NVIDIA Open-Sources TensorRT Library Components

[skeevy420]

Oh Michael, Michael
Oopps, I did tripped it again.
I typed some dumb stuff, got lost in the post
Oh Michael, Michael

[karolherbst]

If you use their SDK you have to defend Nvidia every time we post bad things about the SDK or Nvidia in general.

It's funny how you quote something, totally didn't understand it and then fail to summarize it...

[karolherbst]

oh uhhh... it's a troll.. my bad. Sorry for replying

[skeevy420]

oh uhhh... it's a troll.. my bad. Sorry for replying

I made a joke and you replied by name calling. You're the one trolling in this context.

[betam4x]

they throw you a carrot, but they still keep the leash.

Disclaimer: I own primarily NVIDIA GPUs, so I'm no AMD fanboy.

That out of the way, of course they aren't going to start open sourcing stuff. Then AMD and Intel will realize how stupid easy it is to integrate on current gen hardware. As it is, NVIDIA already does a LOT of stuff in software. Even the DirectX 12 implementations for certain GPUs are done mostly by software. That is one of the reasons why NVIDIA drivers are proprietary on Linux today and you can't find a decent open source equivalent. If NVIDIA created an open source driver, all those dirty little driver secrets would be spilled out.

Note that what I've said above is backed by multiple articles around the web, but I don't have time to find them all and link them here. Just Google it. Everything from scheduling to API implementations is primarily implemented at the driver or development library level. With newer cards this has changed a bit, but much of the stuff is _still_ software. The reason this matters is AMD has had hardware implementations for DirectX 12 and Vulkan for many generations. It wouldn't take much to add a 'software' layer onto this stack, implementing any open source code that NVIDIA were to make available. If CUDA were open sourced (assuming it's not patented), AMD would likely have a decent working implementation in short order, one that might even perform better than NVIDIA. As it is people have attempted to come up with CUDA translation layers for AMD cards.

The situation is getting bad for NVIDIA, however. On Linux, AMD is gradually beginning to beat pascal cards (including the 1080ti, which I have in my machine right now) in frame times. In some cases, even average frame rate is higher on a Radeon VII. Next-gen NAVI products (not the ones launching on the 7th, but the ones launching early next year) are going to continue making this worse for NVIDIA. Linux may be a niche market for gaming, but it IS a growing market. Eventually NVIDIA is going to have to rethink it's approach. If the whole Wayland debacle has shown us anything, it's that NVIDIA is lagging behind on the Linux front.

[finalzone]

...

In other words, Nvidia is delaying the inevitable after spending time using dirty tactics for nearly a decade with their hairwork, gamework and an attempt to force OEM to exclusively use their products by preventing competition.
Nvidia knew they were at serious disadvantage in term of hardware features found on AMD graphic card they went they way preventing developers to fully take advatange of such functionalities. Majority of software notably PC gaming relies to Nvidia approach rather than effectively use a cross-platform code allowing better hardware optimization (looking at Unigine for specifically use Nvidia approach and limiting others as an example).
What Nvidia showed is the nearly inertia from both developers and vendors unable to either call out their dirty play or expose them.

[llukas]

The situation is getting bad for NVIDIA, however. On Linux, AMD is gradually beginning to beat pascal cards (including the 1080ti, which I have in my machine right now) in frame times. In some cases, even average frame rate is higher on a Radeon VII.

I'm bit puzzled. Are you saying that Radeon VII just started to match 1080ti in performance and present this as positive for AMD? Radeon VII is 7nm part released two years after 1080ti 16nm part. Am I missing something?

[starshipeleven]

I guess open-source isn't the right way of putting it. The last paragraph in this article sums it up:
https://www.anandtech.com/show/5238/...source-kind-of

By that logic, anything that is under NDA is "open source".

[starshipeleven]

There goes my plans for a CUDA enabled rail gun. It needs CUDA to quickly calculate the Coriolis effect for long-range shooting.

"Unless you have an agreement with NVIDIA for this purpose"

So contact a sales rep to arrange a deal.

[betam4x]

I'm bit puzzled. Are you saying that Radeon VII just started to match 1080ti in performance and present this as positive for AMD? Radeon VII is 7nm part released two years after 1080ti 16nm part. Am I missing something?

The Radeon VII is based on a refined (mostly a die shrink) Vega architecture. Therefore it is much slow than NVIDIA present offerings (under Windows). The fact they are catching up with NVIDIA under Linux clearly shows that NVIDIA is falling behind. If you take those same games and run them under Windows with a Radeon VII vs anything 1080/2080 and up, the Radeon VII will struggle to keep up. It's NVIDIA's Linux support that is lacking and is making the Radeon VII a better choice for gamers on Linux. NVIDIA will eventually deal with these issues, but not until it becomes such a horrible problem that AMD or Valve start promoting AMD products over NVIDIA ones. It took quite a while to even get NVIDIA to implement EGL Streams for Wayland. In the end they admitted it needed to be done, so they decided to eventually implement things. However no timeline was given.

NAVI is AMD's first real architecture break in many years. AMD's big NAVI early next year will likely surpass even the 2080ti under Linux, and possibly even under Windows.