( ESNUG 560 Item 3 ) -------------------------------------------- [04/29/16]
Subject: The Nvidia stealth benchmark of CDNS Voltus vs. ANSS Gear RedHawk
SCOOP! - Word is in addition to Cadence Voltus, Ansys Apache RedHawk
now has a new IR-drop and electromigration competitor from a John Lee
start-up called Gear Design Solutions.
John Lee is a big name in DFM. He created Mojave Quartz DRC/LVS which
gave MENT Calibre major headaches back in 2006. (See ELSE'06 #18).
"Big Data techniques for IR-drop/EM!" is Gear's new claim to fame.
I've heard conflicting rumors that Gear either 100% won or 100% failed
in a recent RedHawk/Voltus/Gear benchmark at Broadcom.
- http://www.deepchip.com/items/0550-04.html (DeepChip 05/15/15)
From: [ John Cooley of DeepChip.com ]
Hi, DeepChip readers,
There I was in a late afternoon session at CDNlive in a dark room where a
sharp female Indian Phd from Nvidia was giving a talk titled:
"First Pass Silicon-Accurate Power Grip Signoff
for High Performance GPU's with Voltus-DP"
I'll be honest. I hadn't read the abstract and I had just skimmed the title
so I expected it to be a talk on mult-threading EDA tool runs through the
Nvidia GPU's. (Based on the words "GPU" and "Nvidia" in the title.)
And sure enough the first slide confirmed what I was expecting...
(click on pic to enlarge image)
|
WHAT SHE SAID: Blah, blah, blah distributed processing blah blah massively
parallel blah blah multi-treading blah blah 32-CPU's blah blah runs on 100's
of machines blah blah 10X TAT...
WHAT I HEARD: Blah, blah, blah distributed processing blah blah massively
parallel blah blah multi-treading blah blah 32-CPU's blah blah runs on 100's
of machines blah blah 10X TAT...
---- ---- ---- ---- ---- ---- ----
Then the running-EDA-tools-on-Nvidia-GPUs talk took an unexpected left turn
(that was only unexpected because I hadn't read the abstract for the talk)
and quickly morphed into a designing-Nividia-GPUs-using-an-EDA-tool talk.
Oops.
(click on pic to enlarge image)
|
WHAT SHE SAID: Blah blah our team designed a 16nm FinFET GPU blah blah blah
100's of millions of instances blah blah Tempus blah blah Voltus DP blah
blah package and noise problems...
WHAT I HEARD: Our team at Nvidia did a TSMC 16nm GPU (probably ~300 M inst)
with the whole CDNS digital PnR flow -- Innovus/Tempus/Voltus-DP.
---- ---- ---- ---- ---- ---- ----
(click on pic to enlarge image)
|
WHAT SHE SAID: Blah blah blah two pass flow blah blah run flat design
blah blah Tempus blah accuracy blah blah...
WHAT I HEARD: For some reason to do IR-drop with Voltus-DP, it requires a
two-pass flow if you're running it on a big single flat chip... Is this
for capacity or accuracy or what?
---- ---- ---- ---- ---- ---- ----
(click on pic to enlarge image)
|
WHAT SHE SAID: Blah blah Voltus-DP predicted vs. actual waveforms blah blah
real waveforms were attenuated as compared to the Voltus-DP prediction...
WHAT I HEARD: These attenuated actual waveforms are bad news! Bad! Bad!
Bad! Somehow Voltus-DP messed up somewhere!
---- ---- ---- ---- ---- ---- ----
(click on pic to enlarge image)
|
WHAT SHE SAID: Blah blah because our test mode had a very high simultaneous
switching activity, our L*di/dt was significantly more than our IR-drop...
blah blah derate blah blah accuracy blah blah Tempus blah blah...
WHAT I HEARD: The high inductance of the damn package wire was swamping out
our on-chip IR-drop!!! We fixed that with a clever derate scheme that has
somehow involved Tempus for accuracy.
---- ---- ---- ---- ---- ---- ----
(click on pic to enlarge image)
|
WHAT SHE SAID: Blah blah package model blah derate blah...
WHAT I HEARD: See? Our clever derate scheme worked!
---- ---- ---- ---- ---- ---- ----
(click on pic to enlarge image)
|
WHAT SHE SAID: Blah blah instance-based derate blah Tempus blah blah blah...
WHAT I HEARD: Voltus does instance-based derates. (I think that's something
that only Voltus can do... but only if you use it with Tempus...)
---- ---- ---- ---- ---- ---- ----
AND NOW FOR THE BIG SURPRISE SHOCKER ENDING !!!
(click on pic to enlarge image)
|
WHAT SHE SAID: Blah blah after we did the derates, our die-package Voltus-DP
simulations correlated execeptionally well with final silicon blah blah blah
blah blah we used Voltus because the other commercial IR-drop tools couldn't
handle our chip because it was too big for them blah blah and yes, of course
we looked at all the other IR-drop tools in the market blah blah...
WHAT I HEARD: In the end, the Voltus-DP die-package simulations matched up
with real silicon perfectly. Yes, we looked at the Gear flavor of Ansys
RedHawk and it crashed! Same for SNPS PrimeRail! The only tool that could
IR-drop our big ass chip was Voltus-DP. We benchmarked Voltus-DP and it
trounced both Apache Gear RedHawk and SNPS PrimeRail in capacity. ("And
your ~300 M inst estimate is waaaay off, you idiot, Cooley!! Nyah! Nyah!")
And Voltus-DP did this chip flat! Whoa, mama! Not hierarchical, but flat!
---- ---- ---- ---- ---- ---- ----
(click on pic to enlarge image)
|
WHAT SHE SAID: Blah, blah, blah distributed processing blah blah massively
parallel blah blah multi-treading blah blah 64 machines of 8-CPUs each blah
blah blah blah simulated 35 nsec in 24 hours of clock time...
WHAT I HEARD: See? I actually did a distributed processing talk after all!
But it was on how fast Voltus-DP was on a honking gazillion instance chip
(which I ain't going to tell you the true inst count of, Cooley!) that both
the ANSS Apache Gear RedHawk and SNPS PrimeRail tools both choked on!
---- ---- ---- ---- ---- ---- ----
That was fun.
- John Cooley
DeepChip.com Holliston, MA
Join
Index
Next->Item
|
