reduction in ASICs.
● If a multi-Vdd process is not available, consider using
a reduced Vdd operation.
 Consider using long-channel devices.
 Consider power gating.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
34
More tricks from ASICs…
 Consider using a triple-oxide-process, if
available:
● Good for gate leakage.
● Extra mask, extra cost.
 If available, use a triple-well process to take
advantage of well biasing.
● Again good for leakage, but costly.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
35
And yet more…
 If you are really desperate, as a last resort,
consider stacked gates.
● Very costly (in area) for incommensurate returns in
leakage reduction.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
36
And then, get help from SW…
 Use all power-driven synthesis, tech-mapping,
and place & route.
● If you used well-biasing, or multiple VT, or multiple
Vdd, let place & route arrange it as best as it can (in
different parts of the FPGA) for the best power.
 At the end, do power based bit-mask
selection for configuration bits.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
37
When you have adopted most of these
tricks from ASICs…
 You have far exceeded the level of “low
power” derivatives of the current popular
FPGA architectures.
● You have definitely improved static power, but not
sufficiently.
● You kept the changes to a minimum so that you can
amortize the development effort for the original
architectures.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
38
This is good, but not enough…
 We need to explore fancier process
technologies, circuit design methods and
FPGA architectures.
● The new FPGAs must be fully custom tailored for the
portable market.

12/12/2007, IC-FPT,
Kitakyushu, Japan
Even if that means they will not be very good for datacom
applications.
Sinan Kaptanoglu, Actel Corp.
39
Promising technologies for FPGAs
specifically targeted to portable
devices
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
40
The trend is not good…
 CMOS process and portable devices seem to
be heading in opposite directions.
 The transistor leakage has increased with
every CMOS shrink to date.
● For source-drain and gate leakage both.
 As we move down the silicon dimensions the
leakage will only get worse.
● It’s just physics.
● Using the exact same type of materials, a shrink
will leak more than the previous generation, unless
the voltage is scaled.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
41
Vdd is not scaling!
 Most 130nm low-power CMOS processes use
1.2V core operation.
 So do most 90nm and 65nm low-power CMOS
processes.
 Most 45nm low-power processes will use
1.0V, or 0.9V at best.
● Even those which use high-k dielectric materials.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
42
This is not very likely to change
 Vdd will not scale very much in order to
maintain a modest performance and in order
to keep the gate leakage under control.
● With or without high-k materials.
 But the source-drain leakage will keep
increasing from each process node to the
next.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
43
Future technologies
Three examples of promising
technologies for future FPGAs for the
portable device market
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
44
Here is an incomplete list…


In increasing order of maturity:
1) SOIAS (Silicon-on-Insulator-with-an-ActiveSubstrate) technology.
2) Very high-k gate-oxide technology, with very small
gate leakage.
3) Embedded flash-switch technology.
It is possible to combine any two or even all
three.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
45
SIOAS technology
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
46
SOIAS based architectures
 Let’s start first with the basic SOI:
Metal-1
bulk-SiO2
Metal-1
gate
(poly/
metal)
n+
p-
gate oxide
n+
Insulator (Si02)
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
47
SOIAS adds a back-substrate
 From Yang et al. (1995, IEEE-IEDM)
Metal-1
bulk-SiO2
Metal-1
gate
(poly/
metal)
n+
p-
gate oxide
n+
Insulator (Si02)
p+
i-poly
Insulator (Si02)
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
48
Fine-grain dynamic back-gate biasing
 Fine grain back-bias:
● In space and time both.
● We can back bias individual logic modules, FFs or
individual routing buffers.
● Change biasing frequently.
● Order of magnitude reduction in static power
compared to low-power bulk-CMOS.
 Need a highly sophisticated power driven
synthesis, place and route.
● Even ASICs can’t do this very well yet.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
49
SOI uses less dynamic power
 Dynamic power is lower for any SOI due to
lower gate and junction capacitances.
 Many standard cell designs claim 50%
dynamic power reduction compared to bulk
CMOS (SOITECH claim.)
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
50
Very high-k gate oxides
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
51
Very high-k dielectric gate oxides
 Literature is full of promising new gate
insulators.
● Penryn is shipping with hafnium oxide.
● Other’s examples will follow suit soon.
● Feasibility is proven, but reaching maturity will take
time.
 These materials will allow us to play a onetime game with VT and Vdd.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
52
Ready for the next shrink?
 Suppose you are at a process node X, with
operating voltage V, threshold voltage T, and
gate oxide thickness G.
 Ideal shrink is at sX, sV, sT, and sG, where s
is the shrink factor.
● Often s = sqrt(0.5) = 0.7.
 All’s good here, except performance, which
is roughly proportional to (V-T)2, which now
becomes s2*(V-T)2 for the shrink.
2
● Your new performance is only s that of the old one,
which is unacceptable.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
53
Restoring the performance…
 To counter, you go with sX, pV, qT, and pO,
with q < s < p < 1
 Now the speed of the shrink will go like (pV –
qT)2.
● Choose (pV – qT) >= (V – T).
● For example V = 1.2, T = 0.4, p = 0.83, q = 0.5, and
V-T = 0.8 before and after.
 But now your source-drain leakage is
unacceptably bad!
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
54
Now fight the leakage…
 If you don’t have very high-k gate oxide, you
are out of luck. You have to live with the
worsened leakage.
 With a high-k gate oxide however, you can do
sX, V, T, rO, where r < 1.
 Now the performance is even better than
before, and the leakage (both kinds) are no
worse than before.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
55
Starting from low VT and high source-drain
leakage…
Reduce
sourcedrain
leakage
Increase
VT
12/12/2007, IC-FPT,
Kitakyushu, Japan
Increase
gate
leakage
Decrease
gate
leakage
Use
High-k
dielectric
Increase
Vdd
Reduce
Idsat
Increase
Idsat
Reduce
speed
Increase
speed
Sinan Kaptanoglu, Actel Corp.
56
Flash-switch technology
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
57
Flash-switch technology
 The most mature technology today for low
static power.
 Not the same as flash-memory-on-the-side
(FMOS) which loads SRAM configuration bits
in the FPGA.
● FMOS approach may make a good FPGA, but not a
good low-power FPGA.
● Most Lattice products and MAX-2 from Altera has
FMOS.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
58
Actel’s combined flash cell
Flash Cell
Flash-switch
NOR-Flash
memory-bit
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
59
Actel’s flash cell
 Very low leakage: Each switch leaks only a
few pico-amps!
 This technology combines the flash memory
bit and the flash-switch to a single very low
leakage cell.
● This cell is used directly in the signal path, not for
controlling SRAM bits.
● The technology is ideal for configuration, but the
leakage of the CMOS logic and routing buffers still
needs to be dealt with.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
60
Not just some futuristic technology
 It is shipping today in IGLOO™.
TM
 Actel IGLOO
is the lowest power family of
reprogrammable FPGAs in the market.
 Depending on the expected idle time, it can
extend the battery life by more than a factor
of 10 compared to any other FPGA in the
market!
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
61
Static power comparison
150 mW
160

140
Igloo has the lowest
static power.
●
100
80
mWatts
120

60 mW
60
40 mW
40
20
0
Its leakage in “flash-freeze”
state is 300 to 1000 times
better!
Igloo advantage comes
from flash switches in
signal path!
0.12 mW
Igloo
12/12/2007, IC-FPT,
Kitakyushu, Japan
Competitor "A"
Competitor "B"
Competitor "C"
Sinan Kaptanoglu, Actel Corp.
62
Battery life comparison at 95% idle
95% idle, 5% at 100MHz, Smart-phone application
low cost FlashSRAM hybrid
low cost SRAM
fpga
low power CPLD
anti-fuse fpga
Igloo
0
500
1000 1500 2000 2500 3000 3500 4000 4500
Hours of Battery life
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
63
Are we already there then?
 We are indeed very close, IGLOO™ is a great
product for portable devices.
● But we can do even better. A future FPGA using an
order of magnitude less “idle” power than IGLOO™
is possible.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
64
A promising circuit design technique
On-chip low-swing differential signaling
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
65
Low-swing differential signaling
 Lowest dynamic and static power!
● But costly in silicon area!

One needs sense-amps at every input of the logic
modules and FFs.
● Also costly in development effort.
 Immunity to noise and immunity to SEU is
unproven.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
66
Low-swing differential signaling
 Hybrids are possible:
● Some regular routing resources, plus some (mostly)
differential signal routing resources.

Let a smart place and route make the right choices for the
power and performance.
● Can also be combined with SOIAS and flash-cell
technologies.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
67
Examples of promising
architectural changes
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
68
Overall architecture optimization
 Current FPGAs are architected to minimize
[(Area)a * (Delay)d].
● The exponents a and d vary by a factor of two: 0.7 <
a, d < 1.5.
 Instead, we need to minimize:
[(Area)a * (Delay)d / (Battery-life)b]
● With 0.7 < a, d < 1.0 < b <= 2
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
69
Current architectures are very heavily
DSP-oriented
 Market leading low cost FPGAs come with
many hard multipliers.
 This makes good sense in many markets but
not for portable devices:
● Very few portable apps will benefit from a hundred
18X18 multipliers on FPGA.
● For most portable applications just a few multipliers
are enough.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
70
Ratio of used routing buffers
 FPGAs use unidirectional buffers nowadays.
 For a typical design, more than half of these
remain unused after place and route.
● They contribute to leakage in “Idle”.
 It is possible to redesign the interconnect
architecture such that only 20% of the
buffers remain unused for typical designs.
● Then the total number of buffers can be reduced,
with a corresponding decrease in leakage.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
71
A hand-waving explanation
Ratio of used routing buffers
 Doing this without increasing the mux area
very much and without reducing routability
is the key!
● Different interconnect topology is needed!
● Changes are not simple.
● There is a performance penalty, but it could be kept
well below 10%.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
73
Ratio of used carry chains
 In a typical design 70% to 90% of all ripple
carry chains are unused!
● Carry chain drivers are large by necessity, and they
leak a lot.
● The carry chain can be redesigned to reduce the
leakage by up to 50%.

Performance loss can be kept in single digits.
● Overall impact on total leakage is not that big, but
every little bit helps.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
74
Dual port RAMs
 Dual port RAMs are popular.
● Especially useful to build FIFOs.
● For the argument here “dual port” and “two port”
amount to the same thing.
 However, the dual port RAMs are used as
single port RAMs more than 90% of the
time.
● Except in datacom applications!
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
75
Dual port RAMs
 A dual port RAM used in single port mode
has half of its outputs tied OFF.
● These will leak in the unused state.
 Dual port RAM cells have two or more
additional transistors, which increase
leakage.
 For FPGAs for portable devices, it is better to
provide mostly single port RAMs, and some
dual port RAMs.
● Significant help with leakage.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
76
IO leakage…
 Unused IO drivers do not leak much because
of the thicker gate oxide.
 But there is a huge amount of other circuitry
in front of the pad drivers.
● These leak a lot more, especially because the FPGA
IOs are designed to support many different IO
standards.
 By circuit redesign and by supporting fewer
IO standards the leakage can be reduced
significantly.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
77
Don’t forget the peripherals…
 One rarely worries about PLLs, crystal
oscillators, other clock generators, voltage
pumps, etc...
● They don’t contribute much to the dynamic power,
but they are a major component of the “idle” static
power.
 These mostly analog blocks should be
redesigned for FPGAs serving portable
applications with much reduced idle power
consumption.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
78
Now tackle dynamic power
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
79
Next tackle the dynamic power:
 Even though not as critical as static idle
power, dynamic power is still very important
for battery operated apps.
 We can improve dynamic power by doing
similar things as ASICs.
● All changes combined, we can improve it up to a
factor of two or so.
● Unlike the static power, there is no way to improve it
by several orders of magnitude.

12/12/2007, IC-FPT,
Kitakyushu, Japan
The only way to improve it by an order of magnitude is by
low-swing differential signaling.
Sinan Kaptanoglu, Actel Corp.
80
FPGA specific optimizations…
 For dynamic power, FPGA specific
optimization involves primarily IOs and clock
networks.
● Other parts of FPGAs are optimized similarly to
ASICs.
 IOs use 30% to 50% of dynamic power.
● For IOs, the most effective strategy is integration:
 Low voltage (1.2V) IOs are preferred.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
81
Clock architectures will change
 In current architectures 20% to 50% of the
core power is burned by clocks!
● This is 3 to 10 times larger than ASICs.
 This is a relic of the mid-1990s design:
● FPGAs come with 6 to 24 global distribution
networks, because this was the only reliable way of
distributing low-skew signals on FPGAs in 1990s.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
82
It even affected FPGA synthesis…
 FPGA synthesis generates (relatively) few
clocks, with many clock enables.
● This is a very nice and clean methodology.
● Very friendly for synchronous design.
● But it wastes power!
 We need FPGAs that can support many
(hundreds) small clocks.
● No support for explicit ENA on flip-flops.
● That’s what ASICs have been doing!
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
83
Clocks will make the switch first
 It may take a long time for the first FPGA
using all low-swing differential signaling to
come to market.
● Way before then, however, we will have FPGAs that
use differential signaling in global clock networks.
● It’s the lowest hanging fruit.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
84
Summary
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
85
The second FPGA boom may be on its
way…
 The rapidly growing portable device market
is the best candidate that may start the
second FPGA boom.
 But the FPGAs must meet the static power
requirements of this market.
● We can do this by designing new FPGAs specifically
for portable devices, rather than trying to retrofit old
ones.
● The old FPGAs will not go away, they continue to
serve markets which do not care very much about
battery life.
12/12/2007, IC-FPT,
Kitakyushu, Japan
Sinan Kaptanoglu, Actel Corp.
86
Thank you!