( ESNUG 479 Item 6 ) -------------------------------------------- [02/05/09]
Subject: ( ESNUG 478 #1 ) Magma Talus is a big step up from BlastFusion
> Since Gary Smith warned (in Oct when Fister resigned) that Cadence might
> discontinue its IC CAD division, my boss wants me to find the current
> viable PD replacements for SoC Encounter. Our next chip is TSMC 45GS.
From: [ Beetlejuice ]
Hi, John,
We have been using Magma's Talus Vortex since the beginning of CY 2008. We
successfully taped out a 65 nm full-chip design using it. Most of our
designs are in the 1.5M instances range at the top level, with some blocks
crossing 2M. This design as a whole is 4.5M instances flat. For Talus
top level runs, we need to look at hierarchical instance count. Top level
1.5M instance count assumes 1 instance for a hard macro. One of the hard
macros in our design has 2M instances in it.
Talus has been a major improvement over BlastFusion for the following:
1. Talus runtimes
2. Buffer count reduction
3. CTS QoR improvement
4. Routing improvements
5. DRC Closure
6. MCMM Optimization
7. Good congestion/hotspot prediction
The details:
1. Talus Vortex runtimes.
Talus' speed of post-layout-optimization is way better than BlastFusion was.
For a 2M instance post-layout top level design, with crosstalk and OCV on.
BlastFusion : over 20 hours
Talus : 6 hours
Talus has multithreading and single-CPU extractor/timer runtime improvements
that BlastFusion never had.
2. Buffer count has reduced tremendously.
Talus CTS gives us improved Insertion Delay/Skews/Buffer Counts. The buffer
count increase in BlastFusion has always been a big concern for us. On one
of our designs, it showed 300K buffer increase. On the same design, Talus
delivered similar timing results with only a ~100K buffer increase. This
was mostly possible with code improvements from R&D.
3. Clock Tree Synthesis (CTS) QoR improvement, reduction in buffer count/area
Talus CTS gave us 45% lower buffer area and 50% lower leakage compared to
BlastFusion CTS, with similar insertion delay and skews. For our 1.2M
placeable instance design (65nm) with close to 40 raw clocks, following are
the CTS stats.
CTS Stats BlastFusion Talus Change
Insertion Delay 10.23 nsec 9.98 nsec -2%
Clock Buffer Count 1735 1083 -37%
Clock Buffer Area 16960 um2 9258 um2 -45%
Clock Tree Leakage 2.99 uW 1.49 uW -50%
4. Routing has come a long way
Talus is able to route our 45 nm 2M instance block, with full crosstalk
avoidance in ~24 hrs. Routing QoR has improved as well. We are getting
sub-2K DRC counts with out-of-the-box fix-wire/fix-DRC in Talus. For us to
get the same in BlastFusion used to take a lot of congestion-relief and
flow customization.
We also are also using Talus on the first 45 nm tapeout in my group. This
is a troublesome design with about 1.5M instances at the top level and the
biggest single block being 2M instances flat.
We're happy with the Talus cycle times we are seeing.
For our 2M placeable inst block, netlist-drop to post-route-timing-close
database, the cycle time is 2 weeks. For 1.5M placeable instances at the
top level, the cycle-time is 10 days from -drop to post-route-timing-closed
database. However depending on the correlation with PrimeTime-SI, we need
to spend extra cycles to optimize timing.
5. DRC Closure - fix-wire (routing) DRC count has come down drastically.
At the top level, our design had around ~40K failing end points coming out
of fix-wire. We called Magma and asked for help. Within a few weeks we are
down to just ~1.5K failing end points. These are the best results we've
seen coming out of routing in a very long time.
6. MCMM Optimization.
We are also actively engaged with Magma in proving their MCMM capabilities
of Talus on our 45nm designs. We use their MCMM for timing closure in our
current tapeout. For the 2M instance block, MCMM hold-fixing was used
in 2 stages:
- Before routing, in global-routed mode, to fix hold violations
across 3 hold scenarios
- After routing, during the post-layout ECOs, to fix a big chunk
of the hold violations, that otherwise would have to be addressed
with manual ECOs.
We had some runtime issues with Talus' earlier MCMM versions. However, some
latest improvements that came in from Magma R&D helped to reduce the MCMM
timer runtime by 50%, which enabled an overnight MCMM hold-fixing run with
crosstalk and native-OCV turned on for 4 scenarios. Specifically, on our
2M placeable instance block (45nm), we started off with a 25 hrs timer
update runtime for 4 active scenarios with crosstalk and native-OCV enabled.
With latest improvements that came in from Magma R&D, this runtime reduced
to 11 hrs, with no loss of accuracy.
7. Good congestion/hotspot prediction
One good example of how Magma's integrated flow helps is, that the same
global-router is used during the placement (fix-cell), CTS (fix-clock) and
final-routing (fix-wire) stages. This gives us better prediction of routing
hot-spots right after fix-cell. Our trials show reasonable correlation
between the predicted hot spots and the eventual DRC hotspots after routing.
Talus provides routability feedback in two forms:
1) a GR-overflow-bucket-count, and
2) a GR congestion map.
The Magma AEs recommend:
a) limiting the GR-overflow-count to under 10K buckets
b) looking at the congestion map to ensure that the above buckets
are spread out, and not clustered very close together.
Small yellow or orange spots are OK as long as they are spread out. The
router does a good clean-up job on these. However, a lot of buckets
clustered together at the same location, means a routing hot-spot, and is
what we call a "blob".
For our 1.5M top level design (45 nm) we had 1.5K overflow buckets after
fix-cell. This increased to about 5K after CTS, due to application of
non-default-rules on the clock nets, and due to the post-CTS slew-fixing
and timing optimization. These 5K buckets were clustered in 2 major blobs.
After crosstalk driven routing, we had ~2K DRCs, almost all of which were
coinciding exactly with the above blobs.
Talus' biggest strengths are turnaround-time, CTS, routing, post-layout
optimization, and DRC closure. The attitude of the AEs is outstanding and
refreshing. Magma's R&D has given us great support as well.
- [ Beetlejuice ]
Join
Index
Next->Item
|
|
