( ESNUG 571 Item 3 ) -------------------------------------------- [04/21/17]
Subject: Anirudh's 32 jabs at Aart de Geus's ICC/ICC2 on his Pegasus launch
Cadence Innovus -- officially launched 15 months ago, is Anirudh's
attack on Aart's ICC/ICC2. Since then there was a major hush-hush
kerfluffle where Apple benchmarked it at 14/10nm against ICC2 and
then switched over to Innovus. (ESNUG 547 #7) Then Innovus won in
another TSMC 16FF+ user benchmark against ICC/ICC2. (ESNUG 550 #1)
Then 5 different PD engineers discussed Aart's patch rev of ICC2 of
Sept 2015, how Innovus works well at 16/14nm, and how Innovus thrives
on the "new double-/triple-/quadruple-patterning digital color layout
problem at 10nm that Aart's ICC2 can't handle." (ESNUG 552 #6)
Then Imec's world's first 5nm tapeout only used Innovus, Tempus,
Virtuoso, MENT Calibre -- no ICC/ICC2, no PrimeTime. (ESNUG 554 #3)
This year Anirudh's Innovus minions have been busy claiming "10-15%
better PPA using new Power Opt", "better placement QoR by 20% with
GigaPlace", and better parallel and distributed "for more capacity".
HiSilicon 20% less area. Toshiba 16% less area & 25% less power.
10nm/7nm. "Deep hooks into Virtuoso for your A/D chips, too!"
Qualcomm, Nvidia, ST, Faraday, HiSilicon, Broadcom, ARM, NXP users
(booth 107) Ask for KT Moore. Freebie: Denali party tickets
- Cooley's Cheesy Must See List for DAC 2016
From: [ John Cooley of DeepChip.com ]
Last week, Anirudh gave a big keynote where he surprise launched Pegasus
to take on Joe Sawicki's Calibre DRC/LVS empire. In that same CDNlive'17
keynote, Anirudh announced how Innovus was changing, too.
What follows is what Anirudh (pretty much) said followed by what in my mind
I heard him say.
- John Cooley
DeepChip.com Hollistion, MA
---- ---- ---- ---- ---- ---- ----
(click to enlarge pic)
|
WHAT ANIRUDH SAID:
"... blah blah I'm proud to say after 24 months, Cadence digital PnR is
now being used in 16 of the top 20 semiconductor houses blah blah ...
preferred tool for CPUs blah GPUs blah blah ..."
WHAT COOLEY HEARD:
This was Anridh saying "IN YOUR FACE!" to Aart in the digital PnR market.
Looking at the IC Insights Top 20 for 2016, it was much easier just to
redline out where CDNS Innovus wasn't strong at...
Intel is easily Aart's biggest ICC/ICC2 customer by far. Neither Aart
nor Anirudh are big at SK Hynix in digital PnR. (For Virtuoso, yes,
but Innovus, no.) Toshiba and Infineon are both bigtime SNPS PnR users.
On the flip side, Samsung, Qualcomm, Broadcom, Apple, TI, NXP, Nvidia,
TSMC, GlobalFoundries, MediaTek, and even ARM are all big Innovus users;
most either at 10nm or 7nm now. What's embarrassing here is 2 years
ago practically all of those logos were in Aart's ICC camp; that is,
until the ICC2 fiasco hit... which got Antun "promoted" to SNPS CTO.
(that's 16 jabs.)
That "preferred solution for CPUs and GPUs" is Anirudh laughing that
ARM optimizes on Innovus PnR in-house, and that his customers get better
QoR on ARM cores when using Innovus vs. ICC2. (that's 1 more jab.)
Anirudh's Genus/Tempus/Voltus mentions are because he has to give them
mentions. Aart's Design Compiler and PrimeTime, and John "Jolly" Lee's
RedHawk still easily dominate their respective EDA tool niches with
ballpark 70% to 80% marketshare. Anirudh only does "me, too" there.
ICC/ICC2 Jab Count: 17 Total ICC/ICC2 Jab Count: 17
---- ---- ---- ---- ---- ---- ----
(click to enlarge pic)
|
WHAT ANIRUDH SAID:
"... we know SOC sizes are growing to over 1B instances. SoC block sizes
need to grow 1M to 5-10M instances. We already have established a PPA
leadership position in P&R with Innovus on single machine. In this
example our full flow PPA delivered 17% die size reduction with 4X
better run time."
WHAT COOLEY HEARD:
As best as I can figure, Anirudh's "reference flow" on the chart on the
right is Synopsys ICC2 plus Synopsys DC-Topo. Remove ICC2, add Innovus
and your die area drops 10%. Then if you remove DC-Topo and replace it
with Genus, your die area drops 17%. This on a single machine. Now
since Innovus block size has grown to 5 to 10M insts, this 5M might have
been run as a single block. (that's 3 jabs)
ICC/ICC2 Jab Count: 3 Total ICC/ICC2 Jab Count: 20
---- ---- ---- ---- ---- ---- ----
(click to enlarge pic)
|
WHAT ANIRUDH SAID:
"... to be able to handle 5-10M+ blocks you need a fully distributed flow
from synthesis to signoff. Genus, Tempus, and Voltus are all fully
distributed today -- but what about Innovus?"
WHAT COOLEY HEARD:
Anirudh is setting the stage to say his Innovus is now fully distributed.
The giveaway is how memory and runtime goes down 4X, when he runs his
new Innovus on 4 servers instead of 1 server. (that's 1 jab)
ICC/ICC2 Jab Count: 1 Total ICC/ICC2 Jab Count: 21
---- ---- ---- ---- ---- ---- ----
(click to enlarge pic)
|
WHAT ANIRUDH SAID:
"Innovus is now the first PnR tool that is fully distributed. It will
be released in May 2017. It's also the first massively parallel flow
from RTL through PnR out to signoff."
"Fully distributed means multi-machine and multi-threads. Expect 2X to
3X faster runtimes with the same (or better) QoR/PPA."
WHAT COOLEY HEARD:
Anirudh's saying he's beating Aart's ICC2 PnR. ICC2 stops at 8 or 16
CPUs all on the same machine. Here Innovus ran a total of 32 CPUs
split up over 4 machines; that's multi-threaded inside the 4 machines.
(that's 2 jabs; 1 for fully dist, and 1 for beating ICC2)
ICC/ICC2 Jab Count: 2 Total ICC/ICC2 Jab Count: 23
---- ---- ---- ---- ---- ---- ----
(click to enlarge pic)
|
WHAT ANIRUDH SAID:
"We have 2 types of speedup in our Cadence Innovus 17.1 release. First
multi-threading efficiency gets an average 1.5X speedup on 8 CPUs.
Secondly, distributed speedup is up 2X with same or better PPA. Net
result is a 2-3X total runtime speed-up for large designs."
WHAT COOLEY HEARD:
Anirudh's saying: "Before, we were fast and got good PPA. Now, with
all the multi-CPU/multi-threading we get the same (or better) good
PPA in 1/2 to 1/3rd the time. Oh, and we have a kickass 7M inst
block size now, too!" (that's 2 jabs; multi-everything & bigger blocks)
ICC/ICC2 Jab Count: 2 Total ICC/ICC2 Jab Count: 25
---- ---- ---- ---- ---- ---- ----
(click to enlarge pic)
|
WHAT ANIRUDH SAID:
"... blah blah machine learning, machine learning, machine learning."
"We did this over 2 human years, or 14 EDA years..."
WHAT COOLEY HEARD:
Anirudh said: "Just like TSMC and Solido, we've been secretly working
on machine learning for Innovus over the past two years. I'm happy
to report a 12% better TNS just from machine learning! As far as we
can see, neither Aart nor Wally have operational machine learning in
their PnR tools today. Maybe in a few years? But not now." (3 jabs)
ICC/ICC2 Jab Count: 3 Total ICC/ICC2 Jab Count: 28
---- ---- ---- ---- ---- ---- ----
(click to enlarge pic)
|
WHAT ANIRUDH SAID:
"... blah blah machine learning, machine learning, machine learning."
WHAT COOLEY HEARD:
"Oh, and we're applying machine learning all over the place inside
Innovus for floorplanning, routing, placement, CTS, etc...."
ICC/ICC2 Jab Count: 4 Total ICC/ICC2 Jab Count: 32
---- ---- ---- ---- ---- ---- ----
On Tuesday of next week, I'll have
A quickie summary of how the Pegasus vs. Calibre war will play out
published. This post and the Pegasus announcement just set the stage.
- John Cooley
DeepChip.com Holliston, MA
---- ---- ---- ---- ---- ---- ----
Related Articles
Anirudh's 19 jabs at Joe Sawicki's Calibre with his Pegasus launch
Anirudh's 32 jabs at Aart de Geus's ICC/ICC2 on his Pegasus launch
A quickie summary of how the Pegasus vs. Calibre war will play out
Join
Index
Next->Item
|
|
