Reverse engineering: more than NVIDIA deserves?

Open source software is written by people with an itch to scratch, and -- unfortunately -- no
amount of philosophical reasoning will provide free graphics drivers for people on NVIDIA
hardware. Thus, the only way these people are going to get high end consumer hardware to work
on their boxes, is to either use the proprietary driver or to write drivers themselves.

It is true that this effort seems to punish NVIDIA's competitors who are either open source's
poster child in this area (intel) or making noises of gradually improving a previously bad
situation (ATI/AMD). But, to resolve this apparent problem, all one has to do is to ask, how
we got into the current state. Why do people have an itch to scratch, that we rather they
wouldn't?

It's because a short while ago, when the nouveau effort started, there was no credible
alternative for high end consumer graphics cards with open drivers. Intel, as the article
correctly points out, occupies the low end of on-board graphics refusing to offer a
stand-alone version of their products that would be able to compete with NVIDIA's or ATI's
offerings. AS far as I know they are planning to enter that market with the Larrabee cards,
planned for release in 2010. As for ATI / AMD there is, to my knowledge, still no open/free
driver for any of their recent chips, although we have been hearing rumors that things are
going to get better Real Soon Now (TM).

So in effect, the punishment that intel and ATI may be feeling now is due to the fact that
they are too late in providing a solution for people's needs in high power GPUs. Had any such
alternative existed say two years ago, I'm sure that people would long have started voting
with their wallets. I know I would have. Instead, lots of intelligent people have bought
NVIDIA hardware. And if you have lots intelligent people sharing the same itch to scratch,
eventually you will get a solution to that problem...

You know, before complaining about NVIDIA's lack of cooperation, we should look at the
respective track records of NVIDIA's closed-source driver team and of the free software
community. And if you take an honest, hard look, you will see that NVIDIA's closed-source
Linux drivers beat *all* existing free drivers, for any piece of graphics hardware, in terms
of how much performance they can extract from the hardware, their reliability, the speed with
which support is added for new hardware, and the rate at which new features are implemented.

I hope that one day, we the community will finally start making drivers that actually work on
all hardware that we claim them to work on, that really do properly support all features that
we claim to support, and that don't cause our machines to lock up for no good reason. But
until that day comes, instead of complaining about NVIDIA, we should be thanking them for
releasing an excellent (yet unfortunately closed-source) driver.

See, I have to disagree about this is "rewarding" NVIDIA or hurting other 
projects.  It's the old "scratch an itch" thing: a group of people like 
the NVIDIA hardware and wanted a free driver so they're working on it.  
Eight people out of the entire community isn't really a resource suck that 
I see hurting Intel and ATI video cards.  People are free to start their 
own projects to try to get more people to work on Intel and ATI hardware.  

I think that what's actually helped NVIDIA to get the interest that it 
does in the community is that it provided binary-only drivers that 
actually WORKED.  Yes, this is an odd situation, but for years they had 
the only cards with almost identical 3D acceleration as Windows, while ATI 
kept putting out that joking monstrosity that they called drivers.  Thus 
more people in the community bought NVIDIA hardware and so it got a large 
user base.  Then comes Nouveau which is doing a great job at making 
totally open drivers since there was a critical mass in the community.  I 
think that a large number of users annoyed with the binary only driver is 
what brought this about.

Linux has more open drivers than any other OS because people decided to 
write free drivers to scratch their own itch.  If we end up with open 
drivers that make the NVIDIA cards work as well as the binary drivers do, 
so what?  People are free to buy whatever hardware they want, and it's 
never a bad thing to have drivers out there.  Besides, instead of spending 
time and energy being grumpy and saying it's rewarding NVIDIA and we 
shouldn't do it, I think we as a community should be proud that these 
eight developers are making an open driver in spite of the lack of 
documentation and corporate help.  To me that says a lot about what's 
often criticized as a group of non-professional hackers from the 
Internets ;)

One can only hope that when sufficient information about the chips gets reverse engineered and
released, Nvidia will open up parts of their driver and/or devote part of their team (which,
from what I'm hearing, they're doing a pretty good job) to the free driver.

IIRC, something similar happened with the NForce driver.

I just don't see any way of telling someone not to scratch an itch.. or even worse enforce
people not to do so. I mean how? Revoke CVS, email privs, make a McCarthy blacklist from
anyone who works on Nouveau? [Yes hyperbole ontop of hyperbole..]

Yes it sucks that NVIDIA has basically screwed Linux over and over again.. and gets some
people to fix their stuff. But in some ways you can say the same about most FLOSS. If RMS
hadn't been pissed about proprietary Emacs and Linus hadn't been peeved about the state of
proprietary Unix.. Would we have openssh if the OpenBSD team had not been peeved about the
propietary nature of later SSH's?  Would we have the amount of Free Software that we have
today?

So I think it is a two edged sword that we need to think of how to better reward companies
that play with us versus looking at how to punish those who 'reward' those who don't.

It depends what you think the pain points for NVIDIA actually are.  Having done similar sorts
of things in the past, I think it's mainly:

1. Running an entire engineering team just to keep up with the latest greatest API breakage so
the binary module will still work
2. Having to make sure the distribution model works on a huge variety of vanilla and vendor
kernels (this becomes a huge validation and installation effort).

Having a reverse engineered Nouveau driver does nothing to address either of those.
Furthermore, for an evolving hardware platform (which the NVIDIA graphics cards are), reverse
engineered drivers are always behind the curve because you actually have to see the hardware
to test and fix things.

Conversely, I don't think Intel (and eventually AMD) actually expect the community to take
over their drivers.  In many ways, companies like to keep control of their products, so they
tend to hire the driver writers who work on them anyway.

The key benefits companies gain from open sourcing are

1. Community assistance:  bug fixes, end test cases (even Intel can't test all their chip
configs)
2. Freedom from distribution pain ... it's just upstream and the distros ship it
automatically.
3. Franchise in the development model.  Imagine what would happen if Nvidia sent an email to
x.org asking for changes in the X server to support their hardware ... now imagine what
happens when Intel does it ...

So Nvidia still has all of the costs and none of the benefits even with the Nouveau driver;
it's hardly a gift to them (in fact, if they actually look at it it might end up being a
trojan horse because they could get GPL contamination of their binary blob).

NVIDIA's 16March'07 10k statment says (p9):

"We expect substantial competition from Intel's publicized focus on moving to selling platform
solutions dominated by Intel products, such as when Intel achieved success with its Centrino
platform solution. ... If Intel and AMD continue to pursue these initiatives, we may not be
able to successfully compete in these segments."

Frankly, the lack of open-source driver development going on in the ATI and Intel camps are
their problem, maybe even their own fault.  I'm not sure how depriving NVidia of open-source
drivers helps this situation.

NVidia has put in considerable effort not only to produce good video hardware, but also good
video driver software, on Linux and on other platforms.  From Matrox, NVidia, ATI, Intel, and
S3, NVidia is the only one I know of that has interesting performance combined with drivers
that actually work on Linux and other platforms.

NVidia have decoupled their video drivers from the kernel interfaces, and provided a
sophisticated installer that downloads and compiles the kernel interface, the Xorg interface,
and the video hardware driver for the specific circumstances of the end user.  This is a big
deal--just about every other out-of-tree device driver I've seen on Linux requires upgrading
the hardware-specific code at the same time as the kernel interface code, which means kernel
upgrades imply device-specific hardware driver upgrades at the same time, which provides an
extra and often unnecessary source of regressions and bugs.  NVidia has solved this problem,
and even provided a mechanism for updating the kernel interface layer over the network at
install time.

About the only thing wrong with the NVidia drivers is the license and lack of source code--in
every other respect, they are a model for how out-of-tree drivers should work on Linux.

If open source developers are going to create an open-source video driver, it only makes sense
to choose the most popular chipset as the target.  The only thing that would change this
choice is the pre-existence of high-quality open-source drivers for the target, which of
course is not the case for NVidia.

Intel and ATI keep alleging that they are developing or have released open-source drivers,
while NVidia has consistently asserted that there will be no open-source drivers for their
chipsets.  Why should the community bother reinventing the wheel when there's a signficant
chipset in the field with no open source drivers at all, not even on the horizon?  Vaporware
FUD is a double-edged sword (even if the wares are not really vapor, and the FUD isn't
intentional).

If you're going to go through the effort of contributing driver code for a chipset, you want
the chipset to still be around when you're done.  Also, you need a large pool of people to
choose developers and testers from, so the biggest existing population of users is a good
place to start.

Certainly I hope no one is proposing that people should refrain altogether from developing
open-source video drivers; however, suggesting that people develop open-source video drivers
for an obscure niche market chipset, or a chipset vendor who might leave the market in the
near future, is the same as suggesting they don't work on a video driver at all.

So this article is missing some info...

1. Intel does get community development on its driver. Red Hat have done a lot of work on the
Intel driver to bring up new X features. Having an open source driver means we can develop new
core X features and things like kernel modesetting. Before I joined Red Hat I wrote a lot of
the kernel modesetting driver for Intel hardware, I also wrote some support for driving the
external chips on those machines. DRI2/modesetting/TTM is all happening on Intel hardware
primarily.

2. AMD have employed Alex Deucher to work on their open drivers, I also contribute code and we
also get major contributions from the community (rs690 support for 3D was a contributor we
have never seen before). They are also paying SuSE to work on radeonhd. However I'm not sure
how much community contribution radeonhd gets. So I don't think AMD have a major advantage
over Intel in this area, they had more community contributors since they stopped helping the
open source community, but they had no real support model for their cards for their enterprise
distros except fglrx. So currently the AMD area is still in major flux and cannot be used as
an example of a stable situation.

3. NVIDIA only care about money. The reasons for Intel/AMD to open up their drivers won't work
on NVIDIA, NVIDIA don't have major money to lose in other areas due to not having open source
GPU products, Intel and AMD do. The biggest problem NVIDIA will have will when the community
developed driver starts being shipped in enterprise distros their enterprise customers will
start to push back on them to support it instead of supporting the nv driver. This happened
before with network drivers. How this will work out is unknown at present, but I think
developing the nouveau driver is not in any way helping NVIDIA, their customers are not you
and me, they are Dell and HP, until they put pressure on them anything the community does is
irrelevant.

I like that people are working on free software Nvidia drivers.. those people are doing some
cool stuff and are bringing in much needed improvements to the whole open source driver stack
that are going to benefit ATI and Intel driver improvements.

I like what they are doing, I appreciate what they have accomplished and what they continue to
do.... So don't get me wrong when I say this:


I'll use Nvidia hardware if I have a need for it. So they'll get my money, and that's all they
care about and I know they couldn't give two craps about my opinion of them. But otherwise I
will avoid them like the plague (and I do.)

Still I tell people that if they need 3D performance in Linux that they should buy Nvidia.
Why? Because that's the truth. And unfortunately Nouveau just isn't going to match the
performance or stability of the Nvidia hardware in any sort of effective timeframe. It's just
very unfortunate. Maybe I am wrong, hopefully I am wrong.

I feel the efforts on creating a open source Nvidia driver is wasted on the end users. So if
the developer's goal is to make open source drivers for people to use on their nvidia
hardware, then that's a total bust from the beginning. They are making a mistake.

Conversely, If their goal is to learn hardware, learn software, and personal satisfaction of
working with very modern hardware (or other things along that line), then that's a entirely
different thing and I wish them all the luck in the world.

The most Nouveau will ever get used for, at this point or any other point in the future (until
the supported hardware is so obsolete that Nvidia cease caring about the people still running
it), is to let Ubuntu boot up with 3D support long enough so that users can click the 'yes'
button to install proprietary drivers. 

Why?

* People who care about open source drivers are going to purchase Intel or AMD hardware. 
(preferably Intel, IMO, since AMD has released no documentation on 3D stuff and has released
no code for it.. while Intel has released full documentation and full source code)

* People who do end up purchasing Nvidia hardware for Linux either do not care about open
source drivers (or don't care about them that much) or require the performance Nvidia offers
for their purposes, and thusly will install the nvidia drivers due over the open source driver
in almost every case. 

By buying Nvidia hardware in the first place they are voting pro-proprietary driver (in the
eyes of the Nvidia corporation) in the strongest manner available to a end user.

* New users to Linux who cluelessly (through no fault of their own) use machines with Nvidia
hardware will simply not have the desire to keep using the inferior performing driver and will
want the same level of performance they are used to in Windows.


That's all. 


I want it to be known that I do not have a problem with them working on it. I don't think it's
going to 'reward' Nvidia in any way.. I don't think that Nvidia gives a crap, and even if they
did I don't. Even if it _totally_ benefits Nvidia, I still ain't going to worry about it. Life
is to short to avoid doing good things because it may benefit somebody you have a problem
with. 


From a purely selfish standpoint I would like them to devote their time to supporting _my_
hardware (Intel GMA stuff), but I can't see how that fact would excite them or motivate them
in any fashion. I mean I would _love_ to have faster graphics and be able to tell people to
purchase AMD or Intel (for better compatibility and stability) without the caveats of low
performance with OSS drivers.

So my vote is for them to drop everything that they are doing and work totally for the best
Intel driver humanly possible in the shortest time frame. ( They should be happy, and probably
are, that my vote has utterly no impact or control over anything they do, nor should it. )

While I see the point of the article, I think some mitigation is missing.  Nouveau people are
probably great at what they do, but they can only support past Nvidia cards.  From my meagre
understanding of how driver development works (and from what I read about radeonhd), each new
generation of hardware needs a very different driver.  So future Nvidia cards won't be
supported right after they reach market, while it's quite possible that ATI and Intel cards
are.

So even if there eventually is a working driver for new Nvidia cards, there will still be an
advantage for the other brands: the other cards will be supported in-tree (and out of the box)
right now, rather than a year later when the card is obsolete (from the marketing point of
view — I'm quite happy with my five year old Radeon 9200).

There is a flip side to this: if the nouveau people weren't reverse-engineering the nvidia
driver, there'd be less incentive for Intel/ATI to do the right thing. 

Personally, I'd love to buy an AGP (or PCI-X) card from Intel, but there's none available. And
I do need a dual-head card, with DVI rather than VGA.
As for ATI, have had endless trouble with fglrx, and am now running my laptop with the vesa
driver. Nvidia's driver is a pain to install, but at least it usually works reasonably well. 

Also, don't forget the Open Graphics Project.

I'm wondering, is there anyone who *doesn't* play 3D games who even cares about this?

I've been using Intel's open-source drivers for years without any problems. I'm in the process
of buying a new notebook, and the first thing I did was eliminate from consideration anything
that doesn't have Intel GMA X3100 graphics.  That still leaves me with a lot of good systems
to choose from.

I guess if I were a gamer I would just dual boot Windows.  If you're going to use proprietary
software for something as critical as a graphics driver, why not go whole-hog and use a
proprietary OS as well?

Off-topic: what's the status of S3 Unichrome and such drivers on on-board Via chipsets?

How (un)cooperative is Via?

There's an open driver for Intel, and the AMD/ATI people are slowly delivering specs - but
where exactly does the contributed drivers/specs plug in?

Until recently a little of it went into the kernel, and then a duplication of effort in X,
with some 2D stuff completely divorced from the 3D stuff that went into an opengl
implementation. And using both frame buffer driver and X.org driver for the same card was
discouraged (sometimes or always? dunno).

Now stuff are evolving into a coherent infrastructure, with Gallium 3D picking up the 3D
interface - and kernel duplication slowly migrating out of X.org (with the vision of not
having to run X as root some time in the future, yay :). If this infrastructure is built
primarily around one hardware implementation, it may turn out not to be as useful as we are
currently hoping. So the more, the merrier. 

I think Nouveau development isn't detracting from the other free implementations, since it
adds in the common parts of free 3D graphics and since it raises the general level of
expertise (sp?). Do we know for a fact that these developers would have been working on
Intel/ATI drivers, if there had been no Nouveau project? (Well, I mean apart from David
Airlie... ;)

I run the nv driver on my laptop and on my desktop.  I use an nvidia card in my laptop because
the Intel chip would steal bus cycles.  I put a spare nvidia card in my desktop machine last
month because the X server driving the Intel chip had begun crashing; thus far it hasn't
crashed with the nv driver.

I don't see any way for the above to be forced into the notion that stopping development on
Noveau would punish Nvidia or help ATI or Intel.  That seems like many interesting and
plausible hypotheses that turn out to fail to connect with anything in the real world.

I am actually planning to buy an upgraded video card since the last 6 months...

Although their drivers are open source and well supported Intel has nothing interesting at the
moment.  And I'm so damm tired of fighting with closed source drivers (either ATI or NVIDIA
particularly with their new legacy ones) that I'm currently waiting for either a working
Nouveau driver and update to a GeForce FX abc or a working Radeon driver and buy an AMD/ATI
xyz.

I'll probably end-up waiting until next Christmas before making a final choice since I presume
one of the two projects (or even both) will have sufficient supported hardware in both 2D &
3D.

I actually dream that, running Linux, I may end-up having that choice to make between the
two... unless nVidia buys-up AMD :)

Keep in mind that both Intel and AMD are bundling 3D accelerators onto new 
processors, and this has been common knowledge for something like a year 
now:
http://hardware.slashdot.org/article.pl?sid=07/03/29/1252203
http://arstechnica.com/news.ars/post/20070211-8810.html

Here's actual product names (Intel's "Nehalem" and AMD's "Fusion") and 
shipping schedules.  It all comes out about a year from now:
http://www.reghardware.co.uk/2007/11/26/intel_cpu_integra...

Ars Technica has a good summary of Intel's follow-up, explicitly comparing 
Intel's plans to IBM/Sony's Cell processor:
http://arstechnica.com/news.ars/post/20070416-intel-offic...
http://arstechnica.com/news.ars/post/20070604-clearing-up...

The basic driving force is that transistor budgets keep going up.  Even 
the venerable Pentium M had 2 megabytes of L2 cache, and Core 2 duo and 
Core 2 Quatro put 2 or 4 entire processors on the die in an attempt to 
soak up the transistor budget and translate that into performance gains 
and profit margins.  But most desktops can't soak up more than a couple of 
processors before it's just not noticeable.  (Servers may be able to, and 
developer workstations can, but most home computers doing email, web 
browsing, watching the occasional youtube video...  The third and fourth 
cores just sit idle, let alone cores 5-8 on an 8 way system.)

Bundling a GPU on the die is something all end user desktops want these 
days, and it just _sucks_up_ transistors.  Now instead of 4 processors 
sharing a pool of L2 cache, you can have 2 processors and a GPU, and your 
average desktop can keep them all running at full speed.  It also means 
you can increase the L2 cache size again (because it's now used as texture 
memory too, so you can go to 16 megs, 32, 64...), and it means that the 
highest clocked hardware in the system is locally processing graphics with 
the fastest possible interconnects with the CPU.  (There isn't even a 
_bus_ between the CPU and GPU anymore.)

Of course if Intel's doing this, AMD is too.  (Why do you think they 
bought ATI?)

Where does that leave Nvidia?  Well, nvidia is trying to stay relevant by 
purchasing a physics engine company and hoping all the game consoles start 
using that:
http://www.pcper.com/article.php?aid=222
http://technologyexpert.blogspot.com/2008/02/geforce-8-gp...

That's really what Nvidia cares about: game consoles.  If they have to 
make a choice between game consoles and PC GPUs, it's pretty clear where 
their money comes from.

They're also trying hard to grab as much volume as they can before Intel 
and AMD bundle GPUs on-die with the CPU, perhaps to get game developers 
hooked on their physics engine before the market for standalone 3D chips 
goes the way of the northbridge:
http://www.reghardware.co.uk/2007/10/29/gpu_market_q3_07/

To which Intel replied:
http://www.reghardware.co.uk/2007/09/17/intel_to_buy_havok/

*shrug*  It's interesting to watch.  The market is driving Intel and 
Nvidia into direct collision (especially as people try to use GPUs as 
general purpose compting devices), and Intel is pretty darn good at what 
it does... 

I think this article is a surprisingly extreme critique of what a bunch of people decide to do
with their own time.

NVidia really isn't getting free drivers out of this. Unless they actually support this
project, they're going to have to keep providing their own drivers. The Nouveau devs benefit
because they're having fun, and the community benefits from open source drivers, but NVidia
gets nothing unless they officially start supporting the Nouveau project. People who choose to
not buy NVidia hardware for lack of openness will still not buy their hardware, and people who
don't care about openness will continue to not care.