Is NVIDIA's decision to license its Kepler GPU core a
stunningly shrewd business move that brings a compact, powerful but low-power
piece of IP into the SoC design world? Or is it a Hail Mary pass destined to
slip past outstretched hands?
The cop-out answer, of course, is "it's too early to tell," but let's walk through the scenarios because NVIDIA's decision is emblematic of the
roiling change in system design today.
NVIDIA Executive Vice President and General Counsel David Shannon
announced the decision in a blog
post on June 18, saying:
"It's not practical to build silicon or systems to address
every part of the expanding market. Adopting a new business approach will allow
us to address the universe of devices. So, our next step is to license our GPU
cores and visual computing patent portfolio to device manufacturers to serve
the needs of a large piece of the market."
A broad strategy will begin with the NVIDIA Kepler
architecture. Its performance and functionality are pretty well known; the
core is at the heart of the next-generation Tegra mobile processor (Logan)
and the current GeForce, Quadro, and Tesla graphics processors. In addition, Shannon
points out in his post that Kepler "operates in the half-watt power envelope."
Cue dramatic orchestral music as the curtain rises on a new
mobile applications market for NVIDIA (mobile).
So there is clearly plenty of potential for the
cores-licensing strategy (and it's not new to NVIDIA, which licensed GPU
technology to Sony and patents to
Intel). But the road to paradise is not always strewn with red roses.
From a business perspective, the licensing of a core is not
the same as selling an IC. For one thing, there's a different revenue stream
from your customer, who has to sell its devices with your cores before you see
Says Semico Industry Analyst Rich Wawryzniak, "One of their
GPUs could go as high as $30-$40 per part. They'll never get that licensing IP." Well, yes and no: ARM's shown you can make it up in volume, but there
aren't many ARMs in the IP business.
But the big challenge is in the design-in process. A
standalone processor solves a big bounded problem for a system architect. And in
relative terms, it's a turnkey solution. Convincing an SoC designer to
incorporate your graphics core into his or her IC is a more complex process not just because it's a competitive space but because the process and methodology of IP integration is still evolving.
"You're not buying the IP like a frozen leg
of lamb from the butcher and all you have to do is pop it in the oven. You need
an infrastructure internally to manage the IP."
While engineering consultant and EDA DesignLine editor Brian
Bailey says "having a design
does not make you an IP company," he also notes that a serious, marshaled IP
effort at NVIDIA will drive the company's long-term R&D. That means new
directions and opportunities.
And "Not all
chips need to be the most powerful; they may need to be the most power frugal,"
Cool graphics, coming to an IP block near you, tell you I do. (source: NVIDIA.com)
This churn is
amid rapid change in system design. Differentation, as Frank Schirrmeister
points out in a cogent post on the System-Level Design Community site, is increasingly
hard for semiconductor vendors. He writes:
"With the IP providers providing more and more complex subsystems and the value in system houses moving upward into applications—essentially making the hardware for phones and tablets delivery devices for applications—the differentiation of semiconductor vendors has become harder and harder."
Simply put, semiconductor vendors are looking to move up into the software stacks at the same time they're considering moving down into the IP space.
"In the future, if you are not a systems integrator, you will be
an IP supplier. There can be systems integrators at the board level, but most
will be at the chip level," says Bailey.
This means tough decisions lie ahead
for silicon providers, and, in the end, the only wrong answer is
staying the course. Technology, business, and methodology change so
relentlessly that the only alternative to change is death or irrelevance (which is the same thing as death in the Silicon Valley). NVIDIA Chief Scientist Bill Dally sounded the call for change when he talked about what he calls the digital productivity crisis in a Design Automation Conference talk this month.
Tensilica co-founder Chris Rowen
sketches out three change scenarios and their challenges for companies in similar situations.
- The first is pivot to the broad SoC
business, building on top of one's core technology.
that requires so much market know-how in all these different segments and a
willingness to mix high-margin and low-margin silicon business together," Rowen
- The second is
just to avoid wading into new markets and stick to your knitting.
- The third is to
adopt a licensing model for low-margin, high-volume segments.
(that) doesn't work in practice because resources and corporate priorities
remain focused on the chip business," Rowen says.
It's rare for a
silicon vendor going 200 mph to pivot gracefully into new models and markets,
but when it happens it's a beautiful and awe-inspiring thing. NVIDIA is chock
full of smart people—led by cofounders Jen-Hsun
Huang and Chris Malachowsky—and they have as a good a chance as any company
to do it. And in 2013, there's an ecosystem around them to help out.
I'll follow up a
post on the impact on EDA, but for now...
- What do you
- Am I overly
- What pitfalls do you see for companies grappling with system-design change today?
NVidia’s Bill Dally: Chip Design Should Take Days, Not Months