Professional Documents
Culture Documents
Prof. MacDonald
Course Overview
! Intro to CMOS circuit design
– standard cell libraries
! combinational
! sequential
! dynamic
– semiconductor memories
! DRAM
! SRAM
! ROMs
– IOs
– Clock trees
– Power Grids
– Phase-Lock Loops (Introduction to)
Course Overview
Flip-Flop
Flip-Flop Layout
Chip with circuit rows
Silicon Art – Slide 2 – UTEP Art
CMOS – Historical Perspective
! the 70’s: NMOS primary technology
– CMOS becomes usable in late 70’s
! CMOS doesn’t consume power when idle – battery applications
! calculators – slide rules disappear
! digital wrist watches
! the 80’s: CMOS dominates and the PC is born
! 1985: accusations of Japanese chip-dumping
– Japanese extend existing photolithographic technology
– US migrates to next generation tools and can’t get it working
! the 90’s: CMOS gets smaller, faster, more integrated
! the 00’s: CMOS scaling continues but leakage
emerges
CMOS – Moore’s Law
! Gordon Moore – co-founder of Intel
! Empirical observation
– advances in semiconductors double every 18 months
– transistors get smaller
– speed increases
– transistor count increases
– self-fulfilling prophecy
– doesn’t hold true for memories – thus caching important
– every year since, the end CMOS scaling 5 years off
– current challenge is gate oxide leakage with Tox below 15 A
! static power set to overtake dynamic power consumption
! lack of static power was original advantage of CMOS over Bipolar
! gate oxides so thin (10 atomic layers) tunneling occurs
CMOS – Moore’s Law – channel length
100
10
Japanese DRAM crisis
Leffective(microns)
my career started
0.1
working on 0.8u DSPs
at Motorola
0.01
1960 1970 1980 1990 2000 2010 2020
Year
CMOS – Moore’s Law - transistors
100,000,000
10,000,000
1,000,000
100,000
10,000
1,000
1965 1970 1975 1980 1985 1990 1995 2000 2005
CMOS – Moore’s Law - speed
35000
30000
25000
20000
15000
10000
5000
0
1990 1995 2000 2005 2010 2015 2020
CMOS – Moore’s Law – DRAMs
100000
10000
1000
DRAM Mbits
1 Gbit DRAM
16 Mbit DRAM
10
1
1990 1995 2000 2005 2010 2015 2020
PPC405LP
Full Custom CMOS design
! Only done when performance is king
! Unit cost is medium - $50 a chip
! Performance is best – 3 GHz processor
! Development cost (one time cost) is very high
– 1000 engineers working for 2 years and costed at $200k/year
– $400 million dollar development cost – wow!
– turn around and sell 100 million chips for $150 each
– $15 billion income – wow! development cost insignificant
– moral of the story:
! get a lot of engineers to hand tweak design for max performance
! but you better sell a lot of them…
ASIC Chips (AKA Semi-custom)
! Mostly logic design in Verilog and VHDL
! Little transistor-level work
! Synthesis/Place/Route cells from logic library - tools
! Library includes
– standard cells (i.e. ANDs, OR’s, Flip Flops)
– memories
– IO blocks (input and output of chip)
– Phase-Lock Loops
– probably developed in another company (i.e. ARM-Artisan)
! Fabricate in a foundry (i.e. TSMC, IBM, AMI, TI)
! 10 to 50 engineers in one year to design chip
ASIC Chips
! Good trade-off between performance / price
! Unit Cost low - $5 to $10
! Performance medium – 400 MHz
! Development cost medium
– 50 engineers for one year at $200k/year = $10 million
– sell 10 million graphics cards - chip adds $10 of value
– $100 million income (minus cost of fabrication)
! Good approach for system level manufacturers
– Cisco, NVidia, Nokia
– Let someone else fabricate it – Fab is $5B capital expense
– need chip to differentiate product
! router, cell phone, graphics card
FPGA Chips
! Greatest Flexibility – update chips in customer’s hands
! Performance low – 100 MHz
! Unit cost High - $50 to $200 per chip
– good for low volumes only – say less than 10,000
– can start with FPGAs and migrate to ASIC later
! Development cost low
– 2 engineers for 6 months at $200k/year = $200,000
– no mask charges which are now approaching $1M for ASICs
! Unit cost are getting better
– some say FPGAs will soon take over the world