( ESNUG 486 Item 8 ) -------------------------------------------- [06/01/10]
Subject: (ESNUG 485 #6) Oops! Quartz had no missed/false errors vs Calibre!
> Below is the list of Quartz missed errors for our "Dr. Jekyll Dirty" chip
> 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 been a few days.
>
> - Adam Datesman
> Availink, Inc. Germantown, MD
From: Adam Datesman <adatesman=user domain=availink not mom>
Hi, John,
I bet this is the first time anyone has rebutted their own review... but it
turns out that Magma's Quartz had no missing or false errors, and Quartz
consistently performed faster than Calibre.
After our initial posting on Deepchip, we discussed the data from our review
with Magma. While Magma technical support is usually good, this set of
issues had not received the normal attention before I wrote this article.
THE CORRECTED MISSED ERROR, FALSE ERROR QUARTZ DATA:
Magma has since shown us how to run on a local disk for single machine runs
and assisted us to review the DRC error markers. We did not change the
runset, nor the software -- we used the same TSMC 65 nm V2.0 runset and
Quartz DRC 2010.2.1. release from the original review. Note: we did not
update the data from the older TSMC 65 nm runset (V1.4), as it is not used
for production any more.
We originally reported 5 types of missed errors for the Quartz V2.0 runset.
But after reviewing the error markers with Magma, we have verified that
QUARTZ HAD NO MISSED ERRORS.
For anyone else running a benchmark, it is worth mentioning that you need
to work with Magma AE's to understand the difference in reporting errors
between the two tools. Many errors are handled the same, but there are
differences that require training.
In our test data, we had overlapping vias that formed illegal via shapes.
Both Calibre and Quartz report this as an error (illegal via shape). But
Calibre double reported the violation in the following 4 enclosure error
types:
VIA1.EN.2_VIA1.EN.3
VIA2.EN.2_VIA2.EN.3
M2.EN.2_M2.EN.3
M3.EN.2_M3.EN.3
Both approaches are fine to us; TSMC accepts both approaches for sign-off.
The 5th type of error difference (VIA4.R.4.M4) was just a rule naming
difference between the two tools that we overlooked.
In addition for our V2.0 runset, QUARTZ HAD NO FALSE ERRORS.
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
THE CORRECTED CALIBRE VS QUARTZ RUNTIME COMPARISONS:
> TSMC Version 2.0 # of Calibre Quartz Calibre's
> Chip / GDS Size CPUs Runtime Runtime rel. speed
> --------------------------- ---- ------- ------- ----------
> Dr. Jekyll (fix-wire-drc) 923 Mb 8 769 1169 34%
> 4 1579 1349 -17%
> 2 3122 2584 -21%
> 1 6626 4978 -33%
Here's our corrected version of our runtime benchmark:
TSMC Version 2.0 # of Calibre Quartz Calibre's
Chip / GDS Size CPUs Runtime Runtime rel. speed
--------------------------- ---- ------- ------- ----------
Dr. Jekyll (fix-wire-drc) 923 Mb 8 769 746 -3%
4 1579 1309 -20%
2 3122 2558 -22%
1 6626 5107 -29%
Our prior data that TSMC 65 nm (2.0) Calibre is 5% slower than Quartz was
wrong; the correct data is Calibre is on average 18.5% slower than Quartz.
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
CALIBRE VS QUARTZ MEMORY FOOTPRINT SIZES:
Even with our new data, Calibre still clearly wins when comparing memory
footprints, using an average 55% less mem than Quartz.
In the case of the tests we did, the GDS file size was small enough and the
memory on our machine was large enough, that the memory usage is essentially
irrelevant. Both tools ran our current full chips using significantly less
memory than our machines have.
Quartz has a "memory controlled" architecture, which may use more memory for
small designs, but automatically limits it usage when you get closer to the
machine memory limits for larger designs. This lets Quartz to get more
performance without exceeding our machine's memory limit.
We don't have data to test this right now, but our upcoming designs are 5X
to 10X larger than these designs. We are curious to see how this Magma
feature works on our upcoming tape outs,
In summary, we have concluded that neither Quartz nor the 2.0 run deck had
any issues. Our corrected results now show that Magma's Quartz had NO false
nor missed errors (our bad!), and it is faster than Calibre. We are now
working to deploy it as a sign-off tool.
- Adam Datesman
Availink, Inc. Germantown, MD
Join
Index
Next->Item
|
|
