NVIDIA Publishes 73k Lines Worth Of 3D Header Files For Fermi Through Ampere GPUs

[mdedetrich]

None of what you said disputes anything I said, and I myself have pointed out the complications involved when I referenced how long it took for AMD to release their stuff. So, I know more about this than you're giving me credit for.
So yes, of course Nvidia has been working on this for years, but they were in no rush to release anything. It is reasonable to assume the hack may have prompted Nvidia to be a little bit more proactive. As you pointed out, what they released isn't much, because there isn't time for them to release anything else. It really oughtn't to take 1.5 years to release something like the GBM interface, so, if push came to shove, that was an easy first step for them to say "see? We're doing something" while it could take many more months for them to release anything else.
So, how about spending a little more time thinking about how the situation works rather than be a dickhead? While I'm not saying Nvidia has catered to the hackers, your arguments against it are falling flat, and your know-it-all attitude just shows arrogance rather than actually critically thinking about why it isn't possible. The funny thing is: the most obvious reason why Nvidia isn't catering to the hackers is one you didn't bring up: Nvidia clearly didn't release enough to satisfy the hackers, and I'm sure they're not going to just sit around for several years for them to meet the condition. However, we don't know what the full conditions are, which is why I brought it up as a possibility.

No its not reasonable to assume that the leak prompted NVidia to release it faster, thats called a conspiracy theory with no basis in reality. NVidia's reasoning for open sourcing their in kernel wrapper is so that their drivers have better compatibility in light of the direction that Linux graphics stack is going with GBM/Wayland along with making GPU issues easier to debug in Linux (since the source is available).

The only thing thing that would effect NVidia's timeline is that, not some code leaks by some open source hactivists.

And in case its not already clear, the reason why your are projecting is that its evident by the content and tone of your post is that you really want to believe that open source "activism" (or w/e. you want to call it) is having some impact on NVidia when in fact NVidia doesn't give a flying f**k about it. NVidia is not an open source company like Red Hat, which means that open source is on the bottom of the list of their concerns.

[c117152]

the vbios literally tells you that, but for that you need to RE the vbios.

It was my impression the newer bios firmwares are encrypted (as opposed to just being signed) so they can't be dumped into something you can reverse as is... Good to know there's exceptions.

[arQon]

(snip)

1) You need to work on your reading comprehension.
2) That was so detached from reality that even if you *had* understood what I wrote I'm not sure we'd be able to turn this into a useful conversation.

The change of tune in your last post does help, but you're still *really* reaching, and the "I was just trolling" line isn't doing you any favors. If you'd called them "crazy conspiracy theories" in the first place, you'd have found a more receptive audience. Since you put them forward as things you thought were probable, let alone credible, you can hardly claim it's somehow "unfair" if people rip them apart. Both that, and you apparently taking those responses personally, are on you though.

[oiaohm]

Last I recall AMD’s consumer GPUs don’t support 30-bit color using OpenGL. While Nvidia enabled support with their Studio drivers for consumer GPUs in 2019. At the end of the day Nvidia GPUs have much more functionality with Studio drivers than AMD’s open source drivers when it comes to professional work loads.

That one been a funny one of horrible. Since 2019 by registry tweek under windows you were able to make AMD consumer cards go 30-bit color. Mesa opengl on Linux has support 30 bit color on Amd stuff for while. Issue been 30 bit color with Vulkan not working that got recently got fixed.

[oiaohm]

So linux should pay to get the license? Ad what about AMD and intel?

AMD and Intel both have used signed firmware for GPU parts. Both provided the signed firmware to the Linux kernel firmware project free of charge under license of their choosing for the hardware that needs it.

https://git.kernel.org/pub/scm/linux/kernel/git/firmware/linux-firmware.git/tree/LICENSE.i915

Above is the intel license on there firmware files. Notice you are not allowed to modify the firmware. You cannot create new firmware for the Intel hardware. But you can legally ship the firmware to anyone who need it so their hardware works.

Yes that is a copyright license written for firmware. Do note that Intel makes it against the rules to use the firmware provided to Linux kernel to be used with a closed source operating system. So you cannot take the firmware from the Linux kernel firmware tree and put it in the Intel closed source drivers on windows and be legal. But all the open source prototype operating systems that exist can use the intel firmware.

Some of the licenses in the Linux kernel firmware list say that the firmware must only be used with X vendors hardware.

Yes there are a few restrictions on what is acceptable license to put stuff in the firmware project. Big one is not hindering distribution and not blocking out any open source operating system. Vendors are allowed to block out closed source operating systems.

Nvidia has been the odd one out here not providing their firmware files to their hardware. The reality here the Linux world mostly would not care if Nvidia provided their signed firmware under a suitable license that it could be shipped in the Linux kernel firmware project. Signed alone is not the problem.

The license Nvidia provides their firmware under is the core problem. The firmware being signed removes means to work around that license problem by reverse engineering and making a replacement.

Something else to remember when you buy the hardware you paid Nvidia. The idea of Linux should pay this is kind of asking Linux users to pay twice. The reality this is a license problem. Yes this could be patent problem depending on who Nvidia has bought patents from.

[WannaBeOCer]

Every action they have ever taken supports that interpretation. What world have you been living on?

Plus the reasoning is quite logical. Many of the features they have made Quadro only are driver features, not hardware features. If open source drivers would become good enough the market segmentation that have been profitting from with Quadro hardware wouldn't work anymore.

They have hardware market segmentation implemented in software alone.

AMD/Nvidia have been locking features to professional cards for years. For example AMD reduced the bandwidth of the Radeon VII’s FP64 performance to 1:4. Nvidia’s consumer GPUs tensor performance when using FP16 w/FP32 Accumulate is reduced to .5x compared to 1x on the Titan RTX/Pro cards. Lack of GPU virtualization from both on consumer GPUs but there was a recent hack for Nvidia that allows consumer GPUs to work.

Last I recall AMD’s consumer GPUs don’t support 30-bit color using OpenGL. While Nvidia enabled support with their Studio drivers for consumer GPUs in 2019. At the end of the day Nvidia GPUs have much more functionality with Studio drivers than AMD’s open source drivers when it comes to professional work loads.

[scottishduck]

Every action they have ever taken supports that interpretation. What world have you been living on?

Plus the reasoning is quite logical. Many of the features they have made Quadro only are driver features, not hardware features. If open source drivers would become good enough the market segmentation that have been profitting from with Quadro hardware wouldn't work anymore.

They have hardware market segmentation implemented in software alone.

Was never aware nvidia was locking higher bandwidth, higher capacity memory and different die config behind drivers. That’s wild, impressive tech.

[MorrisS.]

Its not just that the firmware is signed is the problem. There are lots of firmware the Linux kernel uses that is signed. The problem is the firmware is signed under a copyright license that open source driver developers cannot legally distribute to all end users. Also cannot legally host in many places. Yes downloading the closed source nvidia drivers and extracting the firmware also does not result in a legally useful solution.

The old Nvidia hardware the firmware was not signed so the open source developers were able to work around the firmware license issue by making their own firmware clones. The reality is majority of the time for hardware provided with legally usable firmware open source driver developers don't make their own. Yes legally usable is under a copyright license that that the firmware can be put in the Linux kernel firmware project and can be put in other open source operating systems and can be shipped around by end users as much as they like. Of course the license can say no reverse engineering and the hardware can mandate that the firmware is signed or its not accepting it.

Of course it really useful to have documentation on what interfaces the hardware will provide once the firmware is loaded into the hardware.

So linux should pay to get the license? And what about AMD and intel?

[oiaohm]

My understanding (which could be wrong) is that the NVidia hardware requires the firmware uploaded to be signed with a key. A key which isn't known publicly. So that means the open source drivers can't create their own firmware which does what's necessary, and the only option is either cracking that key or using the firmware that nvidia provides in their proprietary drivers.

Its not just that the firmware is signed is the problem. There are lots of firmware the Linux kernel uses that is signed. The problem is the firmware is signed under a copyright license that open source driver developers cannot legally distribute to all end users. Also cannot legally host in many places. Yes downloading the closed source nvidia drivers and extracting the firmware also does not result in a legally useful solution.

The old Nvidia hardware the firmware was not signed so the open source developers were able to work around the firmware license issue by making their own firmware clones. The reality is majority of the time for hardware provided with legally usable firmware open source driver developers don't make their own. Yes legally usable is under a copyright license that that the firmware can be put in the Linux kernel firmware project and can be put in other open source operating systems and can be shipped around by end users as much as they like. Of course the license can say no reverse engineering and the hardware can mandate that the firmware is signed or its not accepting it.

Of course it really useful to have documentation on what interfaces the hardware will provide once the firmware is loaded into the hardware.

[oiaohm]

Ontop of that, what NVidia is releasing isn't of any real use to anyone about from Linux kernel dev's. All they open sourced is the interface for GBM for their GPU's, the actual drivers are all still proprietary blobs that sit in userland.

No there is more too this. Not just for Linux kernel developers. You have to remember BSD and haiku-os and other open source kernel developers are interested in this information so Nvidia GPU do in fact fire up right. The Linux kernel graphics drivers are used as reference for many operating systems that are not Linux. Yes these are smaller operating systems.

NVidia basically implemented their own in kernel wrapper and moved code from the proprietary driver into their GSP chip.

This is not quite going to plan. Lot of high end GPU users to really upset by GSP being used to lock out cryptocoin mining that result in hitting there data processing by mistake and are now demanding the right to run their own code in the GSP chip or at least totally disable it. At least totally disabled GSP has predictable performance overhead vs doing something tripping crytocoin mining alarm so getting magical not explained slow down that stuff you over with allocated processing time. Think research super computers here where people only get so much allocated time to do their processing and if the party before you tripped the GPU into slow mode then it was still in slow mode when you attempted to do your processing are you not going to be pissed off baddy to find out what you expected to be processed by X time for Y presentation is not processed. (being faster is not the only fact here reliability is as well)

The result of the GSP chip now is a duplication between the proprietary driver and the GSP chip. Yes also you have high end wanting to be able to fire the GPU up with validated code now as well to make sure they are not going to have random performance events.