( ESNUG 590 Item 04 ) --------------------------------------------- [06/04/21]
Subject: Costello takes on Anirudh and Sawicki on per-minute cloud pricing
Virtual DAC'20 Troublemakers Panel
[ Editor's Note: Although this discussion goes in many
directions, this is their main argument. - John ]
Cooley: [to Joe Costello] So alright, you at Metrics have this massive
breakthrough in regressions; you can speed-up really fast using
cloud technology.
The purchasing agents in user companies - they're used to
dealing with Synopsys, Cadence, and Mentor. They don't want
to talk to tiny Metrics. You're just yet another set of
paperwork and headaches to them.
Your Metrics regressions may not get a sale just because you're
too much purchase order bureaucracy to bring in the door.
Costello: Absolutely true. The first blush of any purchasing guy when it
comes to EDA is "no".
It doesn't matter what it is. And they have kind of the same
allergic response to any new EDA technology which is:
"Oh, I know you're going to tell me some engineer loves it
and it's going to save their life and make their chip
better. I've heard that story a hundred times. Trust me,
I don't believe it. Forget it. My engineers have a flow.
We spend a ton of money -- too much money -- on EDA stuff."
And that's kind of the original pattern. So, you've got to let
them vent -- you know let them get that out of their system.
And then we say...
Cooley: Will they even let you in the room, though?
Costello: Oh, we can usually get in the room, because they feel like they
have to have a meeting. Otherwise their engineering guys will
throw a fit that they weren't even given a day in court.
You know you've got to have your day in court.
And when they [the customer's Purchasing Agent] give you your
day in court, it's going to be short.
"Guilty." [laughter]
Anirudh: John, I do have a comment here on verification and the Cloud.
Because verification is always endless problems you can simulate
more and more -- and you know Joe (Costello) knows that -- and
all the users watching know that.
|
VS.
|
|
But we had an earlier discussion on static; static means to me
no excitation -- dynamic which is logic simulation is excitation
or vectors. So, when you do logic simulation, you're inherently
running millions and millions of jobs, which is inherently
parallel.
So, we have a lot of customers running on the cloud; all the
major EDA providers do. Cloud has a huge value to customers,
but by itself, it's not a differentiator in my opinion.
Because logic simulation is very well suited to run on the cloud,
or basically a large number of machines, and that's what users
have done.
Now what the cloud allows the users to do is flex up and down.
But that is true of almost any tool that runs parallel.
To me there is opportunity to manage regression in verification.
Cloud is useful, but I don't think it will be a differentiator
by itself. You have to have other technologies to manage
regression. (See Best'20 #2a)
And, of course, you need to have the best HDL simulator for
each simulation (Best'20 #2b) and then a way to manage these
constrained random regressions, along with using the cloud.
Cooley: Well, basically Costello just made the claim that his company,
Metrics is beating you, Anirudh (Cadence) and beating you, Joe
Sawicki (Mentor), in some way.
And he's using the cloud and you guys are also using the cloud,
so something is different here and I don't know what it is.
Costello: Let me answer, here's one thing that's fundamentally different.
To take maximum advantage of the cloud and truly flex; you've
got to have built your EDA tool in a cloud native way.
Not using the old school, last generation VMware approach to
managing it. You've got to use Kubernetes and container-based
approach to doing that. And that's a new wave.
And by the way the biggest stumbling block when we actually go
into accounts and start to get them to use this kind of
cloud-based regression, the biggest stumbling block happens to
be exactly that. The companies haven't really made the move.
These semiconductor companies are relatively backward when it
comes to IT technology and the move.
And so, we have to spend time with them getting them to move that
way if you want to use cloud in an efficient, transparent
fashion.
Everything Anirudh said is true. You've got to have a good
simulator; you've got to have great tools to manage it. But by
the way, once you have that cloud and flex capability you can
manage your regressions in a completely different way.
And by the way, when we [Metrics] go in and we reduce something
from originally 4 to 5 days -- down to 4 hours -- when they
[the users] actually then look at their regression set and
they re-architect it knowing that they have that cloud flex
capability, they can go even faster than that.
So, it's all true what Anirudh says. You've got to manage it;
you've got to think about it.
But what are the tools in your tool set? If you know you have
essentially infinite compute capacity available -- and the tools
to manage it in a transparent way -- cross in a hybrid fashion
from my local data center into the cloud because I'm using cloud
native technology -- well, that's a very different tool set and
then you can start rethinking things and you can get way better
performance.
Sawicki: I'm not sure about the comment of new tool / old tool and can't
use Kubernetes.
I mean that's all technology that's being leveraged by our groups
inside Mentor when they're putting in place the cloud activities.
But one thing I'd emphasize on Joe's part, is that I think what's
emerging fairly quickly is that as we talked about: I think last
year cloud is becoming the new IT.
It's going to be the IT solution that people are going to be
looking for. They're going to be looking for that in a hybrid
fashion. They're going to look for that to scale across their
internal network, being able to burst out to the outside.
We've got Calibre running inside in Kubernetes; we'll be rolling
that out to customers fairly soon. We've got our simulators
running in hybrid fashion where they can run part on their
internal cloud and part on the external cloud.
This stuff is all happening. It's all happening. On the stuff
Joe is talking about though, it is really all about how I can get
that massive compute capacity... again it's not the simulator
is faster, it's that I can grab enough compute resources to
take my end-of-chip verification -- whether that be physical or
functional -- and take that from a month's worth of work
down to 3 days worth of work.
That's an enormous benefit to the customers. This is happening.
It's going to be happening with new tools. It's going to be
happening with old tools.
Cooley: But you're going to have to do it over 10,000 processors that you
rent or something like that...
Costello: Yes. And by the way when you go to that model, the traditional
license model for software doesn't really make any sense.
The whole move to the cloud is a SaaS kind of model and at the
same time you've got to go that direction.
Going back to your question, John, about the purchasing guys, the
thing that opens their eyes -- the purchasing guys -- is exactly
what Sawicki was saying.
When you say, "just hold on a second this is not just a technology
story. It's not just an engineer coming in trying to get you to
spend more money on a new fancy tool that they like, good or bad."
There's something fundamental going on here -- which is there's
an IT shift. You're moving from your own data center into the
cloud. That's what's happening.
And by the way, that's what gets the purchasing guys going right
away. They go, "yeah, that's interesting we do that in every
other part of our company, but we haven't really done it yet
in engineering -- especially in the semiconductor market. We
haven't thought that through."
And I make the point, with the classic line of Jerry Sanders
(founder of AMD) who said:
"Only real men have fabs." - Jerry Sanders, founder of AMD
Ironically, his company (AMD) now is doing really well because
they don't have a fab -- you know compared to Intel. But I said,
listen that was not a true statement, but you could imagine that
for a semiconductor industry having a fab sounds kind of core to
the business -- at least it's semiconductor related. Now, it's:
"Real men certainly don't have data centers."
There's nothing about a semiconductor company that says you
should be running a data center.
So that's what opens the purchasing guy's mind. Because they see
this model is also innovative for them, and it meshes and matches
what they're doing in the rest of their business and they go:
"Yeah, that makes sense we should be thinking about that."
Cooley: [to Costello] Are you [Metrics] still selling based on the
penny-per-minute model?
Costello: It's a SaaS model -- you have some number of cents-per-minute.
Cooley: Sawicki, are you [Siemens/MENT] doing pennies-per-minute?
Sawicki: To date, in terms of the business model shift, you see interest
in that from smaller houses. You don't see it from larger
houses.
You know we [Siemens/MENT] had this happen last year as well.
And just to be clear that for most of our larger customers,
there is no "traditional" business model out there.
It's not like in the days where we had $100,000 workstations and
they'd get a discount to $80,000. It's a completely different
business model that has the old FRP as an aberration of history.
There's not that big of a transition when those types of folks
want to move into this burst cloud type of model.
Cooley: So, you're doing a per-minute model? You also offer that?
Sawicki: Nope, I very studiously didn't say that, John. There are some
customers who want to go that way. But when you look at the way
people are buying today, I'm just saying that this whole thing of
"oh, they're paying so much money now today for these licenses
because they're buying them on a per-seat basis and the
utilization..." that's really taken out in the company by company
business model.
Costello: But, John, I will tell you that it is not true that's it's only
for smaller companies. Every single customer is interested in
the by-minute pricing, large and small.
I just had a discussion with purchasing guys from a very large
company -- Fortune 50 company -- and that was number one on
their mind. In fact, they made the proposal in per-minutes.
Cooley: Is Cadence doing that type of stuff? Are you seeing Synopsys?
Anirudh: We look at all business models, and we are open to all business
models, but you know when you say SaaS, most of the SaaS
companies I'm working with are for our IT -- like we buy
Salesforce, Workday, these are all SaaS companies -- they
normally have a yearly subscription model just to be clear.
So, like Sawicki was saying most of our [Cadence] customers want
a yearly subscription model -- and that's what most of the SaaS
companies do also. Okay we have a contract with Workday and
Salesforce -- it's a yearly subscription model.
One thing, we [Cadence] should take all credit for the industry
that EDA moved to a subscription model way before the enterprise
software companies did.
Cooley: Yeah, but you're doing subscription on an annual basis, I'm
talking about minute-by-minute where your engineers can rack
up...
Costello: I want to be 100% clear about this -- the large Fortune 50
company, when they asked us [Metrics] for a price-per-minute,
they also didn't want a subscription -- they want to buy
by-the-minute, period.
Part of selling by the minute is the underlying infrastructure
and plumbing. If you're Kubernetes and a cloud native base;
that's really easy to do because it's kind of built into our
model. And if you're not, not it's not.
But, there are times when they want 12,000 licenses and servers
and not the 1,000 licenses and servers that they have.
Sawicki: And I think that last thing where it's like you know where they
want 10,000 versus 1,000 servers -- the presumption is the only
way you can get there is through a by-minute business model. And
that's just patently not true. All our enterprise customers have
models where they can go ahead and do those sorts of things.
I just don't believe that in the end either Anirudh or I are
going to lose this market because of our unwillingness to price
by-the-minute. We'll get the business models our customers need
in place. In our conversations with users, you rarely end up in
a discussion around renewals and on how they want to do a pricing
model.
Far more discussions are how they're going to be able to do
validation of an application stack on top of multiple CPUs,
running in a system context, and doing that validation across
multiple pieces of firmware and an end user software stack.
And how they do that validation.
That goes back to Anirudh's point where it's just a simulator by
the minute. Sounds good. If our customers need it, we'll do
that. But it's not the problem they're trying to solve right
now.
---- ---- ---- ---- ---- ---- ----
Related Articles:
Costello on his Metrics cloud 4 hours vs. SNPS/CDNS/MENT's 5 days
Costello on Metrics cloud Verilog is at right place, right time
Costello on Montana's 5X to 20X speed-up and merger with Metrics
Cooley asks Metrics Tech: "did anyone there think this through?"
SCOOP -- a first look at the pay-per-minute Metrics SV simulator
Join
Index
Next->Item
|
|
