Beyond3D.com interview http://www.beyond3d.com/interviews/jhh/
________________________
A few weeks prior to the launch on GeForce 7800 GTX, based on NVIDIA's
latest G70 chip, NVIDIA held an "Editors Day" to introduce the press to
the
new architecture and capabilities. Whilst that the Editors Day we got the
op****tunity to sit down with NVIDIA's CEO and visionary, Jen-Hsun Huang
and
ask him a few questions...
--------------------------------------------------------------------------------
B3D: Obviously NV40 was somewhat of a turning point for NVIDIA; what
would
you say were the most im****tant aspects behind its success?
Jen-Hsun: In a lot of ways I thought NV30 was a turning point, because
NV3,
NV4, NV5, NV10, NV11, NV15, NV17, NV18, NV20, NV25 was a straight line for
us; we were on a path that was "make it faster, make it better". When it
came to NV20, that was when we invented programmable shaders, which was a
big, big chunk to swallow - we quickly realised that in building
programmable shaders fixed point [integer], FX12 format, shading was
limited
in its capability and we needed to fundamentally take it to the next level
and we took it to FP32 [32-bit floating point per component] with NV30 -
that was a mistake for us. I would say that was a turning point in fact.
When you make a mistake that comes along after 5 or 6 years it really
shakes
you up and that causes you to make you think about what you need to do
different. Obviously, all the things that we did before worked so well and
then all of a sudden you make a chip that is expensive, very difficult to
manufacture and frankly not fast enough, so we had to go back to basics
and
think about how we re-architect our company and the way we do things;
remain
as innovative as we always have been but innovate in a way that is much
more
aligned with industries. So that was actually the turning point, we
started
fixing ourselves right around NV30 and we quickly came back with NV35 and
NV36 as a way to improve on that architecture because it usually takes a
good two years to bring a new architecture to the market; so we fixed
[GeForce] FX and then brought the 6800 [NV40] to market in about a 1 year
timeframe, which was an amazing cycle time for us. So the changes that are
seen on the outside of the company really started a good, solid year and a
half before.
So, we have a few core values - we want to retain out spirit of innovation
and always wanting to push the edge, wanting to do something better, that
nobody's ever done before; on the other hand when we make mistakes we're
intellectually honest about the mistakes, we face up to it and we quickly
make changes. So, if you think about NV30 there were many things that we
did
right and there were many things that we did wrong and this triggered an
avalanche of changes in the company and that's what got you 6800. On 6800
I
would say that there were more things that we got right that were wrong,
and
with 7800 [G70] even more is right.
B3D: With the G70 graphics chip itself you've decided to use the 110nm
process and with over 300 million transistors lends itself to a fairly
large
die size; what are the reasons for not going with 90nm at this point in
time?
Jen-Hsun: We thought 110nm was more effective to use as a volume
manufacturing vehicle right now...
B3D: ...even with that size of die?
Jen-Hsun: Oh, yeah. Our customer don't care about size of die.
B3D: ...but they will be concerned about the price...
Jen-Hsun: My customers aren't concerned about my costs, but they are
concerned about my price. I have to sell it too them at a price that makes
sense, they care about power.
B3D: But 7800 has now firmly exceeded the perceived $499 price barrier...
Jen-Hsun: I'm not even too sure there is a perceived barrier at $499. We
need to price it at a level the enthusiasts will buy it at, that's the way
that we think about pricing. We think about the pricing of this in the
same
way that Sony thinks about Playstation 3's - its not about how much is
costs, its about what is the price it needs to sell at, and we need to
figure out how to make money underneath that.
B3D: So where to do you see 90nm now then?
Jen-Hsun: We're taping out a whole lot of 90nm chips.
B3D: I'm thinking that you'll be going to TSMC's 90nm process; what's the
relation****p with IBM at the moment?
Jen-Hsun: Yes, we're using 110nm and 90nm at TSMC. The relation****p with
IBM is fine, its just that we've known for a long time that they'll be
reducing the amount of foundry work they do. So we've been moving the
majority of our foundry business to other partners - primarily TSMC, some
UMC and some Charter. UMC and Charter have really good low power processes
so we tend to give them the mobile business, but they also have decent
generic high speed logic.
B3D: One of the things that we've noticed is that previously you've fairly
consistently been refre****ng products every 6 months, but in this instance
its been 14 months since there has been a high end refresh at all - do you
envisage seeing the cycle times lengthening more and more?
Jen-Hsun: Just a little. It lengthens and shortens based on how quickly we
innovate and how quickly we design, but also based on the cycle times of
fabs. Just the theoretical throughput for chips at 90nm is up to 70, 80 or
90 days, whereas it used to be about 40 days - so we've already added
about
a month and a half into the cycle time of the chips even if we didn't
change
anything.
B3D: The way things are looking at the moment is that if one chip takes
two
spins, you are already encroaching on your refresh product...
Jen-Hsun: That's exactly right. But, you know, a lot of people say that
the
GPU business is really, really tough and yeah, so is F1 Racing, so is Indy
-
you're out there, driving at the wall at 230MPH and people say "Boy,
that's
really dangerous", but shucks, that's our s****t and this is what we love
doing. We wouldn't go so fast if our customers didn't want it - our
customers want this and our end users want to keep getting that better and
better experience and we've got to give it to them.
B3D: With the way things are at the moment it appears that even as we
speak
at the launch of your latest generation of graphics processor you must be
somewhere around the taping out point for the refresh of this and around
being RTL complete on the next generation architecture?
Jen-Hsun: Yeah, its something like that. Its kind of like a machine gun
where you have more than one bullet in flight.
SLI does bring an interesting dynamic though, as its possible that with
SLI
end users can upgrade by just buying another card and doubling up the
performance that way. This might have the impact of slowing down refreshes
and if it does then we'll slow down, but we don't see that happening.
B3D: I would say there are three elements that are having an influence on
product times: the cycle times for a given process, the development times
for new processes, which is also slowing down as well, and the game
development time. The length of time its taking to design a game an issue
to
you, because we're hearing developer talk about 2-3 years of development
time, and possibly even more now that more content and realism is demanded
of a game?
Jen-Hsun: There's one dynamic where game development is becoming harder, I
don't want to say longer, just harder for now - it is more technologically
challenging and it is harder to create the art. On the other hand the
tools
are getting better; we're seeing things like programmable shading tools
and
UnrealEngine and its editors - all of these types of tools are getting
better. I happen to believe that in the long term game development cycles
can actually shorten. Now, saying that can sound kind of ridiculous, but
let
me give you one anecdote...
I started designing my first chip when I was 20 years old, some 20 years
ago, and at the time the cycle time of that chip was literally three years
-
there is no chip today that will last three years! That chip was 20,000
transistors - if I gave a new graduate now 20,000 transistors he wouldn't
take the job! What took me three years to do, kids can now do in a
weekend,
so we are more productive now and the reasons for that is because of high
level design languages, synthesizers, compositors.
B3D: But to a certain extent, the problem you can see with that model is
that the onus is on the powerful developers and publishers because someone
has to be making those tools and selling them, and they aren't going to be
cheap which would suggest that it still lends itself to the EA's or Epic's
of this world, pu****ng out the smaller developers...
Jen-Hsun: Maybe, maybe. However, another anecdote... The number of chip
designer companies in the world has dramatically expanded in the last
twenty
years, whereas only TI [Texas Instruments], Intel, AMD, Fairchild,
designed
chip twenty years ago - give me an example of who doesn't today!
So, where common sense would say that these things are getting more
expensive and harder to do the fact of history is that smart people come
along, create tools, increase the level of abstraction and with better
technologies divorce the technology from the artists. What Mark Rein says
is
incredibly im****tant, that you don't want everybody to be a scientist and
engineer and figure out how light reflects or refracts off a particular
material, you just want him to think about what he wants to make and as
soon
as you separate that limitation all of a sudden everyone becomes much,
much
more productive.
B3D: One of the key things NVIDIA have been talking about since the
introduction of NV40 is High Dynamic Range Rendering (HDR). There appears
to
be a bit of a disconnect here though as whenever you deal with "higher
quality pixels" they are always going to need more bandwidth and although
you can trade that off to a certain extent by doing more internally in the
shaders, making fewer external calls, there is still a large difference
between the very low end and high end in terms of the bandwidth available
to
there point where the current generation doesn't have HDR capabilities at
the low end, but does on the mainstream and high end parts. How can the
disconnect between the very high and and low end be managed in pu****ng
photorealism two all parts of the spectrum?
Jen-Hsun: That's actually one of the beauties of HDR as you can rendering
in same game two different modes with one mode being photo-real. And, when
you say photo-real, I actually mean professional quality photo-real. Ansel
Adams is famous not because he took a bunch of pictures of rocks, but
because he was able to get the dynamic range of photographs and he
displayed
it in monochrome. But the dynamic range that he captures is unnatural, it
beyond the dynamic range of the camera, and that his genius - he was able
to
take a camera that had basically 8, 9 or 10-bits of dynamic range, lets
say,
but somehow get a lot more than that and that's the kind of images that
we're going to express.
B3D: One of the key inflection points we've had over the past year is PCI
Express; what did this mean to NVIDIA?
Jen-Hsun: High speed interfaces between the graphics chip and the CPU are
so incredibly im****tant and obviously one thing that it enabled us to do
was
SLI and TurboCache was another example of the type of benefit that you can
get. The PCI Express bandwidth is at a level now that is consistent with a
64-bit DDR level framebuffer and so its pretty good if you just thought of
it as a framebuffer interface.
When you have something that fast you can really re-think everything that
you do. You know, if you could fly at 2000MPH or get your car to be able
to
fly you would think about the way you work, you would think about the way
you travel and you'd do everything a little bit differently - PCI Express,
in a way, is like that and you really have to think, fundamentally, about
the way that you architect and what you can do and when we asked ourselves
those questions that's when SLI and TurboCache came along.
B3D: That's enabled you to do something with the PCI Express, but what
about the content side of things? What would you expect to see here and
are
you doing anything to push it forward?
Jen-Hsun: Well, we invest more in content developers than just about
anybody
in the world and the invention of Cg, working with Microsoft on HLSL, the
notion of programmable shaders, all the SDK's that we create, all the
shader
programs we write and give away, all the Developer Technology and
Relations
engineers we have to help them build better content - we deeply care about
the game developers and them being able to use the technology.
Greater bandwidth gives you the ability to download textures more quickly
and that's a big, big, deal and at some level the bus bandwidth limits the
amount of geometry you can send over the bus so there is a level of
fidelity
that you are at a theoretical limit to be able to achieve with every bus,
so
the level of fidelity of geometry and textures is limited by each
generation
[of the host interface]. Now, of course we can get clever with things like
Perlin Noise shader generators, to be able to synthetically create marble,
wood and things like that - procedural textures - and that's en example of
computing replacing content and we'll be able to do those things, but
fundamentally there is still that theoretical limit.
B3D: There's long since been elements of graphics research which haven't
made their way on to consumer PC graphics products yet, one of them being
geometry compression and another being hardware tessellation - which could
be seen as another form of geometry compression...
Jen-Hsun: Higher Order Surfaces, Procedural Textures, they are all forms
of
compression and we're a big fan of high order surfaces. I don't know if
you
remember NV1, but is was based on higher order surfaces because there was
a
concern that the PC would not have a floating point unit and that the PCI
bus that came to be wasn't going to be available and we'd be stuck on the
local bus, which is why we invented curved surfaces. It turned out that
our
tools just weren't ready for it then, but sub-division surfaces is going
to
happen. In games today there is often polygonal detail that we'd like not
to
see and subdivision surfaces will eliminate all of them.
B3D: So how long do you think the PC will remain as your primary business
model?
Jen-Hsun: The way to think about the PC is really our only model for
deployment of GPU technology and products. The game consoles will always
be
a closed system and so we see that as a primary model for IP. With Cell
Phones we're going to see models from both - IP and stand alone products.
Each industry has its own way of working and in the broadest way of
thinking
about it we fundamentally invest in the technology and we will monetise
that
investment in a lot of different ways, whether its entire boards or
subsystems for workstation, or IP for cell phones and game consoles.
B3D: Presumably the model for cell phones will be the trickle down effect,
i.e. the low end, possibly a number of architectures prior to what
currently
exists on the PC side?
Jen-Hsun: I would only correct that in the sense that because the PC
architectures were designed for speed first you can't use that design at
all. With a PC architecture you design it so that with everything enabled
-
when you are doing texturing, with trilinear filtering, with mip-mapping,
with anisotropic filtering, with alpha blending, Z buffered, and fogged
you
want to still be producing one pixel per clock per pipeline, running at
full
throttle - if you were designing a cell phone chip that's a really bad
architecture and the reason is because you are not always using all of
these
things at once ut even if you turn them off in the old pipeline the
registers for them are there and they are getting clocked and you're
burning
power on that thing whether you like it or not. In the handset market you
want to be much more statistically minded and look at how often people use
each of the elements and only "pay as you turn on" and that architecture
is
totally different. So if you look at out lowest end GPU architecture it
draws a few Watts, but our highest end cell phone architecture draws tens
of
mW's - you can't create those new order of magnitude without a
fundamentally
different architecture.
B3D: The levels of granularity in power management is now making its way
back up to the desktop space though isn't it?
Jen-Hsun: Yeah, its an issue of priorities. In the case of cell phones it
power first, price second and performance third, and even there I would
say
its features first and performance last.
B3D: In the case of your PC graphics, though, the desktop parts are also
used for laptops, which power must be one of the primary considerations?
Jen-Hsun: Let me give you a counter to that. In mobile, meeting a power
envelope is first - being absolutely the lowest power is not im****tant,
and
there the difference. In notebook PC's we call it thermal design points,
TDP's - if its a 20W TDP that's what you need to hit; in a cell phone its
lowest power is best and the reason for that is the lowest power gets you
the longest play time.
B3D: You mentioned that you see the console model being IP based, and it
seems as though XBOX may be the last of the "per chip" cost basis, but
initially when the noises about the next generation XBOX were coming
around
you appeared to say that you didn't really like that model then...
Jen-Hsun: Oh, no, no, no, that's not the case; I didn't like the absolute
amount in that model! Even though we are not delivering a physical
product,
but a set of software, effectively, and IP deal should still generate the
right returns.
B3D: So would would it be correct in thinking that its not so much about
the
return on the royalties but more about the cost of what you would have to
do
to meet the demands?
Jen-Hsun: The simple math is this: we have a finite number of world class
engineers and I currently have them all fully loaded doing projects and if
they were doing another project then the wouldn't be doing a project I
currently have them doing. the projects that I currently have for them
have
some estimated return and if a new project has a substantially lower
estimated return, relative to what they are currently doing, then you're
loosing money. So, just because something pays doesn't mean your not
loosing
money.
B3D: I understand you may not be able to talk too much about it right
now,
but I'm just wondering how you expect Playstation 3 graphics to differ
from
what we see here today?
Jen-Hsun: Although I can't talk about it too much there are some things
we
know about the PS3. First, its got a Cell processor, and that affects
graphics a lot, and second the front-side-bus is 7 times faster than PCI
Express.
B3D: Do you think there'll ever be a push to see something like that in
the
PC space?
Jen-Hsun: Well the PC industry is going to have to work pretty hard, but
that's what our job is - compete against ourselves. If I can't deliver
something that is at least better in our next generation then we'd be out
of
business.
B3D: Where do you see NVIDIA expanding next?
Jen-Hsun: Over the course of the last 13 years we've really expanded
peoples notion that GPU is much, much more than just PC graphics. When we
first started the company people used to call us a VGA controller company,
then a PC graphics company, but now they call us a GPU company, which is
kind of nice. Now what I need to do is help people understand that the GPU
is much more versatile than just PC's and computing and that I'm going to
have a GPU in my cell phone, a GPU in my handheld, a GPU in my television.
That's where we're going to expand, we're going to take GPU's into all of
these rich displays. At the highest level, the way to think about it is
that
when we started the company the number of pixels that were manufactured on
earth then verses now is extraordinary - in 1993 there were no LCD panels,
and you certainly weren't going to fit a CRT in your pocket; that's our
market op****tunity!
----------------------
We'd like to thank Jen-Hsun Huang for his time and giving us a little
insight into his thinking and understanding how NVIDIA is driven.
|
