Valve Working On Explicit Sync Support For "NVK" NVIDIA Vulkan Driver

[AlanTuring69]

LOL. They could chose Tegra, they had not.

They could do it, they already did bold moves, Microsoft went PPC for the Xbox 360, Sony went Cell for the PS3.

They have not, because Tegra would only be a serious competitor if Microsoft and Sony were making a Switch competitor.



You fail at thinking about the product as a whole. An APU is a product, it's in competition with other APUs, An APU is not in competition with the parts of it.

That's because Nvidia has no APU that can compete against AMD APU those console makers went the AMD way, even if there exists dedicated Nvidia GPU that beats the graphics solution embedded in the AMD APU. A good example is ray tracing: AMD was behind Nvidia on the ray tracing topic went the XS and PS5 were designed, but Nvidia was even not in the competition of APUs to begin with so none of those console features them.

Nvidia dragsters maybe faster than the cars used by people who buy groceries themselves, but those people don't use dragsters to go buy groceries, and game consoles are sold to people who buy groceries themselves.

It's actually because all of this is right that Nvidia started the Tegra work years ago, to begin with, because people at Nvidia know they will never sell dragsters to game console makers. That's also why they even attempted to aquire ARM, because they know the problem is real and they are afraid about this problem: No one will buy them APUs if they don't have APUs to sell to begin with.

No one will buy them an APU only because they have another product that is not an APU that is better in another topic. Tegra actually exists because Nvidia is behind, Tegra is part of their attempts to close the gap. The example of the Switch is a good example of how the Tegra allowed them to keep a foot in the market of game consoles by targetting a lower level, which was a very good strategy for them, and this experience and such market is allowing them to progress on closing the gap more, because they have to close the gap and they know it well.

I'm not saying Nvidia will never be competitive in that market, neither they may be never be chosen in the future, neither they may not even become better in the future. But this is only a possible future. They're still working at closing the gap, and even Intel has now doubled them: The steam deck competitors in the same form factors that are not AMD are Intel.

Asus ROG Ally? AMD. Lenovo Legion GO? AMD. Ayaneo Kun / 2S / 1S? AMD. OneXPlayer Pro 2? AMD. GPD Win 4? AMD. MSI Claw AIM? Intel with Arc graphics… No nvidia around.

All of them are built to run Steam and PC games.

And if you even look outside of the main topic, which is about Valve and Steam interest, if you're OK to not have Steam and look at ARM where Nvidia Tegra may stands a chance, you get the Logitech G Cloud with ARM/Adreno or the Ayn Odin 2 with ARM/Adreno too. Where is Nvidia if it's so better and those makers can avoid PC compatibility and if they already do ARM?

Ah yes, the proof Nvidia Tegra is the APU of choice for game console products in 2024 and later is that Tegra was used in the Nintendo Switch 7 years ago. Even those choosing ARM for making game console products in 2024 don't chose Tegra today, but why? Nothing is preventing them to chose it if it's better for their need! And now you said the Nvidia Tegra was 10 years ahead 10 years ago? 7 years later, the Switch using Tegra still looks like an accident, and 7 years later the competitors are still used instead, even by those doing ARM…



Then why do Nvidia doesn't offer it? Why do game console makers don't buy Nvidia? Are you aware they can? Microsoft already integrated Nvidia for an Xbox in the past, Sony already integrated Nvidia in a PlayStation in the past. They can do it, they already did it, why do they don't do it today if Nvidia is better? No one is forbiddng them to do it!

You talk like if Microsoft, Sony and Valve have chosen AMD but it would have been better if they had chosen Nvidia instead… But saying this doesn't make sense, because they all could chose Nvidia and they didn't do it. One cannot say “Imagine what would be an Xbox Series X with an Nvidia, a PlayStation 5 with an Nvidia, A Steam Deck with an Nvidia, it would be so better !”, because they all could do it and they all did not do it because it was worst.

lol y u even bother with X11 x10 was da real deal all that new stuff jus overcomplicated fluff simple is better bro and dont get me startd on rust such a waste js is where its at no classes no types jus pure code man OOP more like NOOP amirite

[TheLexMachine]

its taken 10 years for sony playstation and xbox to compete with tegra.

Tegra is not a current offering from nvidia. it is even further back from what a current mobile solution would look from that as a DGX is forward.

Tegra has No AI acceleration, DLSS, tensor, RT cores or cuda support - any of the fancy stuff nvidia "bet the company on" over the last ten years.

A new mobile offering from nvidia will likely have similar realtime graphics capabilities as the CGI used to make

https://youtu.be/cSp1dM2Vj48?si=6hdfIOU_aAjmB6Sb

But only if the games for it are nvidia exclusive, because AMD have neither the hardware nor the software to create anything similar.

Nvidia put CUDA and AI features into Tegra some years ago, as their Nvidia SHIELD products have GPUs with CUDA cores and they switched their hardware development to focus on hardware they could sell to the Chinese. Their latest Tegra hardware is primarily used in BTO solutions for industrial/embedded networked tech, like security cameras and other things that benefit from front-end video processing.

[mSparks]

Nvidia put CUDA and AI features into Tegra some years ago, as their Nvidia SHIELD products have GPUs with CUDA cores and they switched their hardware development to focus on hardware they could sell to the Chinese. Their latest Tegra hardware is primarily used in BTO solutions for industrial/embedded networked tech, like security cameras and other things that benefit from front-end video processing.

Hardly surprising
updates to tegra focused on hpc (dont care as much about power consumption, stuff like RT cores are a waste of silicon), that has very different needs than mobile gaming that wants the bare minimum power consumption applied to as many RT cores that will fit in the package.

But that doesnt mean they cant repackage what they have done for mobile gaming, just that so far there wasnt the business case to do so.

In that scenario, AMDs most expensive silicon doesnt come close to Nvidias cheapest

[tenchrio]

Hardly surprising
updates to tegra focused on hpc (dont care as much about power consumption, stuff like RT cores are a waste of silicon), that has very different needs than mobile gaming that wants the bare minimum power consumption applied to as many RT cores that will fit in the package.

But that doesnt mean they cant repackage what they have done for mobile gaming, just that so far there wasnt the business case to do so.

In that scenario, AMDs most expensive silicon doesnt come close to Nvidias cheapest

The benchmarks you provided are severely outdated and was before Blender implemented HIP-RT (so the HIP-RT cores were not used in the benchmarks you linked and still aren't in the Blender Opendata benchmark, which recently showcased Nvidia having a 10+% performance decrease in Blender 4.0, whether this affects actual render time, no clue, most people render with 400 samples or below with an image denoiser and the image denoiser gained GPU acceleration with Blender 4.1 meaning Blender 4.1 with the right settings is still faster than Blender 3.6 in spite of the worse Opendata Benchmark score).

The performance has been quite similar for some Cycles benchmarks with Blender 3.6. In the default render engine Eevee (which way to few tech sites benchmark, probably as it follows the gaming benchmarks to hard) it is extremely similar performance wise.
And the Vulkan Compute of AMD matches quite well with Nvidia's.

So yeah, not really the silicon, just software optimization and Blender specifically has caught criticism for dropping OpenCL in the past before implementing ROCM, creating a Blender release that only benefited and worked on Nvidia cards.

[mSparks]

The benchmarks you provided are severely outdated and was before Blender implemented HIP-RT (so the HIP-RT cores were not used in the benchmarks you linked and still aren't in the Blender Opendata benchmark, which recently showcased Nvidia having a 10+% performance decrease in Blender 4.0, whether this affects actual render time, no clue, most people render with 400 samples or below with an image denoiser and the image denoiser gained GPU acceleration with Blender 4.1 meaning Blender 4.1 with the right settings is still faster than Blender 3.6 in spite of the worse Opendata Benchmark score).

The performance has been quite similar for some Cycles benchmarks with Blender 3.6. In the default render engine Eevee (which way to few tech sites benchmark, probably as it follows the gaming benchmarks to hard) it is extremely similar performance wise.
And the Vulkan Compute of AMD matches quite well with Nvidia's.

So yeah, not really the silicon, just software optimization and Blender specifically has caught criticism for dropping OpenCL in the past before implementing ROCM, creating a Blender release that only benefited and worked on Nvidia cards.

most is here

https://opendata.blender.org/benchmarks/query/?compute_type=OPTIX&compute_type=CUDA&compute_type=HIP&compute_type=METAL&compute_type=ONEAPI&group_by=device_name&blender_version=4.0.0

AMD starts bottom of page 2.

[mSparks]

The benchmarks you provided are severely outdated

most current is here
https://opendata.blender.org/benchmarks/query/?compute_type=OPTIX&compute_type=CUDA&compute_type =HIP&compute_type=METAL&compute_type=ONEAPI&group_ by=device_name&blender_version=4.0.0

AMD starts on page 2.
Their most expensive current card still only matches the RTX4060 ti

[tenchrio]

most current is here
https://opendata.blender.org/benchmarks/query/?compute_type=OPTIX&compute_type=CUDA&compute_type =HIP&compute_type=METAL&compute_type=ONEAPI&group_ by=device_name&blender_version=4.0.0

AMD starts on page 2.
Their most expensive current card still only matches the RTX4060 ti

And I think I clarified that opendata still does not use HIP-RT before in my previous post. You also have it set on Blender 4.0 when Blender 4.1 is out and available as an option to set as the Blender version. The RX7900XTX is now on the first page and the non-zluda benchmark is already outperforming the RTX 4060 TI, showing it is not the silicon that is at fault but the software.

Not to mention Opendata has questionable results, if you bothered looking at the Nvidia results of the link you provided you would see the RTX 4080 scoring better than the RTX 4080 Super and the RTX 3090 scoring above the RTX 3090 TI.

[mSparks]

And I think I clarified that opendata still does not use HIP-RT before in my previous post. You also have it set on Blender 4.0 when Blender 4.1 is out and available as an option

and yet your link still says
AMD Radeon RX 7900 XTX 4085.24
matches
NVIDIA GeForce RTX 4060 Ti 3868.34

while
NVIDIA GeForce RTX 4090 11304.24

Not to mention Opendata has questionable results, if you bothered looking at the Nvidia results of the link you provided you would see the RTX 4080 scoring better than the RTX 4080 Super and the RTX 3090 scoring above the RTX 3090 TI.

you are probably looking at laptop gpus....

Blender benchmarks are a 1:1 match to the performance /render time people using blender will experience.

Even if AMD had the physical silicon capable of competing, they are still a decade behind in the SDK stuff to make use of it.

Sorry to burst your bubble, but as good as the $1000 7900XTX card is, its still only matches a $400 RTX 4060 Ti in real world performance.

3 times the price or 1/3 of the performance isnt a smart purchase no matter how hard you wriggle.

[tenchrio]

and yet your link still says
AMD Radeon RX 7900 XTX 4085.24
matches
NVIDIA GeForce RTX 4060 Ti 3868.34
while
NVIDIA GeForce RTX 4090 11304.24

Way to keep changing the goal post.
Funny how you neglect AMD Radeon RX 7900 XTX [ZLUDA] 4848.05.

Blender benchmarks are a 1:1 match to the performance /render time people using blender will experience.

Opendata benchmark absolutely isn't, set it to 3.6 and compare it to this Blender 3.6 benchmark I linked before. Unless you also believe that in Blender 4.0 the RTX 3090 could outperform the RTX 3090 TI. And as someone with actual experience in Blender, Eevee is the default render engine and there the performance is a lot closer between the cards relative to their price point as I also linked before, the opendata benchmark does not take into account Eevee or Workbench render engine performance or Cycles with HIP-RT. It is also the first thing after updating Blender that people suggest to render faster, as Eevee does in a second what cycles does in minutes.

Even if AMD had the physical silicon capable of competing, they are still a decade behind in the SDK stuff to make use of it.

More like 8 years, like literally that is the time difference between CUDA and ROCM. And the performance gap is clearly closing with each Blender release, early HIP-RT benchmarks showcase a 31% increase in performance at best and that is in actual benchmarks. A lot of benchmarks also don't have adaptive sampling or OIDN enabled, despite them heavily affecting render time, popular benchmarks like BMW, Classroom or Fishy Cat (which you linked before) have these off by default. If you switch from Eevee to Cycles in a new project these are now on by default.

Funnily enough when downloading and running Fishycat I found that the actual render time was 4 seconds, the remaining 6 seconds of the total render time were spent on compositing nodes, which is btw done on CPU as the GPU compositor for actual renders is still experimental and can only be enabled in Blender Alpha builds for now (but the node setup is taxing so it still took 5 seconds with the GPU compositor enabled in the latest alpha build).

Sorry to burst your bubble, but as good as the $1000 7900XTX card is, its still only matches a $400 RTX 4060 Ti in real world performance.

3 times the price or 1/3 of the performance isnt a smart purchase no matter how hard you wriggle.

Sorry to burst your bubble but that clearly isn't true as I already showed before, the RTX 7900XTX is closer to the RTX 4070 ti since Blender 3.6 in an actual render case.
The benchmark you are referring to is also entirely synthetic and as I said before the RTX 7900XTX performs even better in Eevee, which is the default on any fresh Blender install and go to render engine for many artists especially for animation and NPR.

Considering the $400 price mark, I will also assume you refer to the RTX 4060 TI 8GB which would be impacted by a lack of VRAM in complex scenes, this was showcased in this video by Blender Rookie were an RTX 3060 12GB (2 min 9 sec) beats out an RTX 3070 (4min 52 sec) by utilizing a benchmark that is heavy on VRAM and the difference is a lot bigger than any Nvidia vs AMD advantage. Again something that isn't noticeable on things like Blender Opendata or a lot of the very old Blender Benchmarks that most sites run. This isn't even including Viewport performance where AMD straight up wins, the complexer the scene the lower it gets, just ask any experienced Blender Artist and having some extra legroom is a lot better than having to hide parts of your scene before rendering, potentially losing time as the composition of the scene is messed up and it is hard to align when you start dipping under usable numbers.

[mSparks]

Way to keep changing the goal post.
Funny how you neglect AMD Radeon RX 7900 XTX [ZLUDA] 4848.05.


Opendata benchmark absolutely isn't, set it to 3.6 and compare it to this Blender 3.6 benchmark I linked before. Unless you also believe that in Blender 4.0 the RTX 3090 could outperform the RTX 3090 TI. And as someone with actual experience in Blender, Eevee is the default render engine and there the performance is a lot closer between the cards relative to their price point as I also linked before, the opendata benchmark does not take into account Eevee or Workbench render engine performance or Cycles with HIP-RT. It is also the first thing after updating Blender that people suggest to render faster, as Eevee does in a second what cycles does in minutes.


More like 8 years, like literally that is the time difference between CUDA and ROCM. And the performance gap is clearly closing with each Blender release, early HIP-RT benchmarks showcase a 31% increase in performance at best and that is in actual benchmarks. A lot of benchmarks also don't have adaptive sampling or OIDN enabled, despite them heavily affecting render time, popular benchmarks like BMW, Classroom or Fishy Cat (which you linked before) have these off by default. If you switch from Eevee to Cycles in a new project these are now on by default.

Funnily enough when downloading and running Fishycat I found that the actual render time was 4 seconds, the remaining 6 seconds of the total render time were spent on compositing nodes, which is btw done on CPU as the GPU compositor for actual renders is still experimental and can only be enabled in Blender Alpha builds for now (but the node setup is taxing so it still took 5 seconds with the GPU compositor enabled in the latest alpha build).


Sorry to burst your bubble but that clearly isn't true as I already showed before, the RTX 7900XTX is closer to the RTX 4070 ti since Blender 3.6 in an actual render case.
The benchmark you are referring to is also entirely synthetic and as I said before the RTX 7900XTX performs even better in Eevee, which is the default on any fresh Blender install and go to render engine for many artists especially for animation and NPR.

Considering the $400 price mark, I will also assume you refer to the RTX 4060 TI 8GB which would be impacted by a lack of VRAM in complex scenes, this was showcased in this video by Blender Rookie were an RTX 3060 12GB (2 min 9 sec) beats out an RTX 3070 (4min 52 sec) by utilizing a benchmark that is heavy on VRAM and the difference is a lot bigger than any Nvidia vs AMD advantage. Again something that isn't noticeable on things like Blender Opendata or a lot of the very old Blender Benchmarks that most sites run. This isn't even including Viewport performance where AMD straight up wins, the complexer the scene the lower it gets, just ask any experienced Blender Artist and having some extra legroom is a lot better than having to hide parts of your scene before rendering, potentially losing time as the composition of the scene is messed up and it is hard to align when you start dipping under usable numbers.

I'm not sure what point you think you are making, but it is certainly no longer that nvidia GPUs cant compete with AMD.

it almost seems to be sinking in that AMD finally being able today with the nvidia chip in the 2017 switch doesnt make AMD competitive.