Transcript No Slide Title

TKT-9627 Digital and Computer
Systems Seminar
Closing the Gap Between
ASIC & Custom
JET 2006
TKT-9627 Digital and Computer Systems Seminar
1
Course content

Literature:
David Chinney and Kurt Keutzer:
“Closing the Gap Between ASIC & Custom”
Kluwer 2002, 407 p., ISBN 1402071132
–

Topics
–
–
–
–
–
–
–
JET 2006
Tools and Techniques for High-Performance ASIC Design
Improving performance through microarchitecture
Timing-driven floorplanning
Controlling and exploiting clock skew
High performance latch-based design in an ASIC methodology
Automatically identifying and synthesizing complex logic gates
Automatic cell sizing to increase performance and reduce power
Controlling process variation
TKT-9627 Digital and Computer Systems Seminar
2
Course Info and Requirements

Time and place: Wed 14-16, TC165
Instructors: Jouni Tomberg and Olli Vainio

Requirements:

–
–
–
JET 2006
Presentation of the course book topic
Active participation (>60%)
Exam
TKT-9627 Digital and Computer Systems Seminar
3
Table of Contents (Book chapters)
1.
2.
3.
4.
5.
JET 2006
Introduction and Overview of the Book (32 p.)
Contributing Factors
Improving Performance through Microarchitecture (24 p.)
Reducing the Timing Overhead (44 p.)
High Speed Logic, Circuits, Libraries and Layout (44 p.)
Finding Peak Performance in a Process (24 p.)
TKT-9627 Digital and Computer Systems Seminar
4
Table of Contents …
Design Techniques
6.
Physical Prototyping Plans for High Performance (19 p.)
7.
Automatic Replacement of Flip-Flops by Latches in ASICs (22 p.)
8.
Useful-Skew Clock Synthesis Boosts ASIC Performance (16 p.)
9.
Faster and Lower Power Cell-Based Designs with Transistor-Level Cell Sizing
(16 p.)
10. Design Optimization with Automated Flex-Cell Creation (28 p.)
11. Exploiting Structure and Managing Wires to Increase Density and Performance
(20 p.)
12. Semi-Custom Methods in a High-Performance Microprocessor Design (16 p.)
13. Controlling Uncertainty in High Frequency Designs (18 p.)
14. Increasing Circuit Performnace through Statistical Design Techniques (22 p.)
JET 2006
TKT-9627 Digital and Computer Systems Seminar
5
Table of Contents …
Design Examples
15. Achieving 550MHz in a Standard Cell ASIC Methodolgy (16 p.)
16. The iCORE 520MHz Synthesizable CPU Core (22 p.)
17. Creating Synthesizable ARM Processor with Near Custom
Performance (25 p.)
JET 2006
TKT-9627 Digital and Computer Systems Seminar
6
Introduction (Book chapter 1)

Why are custom circuits so much faster?
–
–
–

Who should care?
–
–
–

ASIC and ASSP designers seeking high performance
Custom designers seeking higher productivity
EDA tool developers and researchers
Definitions in this book
–
–
JET 2006
Routinely 3x to 8x faster when fabricated in the same same process
generation.
The first aim is to explain this disparity in performance.
The second aim is to understand practical ways in which the performance
gap can be bridged.
ASIC design methodology, standard cell library based design, netlist handoff
Custom-design methodology, layout handoff
TKT-9627 Digital and Computer Systems Seminar
7
A Quick Comparison …
•Process technologies
vary in a number of
ways: the channel
legth, interconnect
density and material,
etc.
JET 2006
TKT-9627 Digital and Computer Systems Seminar
8
…Based on These Figures
•Most of the custom designs
use dynamic logic on critical
paths and more pipeline stages
to achieve higher speeds.
JET 2006
TKT-9627 Digital and Computer Systems Seminar
9
Adding Up the Numbers
JET 2006
TKT-9627 Digital and Computer Systems Seminar
10
Improving Areas







JET 2006
Microarchitecture
Timing overhead: clock tree design and registers
Logic style
Logic design
Cell design and wire sizing
Layout: Floorplanning and placement to manage wires
Process variation and improvement
TKT-9627 Digital and Computer Systems Seminar
11
Microarchitecture (1)


Organization of functional units; number, hierarchy, interfaces, pipeline, number
of computational clock cycles, logic for branch prediction and data forwarding
What’s the problem?
–
–
–
–
–
–
JET 2006
Pipelining in ASICs is limited by the larger timing overhead for the registers in the
pipeline.
Custom designs may show superior logic-level design of regular structures such as
adders, multipliers etc. Achieving fewer levels of logic and combining logic with
registers.
In custom designs the pipeline stage balancing can be done more effectively.
ASIC microprocessors tend to have simpler implementations of speculative
executions to reduce the design time.
It is difficult to estimate the precise performance improvement with
microarchitectural changes.
The overheads for pipelining are the register delays, larger impact of clock skew and
clock jitter and unbalanced pipeline stages leading to about 30% overhead for an
ASIC design.
TKT-9627 Digital and Computer Systems Seminar
12
Microarchitecture (2)

What can be done?
–
–

In the book
–
–
JET 2006
Use the best pipelining practices .
ASICs are unable to have the same tight control of the combinational delay
and the timing overhheads. Thus the estimation is a factor of 1.3x between
best ASIC and custom implementations.
Chapter 2 provides a tutorial intro to microarchitecture of ASICs
The design examples in chapters 15-17 give examples of developing
efficient microarchitecture for ASICs
TKT-9627 Digital and Computer Systems Seminar
13
Timing Overhead: Clock Tree Design and
Registers (1)

The timing overhead is the additional delay associated with pipeline registers and the
arrival of the clock edge.
–

What’s the problem?
–
–
–
–
–
JET 2006
It consists of the delay through the registers, the setup time of the registers, the clock skew
between arrival of the same clock edge at the different points of the chip and the jitter between
the arrival of the consecutive clock edges at the same point of the chip.
ASICs have larger clock skew and clock jitter and slower registers than custom designs.
ASIC designers try to avoid races and increase tolerance to noise, as they have far less control
of variation. Thus they design the circuitry to work for the range of possible conditions.
ASICs primarly use edge-triggered flip-flops and high speed registers are not supported in the
cell libraries.
Tight control of the layout allows hold time violations and noise to be avoided. Also custom
designs may run long wires with shielding wires or use low-swing signaling to reduce the
effects of noise.
ASICs don’t use level-sensitive latches because there is a larger window for hold time
violations. ASICs can not use multi-phase clocking schemes with time borrowing in skew
tolerant domino logic for a variety of tool-flow related reasons.
TKT-9627 Digital and Computer Systems Seminar
14
Timing Overhead: Clock Tree Design and
Registers (2)

What can be done?
–
–
–
–
–
–
–
JET 2006
Current EDA tools can verify that hold times are not violated and insert delay
elements to avoid races on short paths.
Slack passing is possible with level-sensitive latches or cycle stealing, by carefully
scheduling of the arrival of the clock edges at different registers.
Multi-phase clocking is not a viable solution in the deep submicron, because of
signal integrity issues and the increasing difficulty of distributing several clocks
across the chip.
As ASICs have large clock skew, latches have substantial benefits for reducing the
clock period.
Level sensitive latches reduce the impact of inaccuracy of wire load models and
process variation. The clock period is not limited by the delay of the slowest
pipelining stage, because of slack passing.
The clock skew can be reduced by using better clock tree synthesis tools or manually.
Latches and typical ASIC clocking techniques are slower than custom techniques
leading still to 1.1x faster custom design.
TKT-9627 Digital and Computer Systems Seminar
15
Timing Overhead: Clock Tree Design and
Registers (3)

In the book
–
–
–
–
JET 2006
Clock related timing issues are tutorially reviewed in chapter 3.
Chapter 7 describes a prototype tool that automatically converts a gate netlist
with flip-flops to use latches leading to 10-20% speed improvements. It also
discusses the timing overhead in the Xtensa mircoprocessor.
Chapter 8 considers issues associated with clock-tree synthesis and use of
carefully adjusted clock skew in clock-tree design for cycle stealing.
Chapter 15 discusses a very high speed ASIC design, where reduction of the
timing overhead was essential. Clock trees were manually routed and both
latches and high speed pulsed flip-flops were used.
TKT-9627 Digital and Computer Systems Seminar
16
Logic Style (1)

Dynamic logic can be used to speed up critical paths within the
circuitry by reducing gate delays.
–

What’s the problem?
–
–
–
JET 2006
It significantly faster than static CMOS logic and has smaller area, but is
much more sensitive to noise and consumes more power.
To be useful in an ASIC design methodology a logic style must be robust in
a variety of circuit conditions, and supported by tools for static timing
analysis and manufacturing testing.
Dynamic logic requires careful design of the power and clock distribution.
For these reasons dynamic logic libraries are not available it is not supported
for ASIC designs.
TKT-9627 Digital and Computer Systems Seminar
17
Logic Style (2)

What can be done?
–
–

In the book
–
JET 2006
There has been some progress in dynamic logic circuit synthesis, but it is not
yet commercially available and it is not likely that the methodological
obstacles will be overcome to enable dynamic logic synthesis in ASIC
designs.
However, the gap between static CMOS and dynamic logic can be reduced
by using custom designed static logic with pulsed inputs. In this case the
dynamic logic may only be 1.2x faster than highly optimized static logic.
Chapter 4 quantifies the performance improvement by using dynamic logic
and includes an example of a high speed 64 bit adder using static logic. Also
the limitations on future use of domino logic is examined and some
alternative high speed logic styles are discussed.
TKT-9627 Digital and Computer Systems Seminar
18
Logic Design (1)

Logic design describes the topology or interconnectivity of the
gates.
–

What’s the problem?
–
–
JET 2006
It determines the choice of adder algorithm (e.g. ripple carry vs. carry lookahead). Instead, the choice to pipeline the multiplier is a microarchitectural
decision.
For random logic both ASIC and custom designers are likely to use logic
synthesis but ASIC designers are not typically as aware of logical design
alternatives as custom designers.
For example, Wallace tree multipliers have a triangular layout – if not
carefully constrained, layout tools will try and fit gates to a rectangular
region, which is sub-optimal and leads to longer wire lengths.
TKT-9627 Digital and Computer Systems Seminar
19
Logic Design (2)

What can be done?
–
–

In the book
–
JET 2006
For datapath designs use the existing pre-designed libraries (e.g.
DesignWare) to get the optimized implementation.
ASIC logic design can come to parity with custom logic design.
Chapter 4 considers logic design with an example of a 64 bit adder and looks
for relative performance impact of good logic design.
TKT-9627 Digital and Computer Systems Seminar
20
Cell Design and Wire Sizing (1)


In ideal circuit, each gate is optimally crafted from transistors and
each transistor and wire is individually sized to meet the drive
requirements.
What’s the problem?
–
–
–
–
JET 2006
One element of the performance degradation of ASIC design is the poverty
of standard cell libraries including limited number of discretely-sized cells.
Many ASICs still not use good standard cell libraries with varied drive
strengths.
A richer library also reduces circuit area.
Custom designs can achieve a factor of 1.4x speed improvement over poor
ASIC design.
TKT-9627 Digital and Computer Systems Seminar
21
Cell Design and Wire Sizing (2)

What can be done?
–
–
–
–

In this book
–
–
–
–
JET 2006
ASIC designs should use rich standard cell libraries with dual gate polarities and
several drive strengths for each gate.
Several tools are available for automating the creation of cells that are optimized for
a design.
Tools of optimizing the widths of individual wires are currently not commercially
available.
By using continuous cell sizing and by using wire sizing custom designs can achieve
speeds about 1.1x faster than the best ASIC designs.
Chapter 4 gives a tutorial overview of issues in libraries.
Chapter 9 looks at commercial prototype tool for providing continuous cell sizing for
an ASIC flow.
Chapter 10 considers automatically finding macro-cells in logic.
Chapter 12 introduces a cell-sizing tool for custom designs.
TKT-9627 Digital and Computer Systems Seminar
22
Layout: Floorplanning and Placement to
Manage Wires (1)

Wire delays associated with global wires between physical modules are
dominant in the sub-micron technologies.
–
–
–

What’s the problem?
–
–
–
–
JET 2006
The primary factor in the wire delay is wire length.
Routing congestion and the position of cells in the layout affect the wire length.
Noise and cross-coupling capacitance between wires must also be considered.
Traditionally the load of gate drives is estimated using “wire-load models” that
estimates the capacitive load as function of block size and fanout.
This is a poor and inaccurate method in wire dominating delays.
Both over- and underestimating the wire loads leads to poor timing.
Carefully partitioned design can increase the speed by factor 1.4x compared to an
ASIC design that has large blocks of gates and uses inaccurate wire load models in
synthesis.
TKT-9627 Digital and Computer Systems Seminar
23
Layout: Floorplanning and Placement to
Manage Wires (2)

What can be done?
–
–
–
–

In the book
–
–
–
JET 2006
Accurate wire-load models, good design partitioning, careful floorplanning
and resynthesis should improve the speed.
Custom ICs are typically manually floorplanned.
Physical synthesis (e.g. Physical Compiler) is one tool improve the
placement.
ASIC designs should be able to achieve parity with custom designs with
respect to floorplanning and detailed placement.
Chapter 4 discusses the impact of wire-load models, partitioning and layout
tools.
Chapter 6 describes floorplanning techniques for ASICs.
Chapter 11 gives an example where manual layout improves the
performance.
TKT-9627 Digital and Computer Systems Seminar
24
Process Variation and Improvement (1)

Traditionally, the semiconductor process is represented as being fully determined
by a given technology generation, identical implementations, different plants and
at different times.
–
–



In the same nominal process technology, the speed varies significantly for the
same ASIC design synthesized to different libraries and different plant processes.
For custom ICs, additional improvements to the process or the design are
possible.
The semiconductor process can not be perfectly controlled, which leads to
statistical variation of many process variables.
–
–
JET 2006
The semiconductor process is also described through a set of worst-case numbers that
abstract the complexity of the actual manufacturing.
ASICs fabricated on a typical process can be 60-70% faster than the worst case
speeds quoted by ASIC library estimates.
Batch-to-batch, wafer-to-wafer, die-to-die and intra-die
This variation decreases as the process matures.
TKT-9627 Digital and Computer Systems Seminar
25
Process Variation and Improvement (2)

What’s the problem?
–
–
–

What can be done?
–
JET 2006
The ASIC design and manufacturing industry is organized around a
“handoff” point. Golden toolset and libraries guarantee for the customer the
product speed.
This traditional approach to worst-case timing analysis leaves significant
performance potential of semiconductor processes unharvested.
Sticking to worst-case process numbers and lacking the ability to exploit the
process improvements will lead to 1.6x – 2.2x speed difference between
ASICs and custom ICs.
As a result of relying on pre-characterized cell libraries, ASICs are typically
easy to migrate between technology generations. Thus synthesizable ASICs
can be easily switched to use the best fabrication plants available for ASIC
production. The custom ICs are more technology dependent.
TKT-9627 Digital and Computer Systems Seminar
26
Process Variation and Improvement (3)

What else can be done?
–
–
–

In the book
–
–
–
JET 2006
New ASIC libraries should be available from the ASIC vendor when process
improvements are done.
At-speed testing to harvest the worst-case ASICs are not usually available
from the ASIC vendors (commercial reasons).
Due to process variation ASICs will lose a factor of 1.2x performance to
custom designs.
Chapter 5 gives a tutorial overview of process related issues.
Chapter 13 describes a very practical way to manage the impact of process
variation.
Chapter 14 gives more speculative approach to exploiting processing.
TKT-9627 Digital and Computer Systems Seminar
27
Summary and Conclusions

There is a significant (3x to 8x) performance difference between
ASICs and custom ICs.
–
–
–

JET 2006
Influence of the factors of floorplanning and circuit design, while significant,
are relatively overstated in their importance.
The two factors of equal or greater significance are pipelining and process
variation.
The use of dynamic-logic families is a third significant influence.
ASIC designers must become familiar with microarchitecture,
physical design, clocking schemes, and sources of semiconductor
process variation.
TKT-9627 Digital and Computer Systems Seminar
28