( ESNUG 485 Item 6 ) -------------------------------------------- [05/27/10]
From: [ The LOST Smoke Monster ]
Subject: False & missed errors trouble Quartz in user benchmark vs. Calibre
Hi, John,
Due to crazy inter-company politics I have to be anon.
We have both Mentor Calibre 2009.1_35.24 and Magma Quartz 2010.2.1. I ran
in-depth comparisons of the two tools over TSMC 65 nm LP ver 1.4b and 2.0.
In this benchmark we used two chips:
Dr. Jekyll ~975 Mb ~500 K placeable instances
Mr. Hyde ~1.4 Gb ~1 M placeable instances
CALIBRE VS QUARTZ RUNTIME COMPARISONS
I concluded from my benchmarks that Calibre and Quartz were more or less
runtime equivalents. Based on the numbers below, Calibre was technically
10% faster overall, but this was not a noteworthy difference.
Calibre runtime relative to Quartz:
TSMC Version 1.4b # of Calibre Quartz Calibre's
Chip / GDS Size CPUs Runtime Runtime rel. speed
--------------------------- ---- ------- ------- ----------
Dr. Jekyll (tapeout) 997 Mb 8 1854 1900 2%
4 2574 3413 25%
2 4517 4693 4%
1 8205 9675 15%
Dr. Jekyll (fix-wire-drc) 923 Mb 8 1268 1916 34%
4 2093 2390 12%
2 3748 4071 8%
1 7089 8035 12%
Mr. Hyde (tapeout) 1.4 Gb 8 1721 2605 34%
4 2833 3318 15%
2 5181 6125 15%
1 12039 12141 1%
TSMC Version 2.0 # of Calibre Quartz Calibre's
Chip / GDS Size CPUs Runtime Runtime rel. speed
--------------------------- ---- ------- ------- ----------
Dr. Jekyll (tapeout) 997 Mb 8 820 1047 22%
4 1708 1767 3%
2 3359 3296 -2%
1 7915 6315 -25%
Dr. Jekyll (fix-wire-drc) 923 Mb 8 769 1169 34%
4 1579 1349 -17%
2 3122 2584 -21%
1 6626 4978 -33%
TSMC 65 nm (1.4b) Calibre 15% faster than Quartz
TSMC 65 nm (2.0) Calibre 5% slower than Quartz
Overall Calibre 10% faster than Quartz
CALIBRE VS QUARTZ MEMORY FOOTPRINT SIZES
Calibre clearly wins when comparing memory footprints, using an average
55% less mem than Quartz. But given our small design sizes, this was a
non-issue. I am curious to see how larger designs would be handled.
Memory Footprint Memory (Gb) Memory (Gb)
TSMC 1.4b Calibre Quartz
Dr. Jekyll (tapeout) 997 Mb 1.0 2.2
Dr. Jekyll (fix-wire-drc) 923 Mb 0.8 2.2
Mr. Hyde (tapeout) 1.4 gb 1.2 1.9
TSMC 2.0 Calibre Quartz
Dr. Jekyll (tapeout) 997 Mb 0.8 2.0
Dr. Jekyll (fix-wire-drc) 923 Mb 0.8 2.0
Average memory size 0.9 2.1
QUARTZ FALSE ERROR FLAGS
I then looked thoroughly at the errors that Quartz flagged compared to our
Calibre results for 3 different test cases. With the early version 1.4b,
Quartz flagged numerous false errors, which they improved and mostly
eliminated in the later version.
Quartz False Flags TSMC 1.4b
----------------------------
Dr. Jekyll Dirty 17
Dr. Jekyll Clean 16
Mr. Hyde Clean 16
Obviously, false positives are bad. Why does this happen? Perhaps the most
straightforward answer (as I understand it) is that Calibre DRC decks are
co-written between Mentor and TSMC, and then qualified through TSMC process.
The Quartz deck is written by Magma, run through regression, and then
shipped to TSMC for evaluation. As long as this setup continues, Quartz
deck availability and qualification will always lag that of Calibre. I
believe the delay between releases was a few months.
Below are the set of Quartz's false errors for our "Dr. Jekyll Dirty." By
"dirty," I mean a first pass route in a place-and-route environment that has
not been through the DRC cleanup process.
False flags in Quartz TSMC 1.4b
# of instances false flag
-------------- ----------
4 G4_M2
3 G4_M5
2 M1_S_2
4 M1_S_2_1
3 M3_S_2_1
9 NW_S_6
14 PO_DN_2
198 PP_S_1
1 AP_DN_1
4 DUMM3_S_2
52 DUMM4_S_2
7 DUMM5_S_2
10 M6_DN_1H
74 M6_DN_2
12 M7_DN_1H
1 MOMTEST
39 PO_DN_3
We determined false violations by comparing the trial deck (Quartz) against
a signoff one (Calibre). We would need to review each error if we wanted
to waive it for tapeout. If we couldn't easily determine if the errors is
a false positive, we would relay them back to TSMC for their review, then
derive an updated rule for DRC analysis.
QUARTZ MISSED ERRORS
The missed errors below were all real violations related to via enclosure or
reliability rules, and required fixing prior to tapeout.
Quartz Missed Errors
Chip TSMC 1.4 TSMC 2.0
---------------- -------- --------
Dr. Jekyll Dirty 5 5
Dr. Jekyll Clean 1
Mr. Hyde Clean 2
Below is the list of Quartz missed errors for "Dr. Jekyll Dirty" for both
TSMC 1.4 and 2.0:
# of instances Quartz missed errors
-------------- --------------------
1208 VIA1.EN.2_VIA1.EN.3
1649 VIA2.EN.2_VIA2.EN.3
3 VIA4.R.4.M4
1649 M2.EN.2_M2.EN.3
1075 M3.EN.2_M3.EN.3
Our contract with TSMC specifies that TSMC review our DRC/LVS data by
rerunning the signoff deck using Calibre on their side, so if we had gone
to mask, TSMC would have caught these violations in a Calibre run and
requested that we fix them -- the penalty would have likely been a few
days. However, since this was just an evaluation on our end, I did not
go through that part of the process.
In summary, our runtime comparisons indicated Calibre was 10% faster on
average over Quartz, used 55% less memory than Quartz, and that Quartz
presented false and missed errors which Calibre did not have.
Because TSMC co-develops Calibre decks and uses them for their own signoff,
we know that Calibre will not have false or missed errors. Our final
conclusion was that Calibre is still going to be our signoff deck.
We basically have a set of free Quartz licenses as a part of our agreement
with Magma because we use their Talus P&R. Since we currently build
multiple chips in parallel, we will use Quartz for scrubbing data, but not
for sign-off.
We purchased a set of hierarchical Calibre DRC/LVS licenses and will use it
both for scrubbing and for sign-off.
- [ The LOST Smoke Monster ]
Join
Index
Next->Item
|
|
