( ESNUG 482 Item 3 ) -------------------------------------------- [06/30/09]
From: [ Count Chocula ]
Subject: Mentor Calibre nmDRC benchmarks 6X faster than Synopsys Hercules
Hi, John,
I've been using Mentor's Calibre nmDRC and their LVS tools since late '07.
My company has multiple internal design groups around the world. Usually we
do the physical design, but in this instance a far group had responsibility
for the physical design and we were the interface to the fab. The far group
used Synopsys Hercules DRC/LVS, while we used Mentor Calibre nmDRC/LVS.
This gave us with a unique opportunity to compare Calibre vs. Hercules.
The far group was building a 35+ million gate chip and were running their
entire chip at once through DRC. The Hercules runtimes were extremely slow
(48 to 52 hours), and its results took a long time to decipher. We ran the
same design on Calibre, and it was 6x faster (details below).
Although it took 5 hours to upload and download the far group's GDS data via
FTP, it was still much faster for us to use Calibre and give the far group a
log file than for them to work with Hercules directly.
Here's our Calibre with HyperScaling benchmark on our 35+ M gate, 90 nm ASIC
chip with ~2.0 M flip-flops.
Tool Runtime CPUs
------- ------------ -----
Hercules ~50 hours 1
Calibre 8.5 hours 8
Calibre 11.6 hours 4
You can see Calibre scales well with more CPUs. We ran it on an Intel Quad
core 2.66 GHz Clovertown processor with SPECint of 16. The far group ran
Hercules on a similar machine.
Hercules Debugging Nightmare:
As I said before, the far group was also having problems deciphering the
results from the Hercules DRC log files. The Hercules error stepping tool
could not handle non-rectilinear blocks, and in those cases, it would step
to the wrong location. This meant the far group had to manually open a text
window, cut-and-paste from the Hercules log file, then use the GUI to jump
to the correct x, y coordinate. Given that a typical run has 1500 errors,
this made a simple process very tedious.
Since Calibre recognized non-rectilinear blocks, I could step to each error
and its location in ~1 second. I ended up running Calibre DRC, grabbing
the screen shots for the errors, and putting them in PowerPoint. For
example, "Metal 5 violation here, please move this wire". We were able to
successfully identify DRC errors this way in less time than what they were
doing with Hercules.
Another thing I liked about Calibre was that after I first ran DRC and got
the error list, I could create an initial waiver file that could be used
incrementally. Then when I ran new design versions through DRC, the
waivered errors were displayed in gray so I didn't have to wade through
them all over again. One deficiency here was that Calibre did not have a
way to hide previously waived errors vs. all errors.
The best part of Calibre DRC is its GDS viewer. It reminds me of the "mouse
gestures" in Firefox. You can move the mouse in one direction to zoom in,
then move it in the other direction to zoom out.
Finding shorts with Calibre LVS:
I also had an opportunity to work with Calibre LVS during this project. I
liked its Short Finder the best. It locates when two nets that are shorted
together in your GDS, such as when your floorplanner or router crosses two
signals on top of each other.
Before Calibre, I would need to manually look at two netlists without layout
information. I would highlight the two nets and scroll around manually to
find the overlap. For example, for a power to ground short, I would
highlight the power net and highlight the ground net. The whole chip would
light up, making it very hard to find the short! It took an inordinate
amount of time to find it -- sometimes up to two days for one bloody short!
By running the Short Finder, I typically find the shorts in under an hour.
Calibre integrated its GDS viewer in to make it easier to find the shorts.
It finds the two nets in GDS and automatically searches for the most likely
places of intersection, then zooms into that location so you can see it.
Calibre automatically breaks up each segment and assigns a name to it.
This is especially useful for large nets, where finding shorts is a needle
in a haystack problem. This is especially true with gigantic power/ground
meshes, and you must otherwise scroll to see every connection.
I only took the crash course on Calibre, yet the first time I ran Calibre
LVS I was up and running with real results in 1 week for our full chip. It
only required a minimal configuration on my part to extract a good GDS
netlist.
In contrast, a colleague using Hercules needed 3 weeks to get going the
same design.
Please keep me and my company anonymous on this. I don't want the Synopsys
marketing people hounding me.
- [ Count Chocula ]
Editor's Note: I asked Chocula why Hercules was single CPU, he replied:
"The far group did not run Hercules multi-CPU because they did not have
multiple-license use of Hercules. Calibre allows multiple CPUs to be
run by calling additional Calibre licenses, at a ratio of one license
per every 4 CPUs." Sounds like Herc use is more restrictive. - John
Join
Index
Next->Item
|
|
