Calling all low and mid GPU owners - shall we swap RT for more performance or lower prices?

[Apocalypsee]

Thought this happen before when Nvidia release two different DirectX gen on low and high end tier (GeForce MX and Ti DX7 vs DX8) back then, or GeForce 1600 series compared to RTX 2000 series. It didn't end well. I don't know if GeForce 1600 support DLSS or not. RT is in DirectX 12 ultimate specification so you can't say no to that.

IMO it's better to have support for it like Nvidia DLSS require those cores to be present. It will benefit low end card to achieve higher frame rate at the expense of image quality. It does makes it blurry but at least you can adjust it quality to suit your eyes.

 

[wolf]

Game releases were awesome, imo.

I'd say some of the games themselves were, but lawd there were some terrible game launches this year, some games still in need of a lot of work.

 

[Redwoodz]

I wonder how long threads like these will continue to fight the future of GPU design. Thankfully, the people actually designing GPUs understand what that future should be.

As long as it takes. What we are fighting for is the future of "gaming" design, not AI designing disguised as gaming "advancement". If we are not lied to, we might support it.

 

[Dorek]

I'm not on very low end but i think id swap out the rt cores for more performance and perhaps price ratio.

 

[Onasi]

I am not a high-end user exactly, I usually hover around the 400-500 mark in terms if my GPU purchases which previously meant a x70 class mostly. Went from a 570 to 770 to 1070. Planning to probably upgrade to a 4070S if the price won’t be ridiculous. More than I would like to pay, but them the breaks.
As far as the question goes - no. It was tried with Turing and the 16 series and the result was meh. There really isn’t any point, too. Raster performance is harder to increase with each gen not only because “NVidia bad” or whatever, but, frankly, because all viable pathways to it have been already taken. We are essentially polishing up tech that’s nearly tapped out. RT is the future, not because of marketing, but simply because this is the only way to really push forward the improvements in visuals many seem to expect. Now, I couldn’t care less, I am of the opinion that we had reached “good enough” graphics like a decade ago and if the GPU industry suddenly just froze at the current performance level I would still say that we are fine. But I am not the majority. People want photorealism, they want more detail, more volume, more effects. Misguided? Perhaps. But if they want that they will have to swallow down the RT pill with it. As such, hindering the RT performance for some minor raster gains (if that dilemma was even real, which I am not sure of) is counterproductive at best, pointless at worst.

 

[Shihab]

RT cards are inefficient because they still rely on rasterization heavily as the main form of rendering.
Removing the rasterization will make them much faster, however they won't be backwards compatible with games that have rasterization, that's the main problem.
it requires a paradigm shift in development, gaming, & even on the OS side would need to support RT only designed cards with no rasterization.

I don't think it's an either or situation. Even production-grade ray/path tracing has a lot of tricks (and some fundamental pipeline stages) that would require the same hardware traditional game renderers use.
Resource allocation ratios in the hardware would change for an RT-centric world, granted, but I don't think we're ever going to "remove the rasterization."

To be fair everything has gotten meh in 2023

 

[TumbleGeorge]

RT is the future, not because of marketing, but

I definitely don't need any more visual effects than what we already have with dx12_1@dx11. All in all, I can't see it out even on a static image and I think everyone who claims it exists is lying.

 

[wolf]

I definitely don't need any more visual effects than what we already have with dx12_1@dx11. All in all, I can't see it out even on a static image and I think everyone who claims it exists is lying.

That's quite a take there. For sure some games it can be very subtle (or even very natural, so natural it isn't really even striking because it's what your eyes and brains expect to see), but man some games are night and day transformative been rast only vs RT/PT effects. From my perspective if you think you've never seen the difference when even shown side by side, I'd say you're lying.

I can respect people saying they don't care, or it's not for them, but can't [ever?] see the difference... that's a stretch.

 

[Onasi]

I definitely don't need any more visual effects than what we already have with dx12_1@dx11. All in all, I can't see it out even on a static image and I think everyone who claims it exists is lying.

And I don’t need the games to look much better than something like Half-Life 2. Or, if we go a bit more modern, MGS V. Or, if we talk stylization - Prince of Persia 2008. Hell, my main drug of a game series is freaking Ys and Trails, where even the latest entry on a “new” engine looks just like a default anime JRPG with not much fidelity due to it kinda having to run on a Switch. But personal preferences are irrelevant when the crowds are clamoring for better graphics at any cost.

 

[TumbleGeorge]

But personal preferences are irrelevant when the crowds are clamoring for better graphics at any cost

I couldn't understand this either, in my city there is not a million protest in the street demanding the introduction of rt effects at a cost as much as the annual net value of minimal wage in my country. Where are these crowds you claim?

 

[Onasi]

Are… are you serious? Just recently we saw people crapping all over Starfield for looking like a previous gen title. Tekken 8 is now catching flak for not looking as good as the reveal trailer despite, and I quote, “but it’s on UE5, where Lumen, where Nanite?”. Hell, I have seen people repeatedly claiming that BG3 looks “bad” because of blurry ground textures. In an isometric CRPG made by a AA developer house. The whole hype behind the new, now current gen, consoles was the fact that they can do RT now. People don’t literally riot in the streets with their demands, of course, but the moment something doesn’t look like their, however misguided, perception of next-gen graphics they instantly and inevitably start to moan.

 

[dj-electric]

The title poses a bit of a trick question. The reason it is, is because it needs to be understood why silicon is designed the way it is - both from a technical and from a marketing and manufacturing perspective.

When a company like NVIDIA approaches silicon design, things like RT cores are not being considered 'extra' or 'bonus' or 'optional'. These things are in the very basic building blocks of a unified and scalable architecture. The outlier being GTX 1660 and the transition period between 10 series and 20 series.

Typically, when making smaller dies, the question isn't if RT core real estate can be replaced with more traditional CUDA cores, but typically what size silicon and what performance output you want out of a product - since the building blocks already exist and include those RT cores. Not using them or designing special small cores that do not include them might in fact be very cost inefficient. You will have to reroute and redesign the architechture to not leave any wasted space. The only decision left for the silicon design team is when to press the go-ahead and implement these kinds of cores as a part of your building blocks - a decision made by NVIDIA earlier, and by AMD a bit later with RDNA.

As a result, the entire stack is ought to get RT cores, regardless of how viable the performance is when we talk about more entry level products such as the RTX 3050 or even the RTX 4060. The idea behind persisting with having all the none-traditional type of compute transistors in the silicon is that as the technology matures, and as using the same building blocks help shaping products, what didn't use to be that viable will start being somewhat viable. There's quite a difference between how an RTX 2060 deals with Raytracing and how an RTX 4060 does in terms of end-result.

So yes, it sucks that my 200mm^2 GPU can't really handle Raytracing to a satisfying level (to me) and instead of more theoretical CUDA cores i get silicon space dedicated to RT. But, that's how a master-design works with scalable silicon, and that's NVIDIA's (or AMD's decision, if we were to talk about RDNA/2/3) in order to be able to properly work their design teams and their rented fab nodes.

Eventually, more and more none-gaming applications are also starting to use these silicon real estate parts, like modeling, architecture (houses and apartments in this context) and many other types of CAD, simulation and engineering tasks. By now, everybody is implementing RT-intended hardware in their silicon including ARM based mobile SoCs. Its the current natural progression of GPUs.

 

[TumbleGeorge]

Are… are you serious? Just recently we saw people crapping all over Starfield for looking like a previous gen title. Tekken 8 is now catching flak for not looking as good as the reveal trailer despite, and I quote, “but it’s on UE5, where Lumen, where Nanite?”. Hell, I have seen people repeatedly claiming that BG3 looks “bad” because of blurry ground textures. In an isometric CRPG made by a AA developer house. The whole hype behind the new, now current gen, consoles was the fact that they can do RT now. People don’t literally riot in the streets with their demands, of course, but the moment something doesn’t look like their, however misguided, perception of next-gen graphics they instantly and inevitably start to moan.

All the described problems are either technical bugs or bad work by the artists in the team that created the games.

 

[Onasi]

Your takes are getting weirder by the minute. First was that RT is actually a lie and impossible to see, now we are going with “game not including certain new graphics tech is a bug and/or the artists were lazy”.
Uh. Okay then. I don’t even know how to argue with that.

 

[N/A]

RT needs to operate for free on top of what cuda cores are doing, instead it leads to a 50% performance drop for no good reason. RT should be doing its own computation entirely. and instead perhaps it causes all kinds of internal latencies sending part of the work to the cuda cores or even sit idle. Maybe if they double the RT it would lead to the 30% boost. But for now i'd add just one more 64 bit PHY. 4060 -> 192 bit. 4070 -> 256 bit. 4080-> 320 bit. only takes 13mm2 so simple.

 

[Sithaer]

Personally I kinda like RT, mainly the light/reflection related settings.
Tho I don't really require high fps to enjoy my games, anything around ~50 is all good with me so as long as I have headroom for some RT around that perfomance level then I will enable it or the very least reflections if it has a separate option. 'cause once I notice the standard screen space reflections turning itself on and off the moment I move the camera I simply cannot unsee it..'

I did finish some games with RT enabled or at least partially enabled with my 3060 Ti 'Cyberpunk/Guardians of the Galaxy/Control for example' + I also like those older remastered/tweaked games with RT.

Dunno, I mean its not a must have for me but I do like it and I'm not against new tech. 'DLSS/DLAA is almost a must have for me at this point so that I wouldn't want to drop'

 

[KrazyT]

I'll take the card in my $ range ...
If I can enable RT on game with a stable V-Sync 60 Hz, it's OK, if not, well nevermind RT.
I always prioritize texture quality, geometry and viewving distance ...
Lightning and Shadows are lowered if the game runs bad.

But to be honest, 75 % of the games I'm playing now are more than 10 years old, so no RT and running good

 

[londiste]

Would you be open to sacrificing the capability to run Ray Tracing on mid and low-end(under U$500) GPUs in exchange for either a 30% boost in performance or a 30% reduction in price, while maintaining the same level of performance as the current lineup?

This is an incorrect question to start with. Pretty early in RTX arguments the die shots have been analyzed resulting in a best guess that both Tensor and RT cores take up ~11% of the die area. I might remember the numbers slightly incorrectly but Tensor Cores took ~8.2% and RT Cores ~2.7% of die space. The numbers have not changed all that much with newer generations. If anything the percentage would be lower as both CUDA Cores have been made larger and large amounts of cache has been added while RT Core count has not been increased (functionality has but not by that much). On AMD side the things are different - RT functional units are even more integrated into the shader pipeline.

If you scrap RT Cores and use the same space for CUDA Cores, you would get up to 5% more of them (most likely less) and adjustment to power levels and clocks might be needed as RT Cores are idle more than CUDA Cores would be. On AMD side, RT functionality is claimed to have much lower area cost, so even less there.

Tensor cores are a different argument. AI is a big thing in the industry and these seem to be architecturally integral enough that cutting these out would not be an easy thing to do. Intel has XMX units and AMD has also started adding hardware bits to support this stuff, next gen will likely bring something they are confident in to call separate units.

And even cutting out both RT and Tensor cores would not bring performance boost to 30%. Price is a whole different discussion though, especially in that market segment.

 

[iameatingjam]

The irony, you can't get a 1440p card to this day for under $780 on the Nvidia side and that's according to Nvidia's marketing itself.

Um, what about the 4070? Its $550, and nvidia themselves say it "lets you max out your favorite games at 1440p."

Not that I want to be put into a position of defending nvidia, that statement just seemed a little... not true, is all.

 

[Vayra86]

Who knows, maybe if @W1zzard puts this poll on the main page, one of the Nvidia employees will see it and say "Look boss, those guys over there are right. The performance/dollar of GPUs in this segment hasn't advanced almost at all because we mainly focus on shoving RT down people's throats."

As if they don't know.

The RT push is intentional, Nvidia made a bet on its high market share. They want stagnation because that's where GPUs are going, while they still need a selling point to upgrade.

My laptop's RTX 3050 does a pretty decent job, considering what it is and that it has 4 GB of VRAM.



You nailed it. Personally, I have the view that the RT vs. no RT debate is largely fueled by clubism and the whole AMD vs. NVIDIA feud. It has been painfully obvious that AMD's hardware has historically been brutally incompetent at handling this type of workload, they were years late to the party, remain a full generation behind and even with RDNA 3, it's not quite there yet, generating resentment that there's such an insane gap between Nvidia and AMD in this regard - and Nvidia has actively exploited this in their marketing material, and took advantage that they were the only vendor that offered complete DirectX 12 Ultimate and hardware accelerated RT to keep their ASP high during the twilight years of GCN and the bug festa that was the original RDNA.

Couple that with the frame rates being relatively poor even on high end hardware (at least until high-end Ampere came about), budget-conscious customers that shop at performance segment or below (which is the only market where AMD has any meaningful presence) as "superfluous", and often baseless arguments are made against the tech, in a futile attempt to "reject" this change.

But it's the future, whether one likes it or not. Games will be heading towards using ray and path tracing, ray traced global illumination, etc. - because that is the only way to improve graphics any further from where they currently stand. True that graphics don't make the game - but some games do need the graphics, and I will admit that I like having games with eye candy and a high degree of immersion... or i'd just have a RTX 3050 on my desktop, too.



In the Navi 21 design, the RT accelerator is bound to a compute unit, each workgroup consisting of two compute units. If you simply doubled it, it'd likely be resource starved and even more inefficient. It'd not necessarily be faster, possibly slower, even.

Non-XT RX 6800 is simply a low-quality Navi 21 die that is 25% disabled. Only three of the four compute slices (each comprised of ten WGPs or 20 CUs) are enabled on this model. Only way to extract more out of this generation would be to enable the remaining units (6800 XT, 6900 XT), and then overclock it further (6950 XT), in an attempt to defy scaling with the tradeoff for power efficiency. If the 6800 achieves 60% of the RT performance of the 6950 XT, then it's already succeeded at what it had to do, but it's not enough: AMD's first-generation RT design in RDNA 2 really sucks compared even to Turing.

Sure its the future. But define 'future'. Might easily take another 10 years.

 

[P4-630]

What if we are in 2030 and are able to run RT on 144fps ultra @4K without DLSS? Are we just going to 8K then or is there something else coming up besides RT?

 

[Onasi]

What if we are in 2030 and are able to run RT on 144fps ultra @4K? Are we just going to 8K then or is there something else coming up besides RT?

Current RT implementation is extremely limited and essentially just boils down to improving some effects and GI. That’s it. We are years, decades probably, from fully ray traced, let alone path traced games. So the question is purely hypothetical.
But going to 8K is arguably a somewhat possible, if insanely expensive, display advancement. Purely because at popular/usable sizes of 27-32 inches 8K gives a PPI that is, essentially, what Apple popularized as Retina. So, theoretically, no further resolution updates would be necessary.

 

[Beginner Macro Device]

But going to 8K is arguably a somewhat possible, if insanely expensive, display advancement.

No one anticipated 4K becoming a go-to option 20 years ago. Today, it's a little bit expensive but accessible for many people, even for those with below average income. Of course having smooth AAA gaming experience at 4K costs a lot of money but not everyone is buying UHD displays for gaming. Some UHD users are never gaming at all.

That said, 8K is about to become the new 4K in a decade or two. Doesn't really matter though since 4K at 27" allows for virtually zero sized pixels and at 32", it's still exceptionally hard to be annoyed.

I wish RT performance will grow faster than it does now. All these RT effects make minimal sense now since nothing can run it (meaning major RT effects and not just "ray traced shadows" tickbox checked), only crawling is possible. But this technology is if not what we need but definitely what is much useful and potent.

 

[Onasi]

No one anticipated 4K becoming a go-to option 20 years ago. Today, it's a little bit expensive but accessible for many people, even for those with below average income. Of course having smooth AAA gaming experience at 4K costs a lot of money but not everyone is buying UHD displays for gaming. Some UHD users are never gaming at all.

The expensive part refers to performance cost in my post, not pricing, since the post I was responding to clearly was talking about gaming performance.
And saying that 4K is becoming a go-to option is a stretch, in my opinion. It’s more accessible than ever, true, but it seems that 1440p is still seen as the more balanced option. And really, I remember when the first 4K displays were coming out and the expectation was that it will become mainstream and the GPU performance will catch up much faster than it actually happened. We are in 2024 basically and 60 FPS at 4K is still something that floors even expensive as hell flagships.

 

[Beginner Macro Device]

And saying that 4K is becoming a go-to option is a stretch, in my opinion. It’s more accessible than ever, true, but it seems that 1440p is still seen as the more balanced option.

1440p has become a go-to option. 4K is still becoming a said option. This is what I meant.

We are in 2024 basically and 60 FPS at 4K is still something that floors even expensive as hell flagships.

Ultra settings don't make as much sense at 4K as they do at 1080p or even 1440p. You don't need that much antialiasing for example. With DLSS/FSR/XeSS advancement, it's also no more mandatory to play at native 4K as it occasionally looks worse than DLSS (and I expect it to become widespread by late 2020s), and yeah, performance gains are much more visible than supersampling artifacts. So whilst techincally being too demanding it's also the most corner cutting forgiving orthodox resolution on the market. With RT on top of that, of course, it's basically unplayable for everyone but maybe 4080+ owners. But RT is currently at its baby steps so it's bold to expect smooth XP with RT.