( ESNUG 563 Item 10 ) ------------------------------------------- [03/09/17]
Subject: Anirudh and Sawicki on iffy Apache IR-drop #'s vs. Voltus/Innovus
DAC'16 Troublemakers Panel in Austin, TX
Cooley: Anirudh, the perception is that Ansys Gear is beating you in
the competition of Voltus versus Redhawk, because of it's
new Seahawk. Aren't you Voltus guys yesterday's news?
Anirudh: So, first of all I have to comment on this "Big Data" thing.
It's just an algorithm. We like big data, we like small data,
we like clean data, we like dirty data.
Dean: There is no discrimination on data. [audience laughter]
Anirudh: We are in this business to use computer science, mathematics,
and algorithms, and we look at all of them. And it's not just
big data algorithms that are important. There are some really
important algorithms for numerical analysis because it's a
power analysis tool, for differential equations, for matrix
algebra. So all of these things are important and they come
together in an EDA tool.
The point is not how many buzzwords you can use to describe your
EDA tool; the point is what does it actually do for the users.
That's what is important. What is important is that your tool
needs to have performance and capacity.
Now for Voltus we have already done the largest chips that are
made, in a few hours. (See ESNUG 560 #3) So it's a matter of
how many CPUs, you can run 100's of CPUs -- how many CPUs you
throw at it.
The second key point is that it has to be accurate, because it's
a sign-off tool. So if you use all these algorithms and you're
not accurate, then there is no point in talking about...
Cooley: So you're saying is Ansys Redhawk is not accurate?
Anirudh: I think what I've seen is some of the new versions our
competitors' [Apache's] tools and some of the new [Gear]
tools have accuracy problems. (See ESNUG 561 #1)
So then what is the point of a new but slightly inaccurate
sign-off tool? It's not clear to me.
Cooley: Joe, you were nodding your head.
Sawicki: Yeah, I don't think the architectural war is done in power yet,
and I think there'll be other entrants that are going to hit
the scene on this.
I've heard some things in terms of power, in terms of one
solution, where if you do a different number of CPU's you get
a different answer for nodal voltage. (See ESNUG 561 #1)
That's not good. It's worse than inaccurate, inaccurate is
kind of eh, compared to what? But inconsistent sucks.
Anirudh: You have to be, first of all accurate and that requires not
just big-small data. It requires matrix algebra and numeric
analysis and all the points. First is performance and capacity.
Second is accuracy. The third is design closure.
Sometimes there is value of a slightly inaccurate solution.
But that is not for sign-off. That is for optimization, because
in optimization like place and route, sometimes we use the same
engines at slightly less accuracy. But the value then is in
the optimization, the value then is in the implementation tool.
So the real value is to tie Innovus and Voltus together so that
Voltus can drive things in Innovus in implementation. That is
not possible with these tools that are sitting outside PnR
implementation.
So one of the big opportunities, I think, at 10nm and 7nm, is
that IR-drop is taking too much real estate on the chip,
especially if the power is high; you have to drive a lot of
current to these things. So if you combine Voltus and Innovus
we are able to do optimizations in Innovus that can dramatically
improve your overall design. That is the real value of...
Cooley: So you're saying that the hooks to your place and route is what
makes the difference.
Anirudh: It is critical. Of course, accuracy is #1, performance is always
critical, and the third point is it has to affect implementation;
because that is the cockpit that people are designing things in.
To have Innovus make changes based on Voltus IR-drop is a
fabulous thing, and we're getting a lot of traction. And even
simple changes in placement can make a big difference in clocking
schemes, can make a big difference in routing, and I think that
will become more and more critical as we go to these newer nodes.
---- ---- ---- ---- ---- ---- ----
Related Articles
Anirudh and Sawicki on IC Compiler II, Innovus, Nitro-SoC, Antun
Amit and Sawicki on Cadence Spectre, Synopsys HSPICE, Mentor BDA
Dean and Sawicki on Big Data tapeout predictors, John Lee's Gear
Raik and Hogan on formal apps, formal engines, and SNPS VC Formal
Anirudh and Sawicki on iffy Apache IR-drop #'s vs. Voltus/Innovus
Join
Index
Next->Item
|
|
