( ESNUG 464 Item 13 ) ------------------------------------------- [03/30/07]
Subject: ( ESNUG 458 #7 ) Synopsys response to the CCS vs. ECSM discussion
> It has a big impact on disk space. A 400 cell single corner Liberty lib
> with 7x7 tables for timing and power data is usually around 4 MB. It
> grows to be in the range of 100 MB with ECSM data and 300 MB with CCS
> data. CCS takes approximately 3 times as much storage as ECSM.
>
> - Mehmet Cirit
> Lib Tech, Inc. Saratoga, CA
From: Richard Trihy <richard.trihy=user domain=synopsys not palm>
Hi John,
Aren't you glad storage is dirt cheap nowadays? When NLDM models were
first introduced 10 years ago, hard disk drive capacity was measured in
dozens of Megabytes and cost hundreds of dollars. Today, for the same
price, you can probably buy a one Terabyte hard disk drive. So even if
current-based modeling increases your library file size by 100X it's still
costing you about the same to store it. ;-)
On a more serious note, I want to give you the skinny on the Synopsys CCS
and Cadence ECSM files size from an engineering standpoint.
Both CCS and ECSM timing models model the transient behavior of the driver
waveform into a load. The key difference between them is that CCS consists
of current samples as a function of time, while ECSM consists of voltage
samples as a function of time.
Theoretically, both models have to be larger in size than NLDM. Typical
CCS libraries are about 8-10x larger than NLDM.
The reason for this is that where NLDM stores 2 numbers (delay and slew)
for each input slew and output load, ECSM and CCS both need to store
waveforms. CCS stores current and time samples. CCS characterization
employs adaptive algorithms to ensure that the sampled waveform provides
the most accurate model of the current vector. This means that for each
waveform, current and time values are saved. On average we see about 10
samples for each waveform. This means storing 10 current values and 10
time values for each waveform, so 20 numbers total. A simple rule of
thumb then is that where CCS has 20 numbers for each input slew and
output load, NLDM stores just 2 numbers. Hence we expect about a 10X
increase in the size of a new CCS timing model compared to an old NLDM
timing model.
The table below shows some example standard cell libraries along with
file sizes of the NLDM (.lib) models and their corresponding file size
when CCS data is added:
Library Tech # of cells NLDM file size NLDM + CCS
A 90 nm 804 19.0 M 168 M
B 90 nm 593 35.0 M 241 M
C 65 nm 330 3.7 M 42 M
D 65 nm 712 28.0 M 331 M
Synopsys has made available detailed characterization guidelines and
jumpstart scripts to allow for CCS models to be extracted and written to
Liberty. Following these guidelines ensures that the models deliver the
best accuracy with the least amount of data.
The open source scripts that we provide to customers show the mechanism
by which the current is sampled to preserve accuracy in both current and
voltage. We have found that fewer samples are needed at the beginning of
the transition, where miller glitches and overshooting is sometimes
seen. The segmentation algorithm skips over these glitches as this
portion of the waveform is not required to accurately model the drivers.
The extra delay that such glitches introduce is naturally captured in
the delay through the cell, represented in the difference between the
reference time and the delay threshold transition of the output signal.
The CCS segmentation algorithm which extracts current samples adaptively
selects the current while keeping the corresponding resurrected voltage
waveform accurate with respect to that predicted by the circuit simulator.
There are some differences between CCS and ECSM. For accuracy modern
delay calculation engines use current to calculate the effective delays
in the parasitic environment. So EDA tools can use the current values
directly from CCS models whereas they would have to derive current from
voltage when ECSM is used, leading to inaccuracy.
Furthermore we have found that accurately capturing the peak current in
the waveform is essential to delay calculation accuracy at 65 nm and
below. Storing voltage samples as ECSM does, makes it very difficult to
recreate an accurate current waveform.
Another key difference between ECSM and CCS is that CCS uses a 2 segment
receiver capacitor model. Having 2 segments allows us to have the
flexibility to model the Miller Effect as well as the loading of the gate
on its driver.
The feedback that we have received from foundries, library vendors and
IDMs point to CCS's superior accuracy for voltage and temperature
scaling. By storing the non-linear current waveform, scaling can be
performed in the current domain and then the scaled current waveforms
are used to accurately predict timing behavior.
Finally I want to let the readers know about a new extension to CCS that
we are calling "Base Curve Technology". This technology was recently
approved by the Si2 Liberty Technical Advisory Board. Base curve
technology allows the CCS data to be represented in a much more compact
form in Liberty. Without any loss in accuracy the CCS libraries are up
to 3X smaller when base curve technology is utilized.
- Richard Trihy
Synopsys, Inc. Mountain View, CA
Index
Next->Item
|
|
