Blender 4.3 Released With AMD HIP-RT Ray-Tracing On Linux, Experimental Vulkan Backend

[Adarion]

(not strictly Blender related)

> Support for Vega in Cycles AMD HIP backend has been removed
Ah, yeah. "Nice."

So there are improvements, but only for those who like to spend 1000+ USD/Euro/whatever on the available high-end models.

I wonder how AMD is going to handle anything once they focus on middle-class dGPUs only (read the news from the recent months).
AMD needs to hire more people to get their software side done in a decent manner. There still are some nasty bugs and regressions, even though mostly the sheer graphics part is pretty fine, but anything even remotely related to computing on GPUs lacks to no end. There must be much broader support and there is no good in not supporting anything older than a year.
And handling / installation must be a breeze, so the majority of people can simply put it to good use. But we are far from that.

On the Blender side of things Blender also has focused on the nvidia monopoly for way too long in the past. This is bad and now we earn the sinister results.

[stargeizer]

For those wondering about AMD...

OpenCL: Is not coming back. Simply put, AMD hardware implementation of it has some severe bugs and limitations in the sillicon that made coding a raytracing engine on it, basically a nightmare. Even with code and funding provided by AMD, neither Blender devs nor AMD devs were able to finish a good enough backend of Cycles on AMD's own hardware. Also, Apple deprecation and dropping of OpenCL, put the nail in the coffin of OpenCL development out there. Even radeon ProRender is using HIP nowadays to avoid OpenCL buggyness. And for the record: Ironically, OpenCL code worked correctly and flawlessly only on nvidia cards. (but quite slower than CUDA, mind you).

Nvidia monopoly: The main problem with nvidia, is their actual "JUST WORKS" policy (i never expected to say this, or to see how the "mafia" became "the good guys" in the coding side) . Something Linux users/developers tend to avoid like a plague. Software wise, nvidia invested quite a large amount of work in their software. That's why coders do work mostly with nvidia. Code, compille and forget. That's all folks.

Vega: It was AMD who deprecated VEGA support on HIP drivers, not Blender devs. There are some bugs on VEGA hardware that aren't going to be fixed by AMD and Blender devs just removed the support from Cycles, since they cannot fix/workaround the bugs. Nothing to do here, unfortunately.

[MillionToOne]

So there are improvements, but only for those who like to spend 1000+ USD/Euro/whatever on the available high-end models.

I believe they want to focus on the main cards people use, with the main cards being Radeon RX series.

On the Blender side of things Blender also has focused on the nvidia monopoly for way too long in the past. This is bad and now we earn the sinister results.

What do you mean? They support all of the major players in the GPU space. They support Intel, AMD and Nvidia. AMD just doesn't know what it's doing half of the time unlike Intel and Nvidia. You shouldn't have million GPU lineups. Only 2 or 3. Server GPUs, consumer GPUs and GPUs for heavy workloads like AI at a large scale.

[Quackdoc]

Please add back OpenCL support to Blender. AMD doesn't support HIP normally. Only a few GPU gets full support.

not gonna happen, gotta wait for vulkan accel in the renderers, RIP old gpus

[tenchrio]

This is interesting but can you test on some demo files: Classroom, Barbershop, Lone Monk etc.?

The 7900 xtx and 3090 are the comparisons I'm interested in, ironically - so, it's intriguing those are your cards? The 7900 xtx used, in my market, is about $100 ish more - assuming someone would reduce their price on the amd gpu. The 3090 prices are all over the place but that's probably because some ppl are hoping for rich ppl to just pay what they are asking - and they think the 24gb gives them the right to hype their cards - plus, it's nvidia. But, they are around $100-$200 used. The 7900 xtx new prices are still high.

The 7900 xtx would be the only amd gpu I'd even consider - but, used. It's disconcerting - at least, to me, though - that the hip-rt - ray tracing - is still not reliable or ironed out - it's still considered experimental. I feel like I'd still be forced to go with nvidia even though it's the older gen/architecture. The 4070 Ti Super might be coming down in price depending....the only 40 series I'd look at - used again, though. I wouldn't go anything lower than 3090 or 4070 Ti S.

Anyway, didn't mean to ramble on - thanks for your write up.

I finally had some time to run these. There were some issues with the Lone Monk render on the RTX 3090:
Screenshot from 2024-11-21 19-52-08.png
This issue seems to be reported for the 4090 as well, disabling "Open Shading Language" and restarting Blender let me render the file.

To be sure I updated the drivers on both machines (so latest ROCM and latest nvidia drivers). Obviously GPU compute.
Second render onwards (else the 3090 would score 2 minutes on classroom, IDK why, it was the last Cycles benchmark I ran and somehow it was the longest overall).
Denoising on (if it wasn't) and "use GPU" for denoising + GPU compositor (different CPUs in the system so less chance on that creating a difference).

	Barbershop	Monk	Classroom	Classroom Eevee
RTX 3090	1m06.33s	1m29.64s	14.98s	3.94s
RX 7900XTX	1m12s	1m45.03s	15.65s	2.25s

BVH building definitely felt longer on the AMD card, I wish I had some time to figure out how much that affects the outcome (like did the actual rendering take longer or does it only seem like it due to the BVH building) but I have little time this week and the next. As said before this issue is know and might be fixed by Blender 4.4. None of the HIP-RT renders looked noticeably off in output, looked about the same as Optix. RTX 3090 also has improved Eevee performance, wasn't expecting that when they announced light linking but hey free bonus.

In my country neither cards are all that prevalent on the used market. New the RX 7900XTX goes for as low as €850, the RTX 3090 for €1200 these are currently the 2 cheapest 24GB cards (quadro cards are insanely expensive especially the ADA ones). If you can get the (used) RTX 3090 cheaper (and you are certain they are still in good condition) it is still worth it. The RTX 4070 Ti Super is also a good card on paper but there has been reports about RTX 40 series cards falling back faster on shared memory (Out of vram) compared to RTX 30 cards (some scenes apparently rendering fine on an RTX 3080 12GB but are failing on the RTX 4090).

Provided that gets fixed, 16GB can be enough but I've had some projects spike well into 20GB, if you're lucky shared memory can save the day (at a hefty speed penalty) but if not you either need to start making sacrifices or lose a lot more time optimizing your render (or fall back to CPU rendering), I'm not the only artist that argues VRAM over speed (the old 4070 TI was a definite no, 12GB is a line you cross fast once you start making complex scenes, 16GB doesn't feel all that safe by comparison). But arguably it also depends on what you make and how you make it (the Geonodes bake node has had me sometimes almost halving the render time but at the cost of greatly increasing the VRAM usage), nothing is worse though than believing you are finished and reading the words "Out of Cuda Memory".

On a sidenote, the RTX 4500 ADA uses the same chip as the RTX 4070 TI S with slightly less CUDA cores (7680 instead of 8448, so it is actually a bit slower) but is blessed with 24GB VRAM. This card goes for a whopping minimum of €2000 (there are a couple more caveats why this card goes for about the same price as an RTX 4090, one being better suited for multi gpu since 2 slot so easily 3x fitable in a case, performance bla bla bla but I need sleep) so yeah Nvidia selling an arguably weaker version of the 4070Ti S with more VRAM for almost thrice the price while marketing towards creators, should be a clear sign I am not kidding about VRAM's importance.

[tenchrio]

Please add back OpenCL support to Blender. AMD doesn't support HIP normally. Only a few GPU gets full support.

All RDNA1, RDNA2 and RDNA3 gpus are supported by HIP. ROCM supports every RDNA2 and RDNA3 GPU (for some reason the AMD docs only mention the more expensive cards out right, presumably for legal/ass covering reasons, the Debian Rocm docs are in my opinion better, Debian's ROCM compile also has RDNA1 enabled for a couple more libraries, AMD's does compile with RDNA1 for the runtimes required to run applications like Blender).

(not strictly Blender related)
And handling / installation must be a breeze, so the majority of people can simply put it to good use. But we are far from that..

I see this sentiment a lot but I don't agree with it.
It might have been true a number of years ago but I don't feel like copy pasting the quick install guide is in any way difficult.
The detailed guide for ROCM is about the same as the one for CUDA and recently a lot of distro's have been adding ROCM to their repositories (so install is even easier). Just like CUDA there are generic install instructions to make it run on a distro that is not listed (provided you meet the kernel version requirement).

HIP/ROCM and CUDA runtimes to run applications can also be installed without the entirety of ROCM or CUDA (as is usually the case, the runtimes tend to be shipped together with the graphics drivers, I've had a couple of times were updating the Nvidia Drivers for some reason did not update the CUDA runtimes, nothing a manual sudo apt install won't fix but there is always that level of confusion when opening Blender, seeing GPU compute grayed out and finding no GPUs listed under Cycles Optix or Cuda).

OpenCL: Is not coming back. Simply put, AMD hardware implementation of it has some severe bugs and limitations in the sillicon that made coding a raytracing engine on it, basically a nightmare. Even with code and funding provided by AMD, neither Blender devs nor AMD devs were able to finish a good enough backend of Cycles on AMD's own hardware. Also, Apple deprecation and dropping of OpenCL, put the nail in the coffin of OpenCL development out there. Even radeon ProRender is using HIP nowadays to avoid OpenCL buggyness. And for the record: Ironically, OpenCL code worked correctly and flawlessly only on nvidia cards. (but quite slower than CUDA, mind you).

Nvidia monopoly: The main problem with nvidia, is their actual "JUST WORKS" policy (i never expected to say this, or to see how the "mafia" became "the good guys" in the coding side) . Something Linux users/developers tend to avoid like a plague. Software wise, nvidia invested quite a large amount of work in their software. That's why coders do work mostly with nvidia. Code, compille and forget. That's all folks.

Vega: It was AMD who deprecated VEGA support on HIP drivers, not Blender devs. There are some bugs on VEGA hardware that aren't going to be fixed by AMD and Blender devs just removed the support from Cycles, since they cannot fix/workaround the bugs. Nothing to do here, unfortunately.

This is just filled with falsehoods. Cycles had a rewrite for version 3.0 with the main purpose being better CUDA performance and Optix, the Blender team then opted to go with HIP and individual API's for every vendor, OpenCL was just dropped plain and simple, no attempt was made to make it work on the then new "Cycles-X" (now called Cycles). When optix was first introduced in Blender 2.81 it sometimes ran worse than CUDA, hence the rewrite (and it is hard to ignore how heavily it favored Nvidia, also "just works" my ass, I've had my fair share of problems with Nvidia gpus, not saying AMD is flawless but Nvidia is far from the flawless experience that non-Blender users make it out to be, the RTX 20 series was especially bad, but for some reason only AMD gets haunted by past driver issues, if you choose Nvidia over AMD because you think there won't be issues you're going to be disappointed (same for the reverse, both have issues, run CPU if you want just works)).

Radeon Prorender still lists OpenCL on its product page as it is used for legacy compatibility. There are also numerous OpenCL based Render engines that can easily be installed in Blender (e.g. Indigo Renderer), all of those also still have Network Rendering and Multi-GPU support, 2 features that Cycles lost when it was rewritten for 3.0 (crazy how that part "Just works" in the older version but hasn't for over 3 years now). The Vega portion is true, for ROCM at least (and you can switch to a different render engine using libraries that Vega still has support for).

not gonna happen, gotta wait for vulkan accel in the renderers, RIP old gpus

Vulkan compute will probably never be implemented in Cycles, apparently it lacks equivalent libraries that the other 5 do have that are essential.

However Vulkan graphics based render engines do exist. Blender's Eevee is an easy example (it's also the default) and AMD's Radeon Prorender engine combines Vulkan compute with HIP (or OpenCL on older machines, note HIP also uses CUDA). These types of real-time render engines are also seeing more popularity, Eevee has been used in a couple of featured length movies by companies aside from the Blender foundation itself. Since Eevee also supports most of the Cycles nodes with near parity it is not that hard of a switch to make and due to the faster performance there is a case for learning and using it even on Nvidia GPUs.

With those render engines it is technically possible to use the RADV drivers and render using GPUs as far back as the AMD HD 7900 series.

[Quackdoc]

Vulkan compute will probably never be implemented in Cycles, apparently it lacks equivalent libraries that the other 5 do have that are essential.

However Vulkan graphics based render engines do exist. Blender's Eevee is an easy example (it's also the default) and AMD's Radeon Prorender engine combines Vulkan compute with HIP (or OpenCL on older machines, note HIP also uses CUDA). These types of real-time render engines are also seeing more popularity, Eevee has been used in a couple of featured length movies by companies aside from the Blender foundation itself. Since Eevee also supports most of the Cycles nodes with near parity it is not that hard of a switch to make and due to the faster performance there is a case for learning and using it even on Nvidia GPUs.

With those render engines it is technically possible to use the RADV drivers and render using GPUs as far back as the AMD HD 7900 series.

khronos is changing up a lot when it comes to how vulkan works, here is a bit of a teaser https://www.youtube.com/watch?v=fQR1hk-Ktjs&t=3273s

I cant say this will fix things, but it might

[Panix]

I finally had some time to run these. There were some issues with the Lone Monk render on the RTX 3090:
Screenshot from 2024-11-21 19-52-08.png
This issue seems to be reported for the 4090 as well, disabling "Open Shading Language" and restarting Blender let me render the file.

To be sure I updated the drivers on both machines (so latest ROCM and latest nvidia drivers). Obviously GPU compute.
Second render onwards (else the 3090 would score 2 minutes on classroom, IDK why, it was the last Cycles benchmark I ran and somehow it was the longest overall).
Denoising on (if it wasn't) and "use GPU" for denoising + GPU compositor (different CPUs in the system so less chance on that creating a difference).

	Barbershop	Monk	Classroom	Classroom Eevee
RTX 3090	1m06.33s	1m29.64s	14.98s	3.94s
RX 7900XTX	1m12s	1m45.03s	15.65s	2.25s

BVH building definitely felt longer on the AMD card, I wish I had some time to figure out how much that affects the outcome (like did the actual rendering take longer or does it only seem like it due to the BVH building) but I have little time this week and the next. As said before this issue is know and might be fixed by Blender 4.4. None of the HIP-RT renders looked noticeably off in output, looked about the same as Optix. RTX 3090 also has improved Eevee performance, wasn't expecting that when they announced light linking but hey free bonus.

In my country neither cards are all that prevalent on the used market. New the RX 7900XTX goes for as low as €850, the RTX 3090 for €1200 these are currently the 2 cheapest 24GB cards (quadro cards are insanely expensive especially the ADA ones). If you can get the (used) RTX 3090 cheaper (and you are certain they are still in good condition) it is still worth it. The RTX 4070 Ti Super is also a good card on paper but there has been reports about RTX 40 series cards falling back faster on shared memory (Out of vram) compared to RTX 30 cards (some scenes apparently rendering fine on an RTX 3080 12GB but are failing on the RTX 4090).

Provided that gets fixed, 16GB can be enough but I've had some projects spike well into 20GB, if you're lucky shared memory can save the day (at a hefty speed penalty) but if not you either need to start making sacrifices or lose a lot more time optimizing your render (or fall back to CPU rendering), I'm not the only artist that argues VRAM over speed (the old 4070 TI was a definite no, 12GB is a line you cross fast once you start making complex scenes, 16GB doesn't feel all that safe by comparison). But arguably it also depends on what you make and how you make it (the Geonodes bake node has had me sometimes almost halving the render time but at the cost of greatly increasing the VRAM usage), nothing is worse though than believing you are finished and reading the words "Out of Cuda Memory".

On a sidenote, the RTX 4500 ADA uses the same chip as the RTX 4070 TI S with slightly less CUDA cores (7680 instead of 8448, so it is actually a bit slower) but is blessed with 24GB VRAM. This card goes for a whopping minimum of €2000 (there are a couple more caveats why this card goes for about the same price as an RTX 4090, one being better suited for multi gpu since 2 slot so easily 3x fitable in a case, performance bla bla bla but I need sleep) so yeah Nvidia selling an arguably weaker version of the 4070Ti S with more VRAM for almost thrice the price while marketing towards creators, should be a clear sign I am not kidding about VRAM's importance.

Your country might be Germany?

Thanks for all that info! Really interesting? I'm not sure about the USA but in parts of Canada -the 3090 and 7900 xtx - USED only - are around the similar price - the 3090 *can* be found (or negotiated, rather) for a bit cheaper - maybe $100 - compared to a used 7900 xtx. I like the AMD gpu for Linux use (mostly) but the 3090 I believe(d) is better for overall production use - for Blender and Davinci Resolve - the 2 programs I'd use, mostly - but, I'm also interested in Stable Diffusion - and again the nvidia gpu is probably preferable. But, this thread is about Blender - and I have tried to research the performance of both cards and this was my main comparison - the AMD gpu's only hope of being 'competitive' (imho) vs the 3090 rested in improved hip/hip-rt performance - in particular, the ray tracing.

As I have posted before - I found 'mixed reviews' - one youtuber surmised that the ray tracing performance - i.e. hip-rt allows the 7900 xtx to be pretty close to the 3090 in several tests/benchmarks - but, I found it hard to believe. Another youtuber said it wasn't much of an improvement in enabling the ray tracing. So, your tests were really interesting. If I was able to have a 2nd PC, I'd definitely like to have an all-AMD PC with that gpu.

I wonder what AMD is going to have when they switch the architecture from RDNA 4 to UDNA? Will that help at all? Hmmmmm....

It appears that Blender performance is pretty adequate - but, on the Blender forums - the last I read is that Linux use and any hip-rt performance was a mixed bag - quite a few reports/anecdotes/feedback posts about crashes and disappointment - perhaps, the newer Blender versions and AMD's hip-rt developments have improved the situation? :-)

Unfortunately, it's still considered experimental.

P.S. it was also my hunch that vram can make a big difference too and is also a bonus - it's there in case you need it which is why I compare(d) these 2 cards. The 4090 is a different animal and price so it isn't/wasn't an option - at least, for me.

[tenchrio]

khronos is changing up a lot when it comes to how vulkan works, here is a bit of a teaser https://www.youtube.com/watch?v=fQR1hk-Ktjs&t=3273s

I cant say this will fix things, but it might

Hard to say from that clip but as far as my memory served it was based on certain extensions not necessarily how Vulkan compute and SPIR-V worked (but perhaps the reason for the lack of said extensions could be related to that).

I can't find exactly which extensions/libraries it was for Blender (it is now burried in a lot of posts about Vulkan Graphics) but I vaguely remember that the reason a ML project went with CUDA and ROCM versions was because Vulkan didn't have a Basic Linear Algebra Subprograms extension like cuBLAS and rocBLAS (Intel's version is called oneMKL, kind of disappointing) and it still doesn't do this day (the first result for "Vulkan BLAS" is Vulkan RT BLAS which stands for Bottom Level Acceleration Structure, unfortunately unrelated to the other BLAS).

Perhaps one day Vulkan will have these extensions and we will see a Vulkan option for Cycles.

Your country might be Germany?

Thanks for all that info! Really interesting? I'm not sure about the USA but in parts of Canada -the 3090 and 7900 xtx - USED only - are around the similar price - the 3090 *can* be found (or negotiated, rather) for a bit cheaper - maybe $100 - compared to a used 7900 xtx. I like the AMD gpu for Linux use (mostly) but the 3090 I believe(d) is better for overall production use - for Blender and Davinci Resolve - the 2 programs I'd use, mostly - but, I'm also interested in Stable Diffusion - and again the nvidia gpu is probably preferable. But, this thread is about Blender - and I have tried to research the performance of both cards and this was my main comparison - the AMD gpu's only hope of being 'competitive' (imho) vs the 3090 rested in improved hip/hip-rt performance - in particular, the ray tracing.

As I have posted before - I found 'mixed reviews' - one youtuber surmised that the ray tracing performance - i.e. hip-rt allows the 7900 xtx to be pretty close to the 3090 in several tests/benchmarks - but, I found it hard to believe. Another youtuber said it wasn't much of an improvement in enabling the ray tracing. So, your tests were really interesting. If I was able to have a 2nd PC, I'd definitely like to have an all-AMD PC with that gpu.

I wonder what AMD is going to have when they switch the architecture from RDNA 4 to UDNA? Will that help at all? Hmmmmm....

It appears that Blender performance is pretty adequate - but, on the Blender forums - the last I read is that Linux use and any hip-rt performance was a mixed bag - quite a few reports/anecdotes/feedback posts about crashes and disappointment - perhaps, the newer Blender versions and AMD's hip-rt developments have improved the situation? :-)

Unfortunately, it's still considered experimental.

P.S. it was also my hunch that vram can make a big difference too and is also a bonus - it's there in case you need it which is why I compare(d) these 2 cards. The 4090 is a different animal and price so it isn't/wasn't an option - at least, for me.

I order from Germany, EU countries have the luxury to just order from the cheapest country and have it shipped with no additional import costs.
If you can get it cheaper, go for it. The RX7900XTX does have better Vulkan performance which is useful if your workflow is centered around Eevee (personally I haven't used Cycles in my own projects since 4.2 but this will depend from artist to artist), it also seems to have better Davinci Resolve performance and scores relatively well in the SD Webui Benchmark (but it isn't as if the RTX 3090 has bad performance in those regards so just go for the card that is cheaper, both are good).

The ray tracing mixed reviews could stem from the fact not every shader is as impacted by the RT acceleration as the next as well as other factors like light bounces and samples. Even with Optix you can have scenes that barely benefit compared to CUDA (and other scenes that greatly benefit from it), this will depend on how you go about making your project (e.g. for glass some just use a glass BSDF, others mix it with a translucent BSDF or principled BSDF and other would just tweak the Principled BSDF until it looks like glass, each one would have you saying "Jup that looks like glass, just a (tiny) bit different from that other one" but each one is handled differently in rendering and on a large enough scale could affect rendering).

This difference includes a difference in output, normally this is barely noticeable (you would need to use program like gimp and use burn/substract/etc on the layer to even notice the difference) but there has been an unsolved issue where the difference between CPU, CUDA and Optix is extremely noticeable (still present from my testing but can be fixed by moving the cube up a little making it float above the surface, regardless a good example of how each one handles shaders differently affecting not just speed but technically outcome).

The 4090 is definitely in a league of its own, I was thinking of picking one up but somehow the price increased from €1700 to €2000+ (used price went up to €1600 at the lowest) so that is off the table for now. Kinda also want an Intel GPU just to experiment with but Black Friday didn't give any great price drops so might wait till Cyber Monday or buy at full price afterwards, they are pretty cheap but still money saved is money saved.

[qarium]

I finally had some time to run these. There were some issues with the Lone Monk render on the RTX 3090:
Screenshot from 2024-11-21 19-52-08.png
This issue seems to be reported for the 4090 as well, disabling "Open Shading Language" and restarting Blender let me render the file.

To be sure I updated the drivers on both machines (so latest ROCM and latest nvidia drivers). Obviously GPU compute.
Second render onwards (else the 3090 would score 2 minutes on classroom, IDK why, it was the last Cycles benchmark I ran and somehow it was the longest overall).
Denoising on (if it wasn't) and "use GPU" for denoising + GPU compositor (different CPUs in the system so less chance on that creating a difference).

	Barbershop	Monk	Classroom	Classroom Eevee
RTX 3090	1m06.33s	1m29.64s	14.98s	3.94s
RX 7900XTX	1m12s	1m45.03s	15.65s	2.25s

BVH building definitely felt longer on the AMD card, I wish I had some time to figure out how much that affects the outcome (like did the actual rendering take longer or does it only seem like it due to the BVH building) but I have little time this week and the next. As said before this issue is know and might be fixed by Blender 4.4. None of the HIP-RT renders looked noticeably off in output, looked about the same as Optix. RTX 3090 also has improved Eevee performance, wasn't expecting that when they announced light linking but hey free bonus.

In my country neither cards are all that prevalent on the used market. New the RX 7900XTX goes for as low as €850, the RTX 3090 for €1200 these are currently the 2 cheapest 24GB cards (quadro cards are insanely expensive especially the ADA ones). If you can get the (used) RTX 3090 cheaper (and you are certain they are still in good condition) it is still worth it. The RTX 4070 Ti Super is also a good card on paper but there has been reports about RTX 40 series cards falling back faster on shared memory (Out of vram) compared to RTX 30 cards (some scenes apparently rendering fine on an RTX 3080 12GB but are failing on the RTX 4090).

Provided that gets fixed, 16GB can be enough but I've had some projects spike well into 20GB, if you're lucky shared memory can save the day (at a hefty speed penalty) but if not you either need to start making sacrifices or lose a lot more time optimizing your render (or fall back to CPU rendering), I'm not the only artist that argues VRAM over speed (the old 4070 TI was a definite no, 12GB is a line you cross fast once you start making complex scenes, 16GB doesn't feel all that safe by comparison). But arguably it also depends on what you make and how you make it (the Geonodes bake node has had me sometimes almost halving the render time but at the cost of greatly increasing the VRAM usage), nothing is worse though than believing you are finished and reading the words "Out of Cuda Memory".

On a sidenote, the RTX 4500 ADA uses the same chip as the RTX 4070 TI S with slightly less CUDA cores (7680 instead of 8448, so it is actually a bit slower) but is blessed with 24GB VRAM. This card goes for a whopping minimum of €2000 (there are a couple more caveats why this card goes for about the same price as an RTX 4090, one being better suited for multi gpu since 2 slot so easily 3x fitable in a case, performance bla bla bla but I need sleep) so yeah Nvidia selling an arguably weaker version of the 4070Ti S with more VRAM for almost thrice the price while marketing towards creators, should be a clear sign I am not kidding about VRAM's importance.

for this vram problem i have a AMD PRO W7900 with 48gb vram in my computer

you can buy it for around 3000€ ... (it started at 4000€ years ago and dropped in price to now around 3000€)

"I'm not the only artist argues VRAM over speed"

for people like this the AMD PRO W7900 fits well... with 48gb vram you are future proof.