Contemplating about FSR 4..

[EdoTensei]

As finally AMD choose the bright side of the force, and officially confirmed FSR 4 fully based on AI, and it's in progress for 9-12 months, whatever that mean with AmD , nevertheless.. even tho AMD didn't specified which RDNA will get FSR 4. 

 

RDNA 4 comes with AI (as Sony claims) so GPUs as well. But the question is, if that will be the 2nd Gen of what is already in RDNA 3, or it's something completely new, and RDNA 3 has only some sort of "AI instructions", especially because there is nothing about it on TechPowerUp which has all information about RT and Tensor cores in Nvidia. 

 

Any thoughts ? 

[ki8aras]

17 minutes ago, EdoTensei said:

As finally AMD choose the bright side of the force, and officially confirmed FSR 4 fully based on AI, and it's in progress for 9-12 months, whatever that mean with AmD , nevertheless.. even tho AMD didn't specified which RDNA will get FSR 4. 

 

RDNA 4 comes with AI (as Sony claims) so GPUs as well. But the question is, if that will be the 2nd Gen of what is already in RDNA 3, or it's something completely new, and RDNA 3 has only some sort of "AI instructions", especially because there is nothing about it on TechPowerUp which has all information about RT and Tensor cores in Nvidia. 

 

Any thoughts ? 

sony said theres tech in the ps5 pro that isnt yet available on the consumer amd gpus

[EdoTensei]

5 minutes ago, ki8aras said:

sony said theres tech in the ps5 pro that isnt yet available on the consumer amd gpus

Meant Ray Tracing 

[Salted Spinach]

1 hour ago, EdoTensei said:

Any thoughts ? 

If you are making a purchasing decision now, base it on what is available now

 

If you are speculating on a big payout in the future from a purchasing decision now, buy some crypto

 

If you want certainty about future events, look for an fortune teller

[ChrisLoudon]

2 hours ago, EdoTensei said:

Any thoughts ? 

Without knowing what the Nvidia 50xx pricing is going to look like, its hard to make any decisions on what to buy for the next round of GPUs.

 

The only 2 things I / we all know for sure is....

 

1. DLSS is better than FSR and that is likely to continue as each team revises their upscaling tech given that Nvidia had a decent head start on AMD.

 

2. AMD are shying away from the high-end with their RDNA 4 cards.

[Eigenvektor]

1 hour ago, ChrisLoudon said:

AMD are shying away from the high-end with their RDNA 4 cards.

Of course we don't know for certain what exactly "high-end" means in this context. No 5070 equivalent? No 5080 equivalent?No 5090 equivalent?

 

Unless you're in the market for a high-end card, it doesn't really matter.

 

If you're in the market for a 5090, you probably wouldn't have looked at a (presumed) 8900 XTX anyway.

[porina]

3 hours ago, EdoTensei said:

Any thoughts ? 

I'm sure it'll eventually happen. People will test it when it does. Hopefully it does address the major image quality problems with current FSR and finally allow it to play on the same level as DLSS and XeSS. 

 

For those not following, an AMD SVP was talking about this at IFA so this is official, not leak/rumour. The context it was presented in was focused on improving mobile battery life, and how to help with that in GPU performance. So we have some open questions. Is this a "mobile first" technology? Maybe other support like dGPU will follow not necessarily at the same time, even depending on gen. IOW what will be supported, and when? AMD have made FSR open source, presumably this will continue.

 

Interesting comment was that current FSR path (without AI) was taken as it was fast to market. 

[Bagzie]

11 hours ago, EdoTensei said:

As finally AMD choose the bright side of the force, and officially confirmed FSR 4 fully based on AI, and it's in progress for 9-12 months, whatever that mean with AmD , nevertheless.. even tho AMD didn't specified which RDNA will get FSR 4. 

 

RDNA 4 comes with AI (as Sony claims) so GPUs as well. But the question is, if that will be the 2nd Gen of what is already in RDNA 3, or it's something completely new, and RDNA 3 has only some sort of "AI instructions", especially because there is nothing about it on TechPowerUp which has all information about RT and Tensor cores in Nvidia. 

 

Any thoughts ? 

I'll be curious to see if it will be active with RDNA 3 aswell since they have AMD's tensor core equivuilent which are currently unused I beleive.

[EdoTensei]

9 hours ago, Bagzie said:

I'll be curious to see if it will be active with RDNA 3 aswell since they have AMD's tensor core equivuilent which are currently unused I beleive.

As far as I can tell, you can use them with some Apps for AI making pictures, or some other stuff strictly related to AI, but Anti-Lag 2 would use AI cores as well, but not sure if only when will be in the Adrenaline software. 

 

So yeah, in gaming those cores are unused for a moment. Let's hope to see if FSR 4 will actually working on RDNA 3, cuz 168 Ai Cores sounds pretty good in my 7900XT imho , that's also few cores less than 4070 has Tensor cores, but it may be different architecture so the numbers may be different as well. 

[porina]

Intel used DP4a instructions on non-Arc GPUs for XeSS. Support for that goes back to Pascal and RDNA1, so no dedicated AI hardware support needed if FSR4 will go back that far. AMD could branch it like Intel did, with a native implementation for recent AMD hardware, and a generic implementation for others. For Yakuza Like a Dragon: Infinite Wealth, I actually preferred the generic XeSS over DLSS (on NV GPU!) due to its better temporal stability.

[Bagzie]

1 hour ago, porina said:

Intel used DP4a instructions on non-Arc GPUs for XeSS. Support for that goes back to Pascal and RDNA1, so no dedicated AI hardware support needed if FSR4 will go back that far. AMD could branch it like Intel did, with a native implementation for recent AMD hardware, and a generic implementation for others. For Yakuza Like a Dragon: Infinite Wealth, I actually preferred the generic XeSS over DLSS (on NV GPU!) due to its better temporal stability.

When i had a 7900XT i used XESS when i could because of its better stability over FSR but the problem was the performance increase from XESS atleast on AMD was quite limited in comparison to FSR.

4 hours ago, EdoTensei said:

As far as I can tell, you can use them with some Apps for AI making pictures, or some other stuff strictly related to AI, but Anti-Lag 2 would use AI cores as well, but not sure if only when will be in the Adrenaline software. 

 

So yeah, in gaming those cores are unused for a moment. Let's hope to see if FSR 4 will actually working on RDNA 3, cuz 168 Ai Cores sounds pretty good in my 7900XT imho , that's also few cores less than 4070 has Tensor cores, but it may be different architecture so the numbers may be different as well. 

It's less tensor cores but then AMD doesn't use its tensor cores for frame generation.

 

Honestly I'm kinda convinced frame generation needing tensor cores is just nvidias way of selling 4000 series , I tried fsr frame gen in Avatar and thought it was just as good as any nvidia frame generation exclusives.

 

FSR upscaling whilst not absolutely terrible is still quite a way behind though , The shimmering it introduces in cyberpunk is a bit of a pain in the ass.

[toasty99]

There's really no reason to have any confidence in AMD. They are saying 9-12 months out for fsr4, but that's assuming no delays. 

 

We know they are leaving the high end with their next product launch. They are framing it as a "choice", but considering Intel now exists in the market occupying that low end space logically it's not a wise choice. This implies the decision is more likely due to their architecture simply not being competitive enough to be at the high end.

 

I'm quite wary that RDNA 3 will get the Vega treatment and get axed really early. I'd avoid AMD at this time. The only cards that are really competitive is last gen's 6700xt which obviously destroys the 4060 in its price class, the 6600 which destroys the 3050, and the 6600xt/6650xt/6750xt/6700 that are either slightly better or slightly worse varients of the above cards. 

 

Once you get to $500, just go Nvidia. Nvidia still supports the Gtx 900 series from 2014, AMD doesn't support the Radeon VII that was released in 2019.

 

 

[porina]

3 hours ago, Bagzie said:

When i had a 7900XT i used XESS when i could because of its better stability over FSR but the problem was the performance increase from XESS atleast on AMD was quite limited in comparison to FSR.

In the title I mentioned, the performance difference between FSR/DLSS/XeSS at same quality preset wasn't significant, but FSR was by far the worst visual quality. There are limits to what I find acceptable and FSR did not meet them. A common problem not limited to that game is character hair fizzes which I find really distracting. I also saw that in Forspoken when I tried that to test out FSR3.

 

3 hours ago, Bagzie said:

Honestly I'm kinda convinced frame generation needing tensor cores is just nvidias way of selling 4000 series , I tried fsr frame gen in Avatar and thought it was just as good as any nvidia frame generation exclusives.

NV FG requires 40 series due to the bigger optical flow unit. It was present in 30 series too but at a much smaller capacity. In a similar way NPUs from AMD before Ryzen AI also existed, but did not meet the current Copilot+ requirement.

 

I haven't done much testing on FSR3 FG myself, but since it got decoupled from requiring FSR2 upscaling it might make it actually viable. In Forspoken it seemed to work well enough but was held back by being tied to FSR2 upscaling at the time.

[Bagzie]

5 hours ago, porina said:

In the title I mentioned, the performance difference between FSR/DLSS/XeSS at same quality preset wasn't significant, but FSR was by far the worst visual quality. There are limits to what I find acceptable and FSR did not meet them. A common problem not limited to that game is character hair fizzes which I find really distracting. I also saw that in Forspoken when I tried that to test out FSR3.

 

NV FG requires 40 series due to the bigger optical flow unit. It was present in 30 series too but at a much smaller capacity. In a similar way NPUs from AMD before Ryzen AI also existed, but did not meet the current Copilot+ requirement.

 

I haven't done much testing on FSR3 FG myself, but since it got decoupled from requiring FSR2 upscaling it might make it actually viable. In Forspoken it seemed to work well enough but was held back by being tied to FSR2 upscaling at the time.

What optical flow unit? Optical flow is hardware accelerated with tensor cores on nvidia and asynchronous compute on cards that don't have it.

[porina]

52 minutes ago, Bagzie said:

What optical flow unit? Optical flow is hardware accelerated with tensor cores on nvidia and asynchronous compute on cards that don't have it.

I never looked at it in detail since the release of Ada, but following nvidia page seems to sum the difference between Ada and previous.

 

Quote

In the NVIDIA Turing and NVIDIA Ampere architecture generation GPUs, most of these algorithms use a compute engine to perform the required tasks. As a result, when the compute engine workload is high, the performance of the NVIDIA Optical Flow Accelerator (NVOFA) could be affected.

 

On NVIDIA Ada-generation GPUs, most of these algorithms are moved to dedicated hardware within the NVOFA, reducing the dependency on the compute engine significantly.

https://developer.nvidia.com/blog/harnessing-the-nvidia-ada-architecture-for-frame-rate-up-conversion-in-the-nvidia-optical-flow-sdk/

 

[Bagzie]

4 hours ago, porina said:

I never looked at it in detail since the release of Ada, but following nvidia page seems to sum the difference between Ada and previous.

 

https://developer.nvidia.com/blog/harnessing-the-nvidia-ada-architecture-for-frame-rate-up-conversion-in-the-nvidia-optical-flow-sdk/

 

Marketing gobbledygook.

 

Why would the 3000 series have an optical flow accelerator if it doesn't have frame generation.

 

Its just tensor cores with an optical flow algorithm.

 

 

[porina]

6 hours ago, Bagzie said:

Why would the 3000 series have an optical flow accelerator if it doesn't have frame generation.

For uses other than gaming FG, such as video interpolation effects. Regardless, the main point remains the hardware supporting it is different with Ada.

 

[Mihle]

I am considering getting AMD instead of Nvidia, even if DLSS is better, because the extra VRAM. I am very sceptical if 12GB is enough.

[Bagzie]

10 hours ago, porina said:

For uses other than gaming FG, such as video interpolation effects. Regardless, the main point remains the hardware supporting it is different with Ada.

 

Interpolation is different from optical flow , My tv can do interpolation.

 

Tensor cores do optical flow , there is no specific optical flow hardware unit.

[porina]

1 hour ago, Bagzie said:

Interpolation is different from optical flow , My tv can do interpolation.

 

Tensor cores do optical flow , there is no specific optical flow hardware unit.

I'm not big into video, but video software has supported the use of optical flow for the generation of additional frames which I called interpolation. If my use of the term interpolation in that context is off, that's on me. Anyway, I gave the previous NV link that stating there is additional hardware supporting optical flow in Ada not present in Ampere or earlier, which still supported optical flow.

 

Edit: I also wonder if we're talking about slightly different things. Looking back, my use of the term unit might have suggested a physical dedicated area. Looking at the NV page again, the wording does, like you said, suggest optical flow in Ampere and earlier was implemented with general compute. It mentions that if you're already using that compute for other things at the same time, then that could be a problem. Ada aimed to resolve that by adding dedicated resource for the purpose. So optical flow may be better viewed as a logical function - is that why we're having this argument?

[WallacEngineering]

On 9/18/2024 at 2:50 AM, EdoTensei said:

Meant Ray Tracing 

I was wondering about this particular part of the new PS5 Pro as well. Its an AMD based system like pretty much all Consoles in recent history and yet Sony was able to come up with what, maybe a Ray Accelerator chip that enhances Ray Tracing performance to be on par with Nvidia? Because we all know one thing AMD isn't good at - is Ray Tracing.

 

This to be suggests that RDNA-4 could possibly close the gap in Ray-Tracing performance significantly with Nvidia. Mind you I already find the Ray-Tracing performance of my 7900-XTX to be quite reasonable and I don't really care about Ray-Tracing features anyways and so I went with AMD for the value per dollar - $900 (after discounts) versus the $1230 of the cheapest RTX 4080 I could find at the time.

 

But basically this tells me that AMD may be hiding an ace up their sleeve, even though they have already announced their departure from high-end GPUs with the next generation of Radeon.

 

But how weird would it be if the successor to the RX 7800-XT (What I believe will likely be their highest tier offering considering the withdrawal from high-end) actually matched the RTX 4070/Super's Successor (RTX 5070) in Ray-Tracing performance? That would certainly come as a "shot from left-field" if you ask me, not something we would expect AMD to be able to pull off.

[WallacEngineering]

13 hours ago, Mihle said:

I am considering getting AMD instead of Nvidia, even if DLSS is better, because the extra VRAM. I am very sceptical if 12GB is enough.

12 GB is enough for any resolution under 4K, although just barely, especially in the case of my particular resolution - 3440x1440p UltraWide.

 

Additional VRAM is indeed another reason I was attracted to the much less expensive $900 7900-XTX rather than the $1230 RTX 4080 at the time. Although I would say 24GB is certainly overkill for anything I would ever do with the card, I do appreciate that Radeon either has the correct amount of VRAM for a given tier or is even generous with VRAM. Thats something Nvidia desperately needs to improve on.

 

Basically, it stacks up like this, 1080p needs 8GB of VRAM, 1440p needs 12GB, and 4K needs 16GB. And thats not leaving any headroom, modern games will absolutely use every MB of that VRAM. For anything particularly extreme like 4K at absolutely Ultra-Everything settings while also using additional features, or 4K UltraWide (5040x2160p) you may actually need 20GB of VRAM, but those are extreme examples.

 

16GB seems to be the ideal target these days for most people. It would give you enough VRAM to pretty much do anything you want in modern games without going excessive. This is precisely why the RX 6800/7800/XT cards are considered the best overall value gaming cards right now and why they are constantly recommended to those seeing a good, but somewhat budget conscious option. Its because they genuinely offer pretty damn good 1440p performance while also having more than enough VRAM to not even have to think about it. Such an option does not exist for Nvidia.

[toasty99]

1 hour ago, WallacEngineering said:

12 GB is enough for any resolution under 4K, although just barely, especially in the case of my particular resolution - 3440x1440p UltraWide.

 

Additional VRAM is indeed another reason I was attracted to the much less expensive $900 7900-XTX rather than the $1230 RTX 4080 at the time. Although I would say 24GB is certainly overkill for anything I would ever do with the card, I do appreciate that Radeon either has the correct amount of VRAM for a given tier or is even generous with VRAM. Thats something Nvidia desperately needs to improve on.

 

Basically, it stacks up like this, 1080p needs 8GB of VRAM, 1440p needs 12GB, and 4K needs 16GB. And thats not leaving any headroom, modern games will absolutely use every MB of that VRAM. For anything particularly extreme like 4K at absolutely Ultra-Everything settings while also using additional features, or 4K UltraWide (5040x2160p) you may actually need 20GB of VRAM, but those are extreme examples.

 

16GB seems to be the ideal target these days for most people. It would give you enough VRAM to pretty much do anything you want in modern games without going excessive. This is precisely why the RX 6800/7800/XT cards are considered the best overall value gaming cards right now and why they are constantly recommended to those seeing a good, but somewhat budget conscious option. Its because they genuinely offer pretty damn good 1440p performance while also having more than enough VRAM to not even have to think about it. Such an option does not exist for Nvidia.

The rtx 3090 does exist on the used market. If you can find one in the $600 range, I'd still recommend that over anything new.

 

At $700+ where it goes on eBay, I would avoid but when the price is right that's the ideal 1440p/4k GPU.

[WallacEngineering]

53 minutes ago, toasty99 said:

The rtx 3090 does exist on the used market. If you can find one in the $600 range, I'd still recommend that over anything new.

 

At $700+ where it goes on eBay, I would avoid but when the price is right that's the ideal 1440p/4k GPU.

True, a used RTX 3090 is indeed a good grab for the right price. However, because it was originally $2000 it rarely goes low enough to be a good deal. Check out this graph, GPU Comparisons from current and previous generations over a 10-game average at 1440p Ultra. You can get a 7800-XT for as low as $460 right now on NewEgg and it nearly matches the performance of an RTX 3090.

 

 

So really, the RTX 3090 would have to be $500 or less to make any sense with current GPU Pricing, and thats not even accounting for the fact that we are talking about a brand new retail card from NewEgg with a warranty versus a used card off eBay. Personally, I wouldn't touch a used RTX 3090 over $400 even.

[toasty99]

1 hour ago, WallacEngineering said:

True, a used RTX 3090 is indeed a good grab for the right price. However, because it was originally $2000 it rarely goes low enough to be a good deal. Check out this graph, GPU Comparisons from current and previous generations over a 10-game average at 1440p Ultra. You can get a 7800-XT for as low as $460 right now on NewEgg and it nearly matches the performance of an RTX 3090.

 

 

So really, the RTX 3090 would have to be $500 or less to make any sense with current GPU Pricing, and thats not even accounting for the fact that we are talking about a brand new retail card from NewEgg with a warranty versus a used card off eBay. Personally, I wouldn't touch a used RTX 3090 over $400 even.

I like it at $600 used for the following reasons:

 

1. 24gb vram. Don't have to concern yourself whether you have enough.

 

2. DLSS. It's better than FSR for now. Sure it lacks Nvidia frame gen, but AMD's option for that is good enough imo and works on 30 series cards. 

 

3. Other options are horribly crippled. Every AMD card (even the 7900xtx) can't really do ray tracing in the more intense games like Cyberpunk, the 3090 can. The 4070, 4070 super, and 4070ti are horribly crippled with 12gb of vram. 

 

4. The options that aren't crippled in some way start at the 4070ti super at $800 frankly. That's a lot more money, and even then you can argue the 16gb is possibly an issue down the road.

 

I can understand warranty being a concern for some, personally I never really worry too much about that because these companies are so awful to deal with I just assume there isn't a warranty lol. Zotac is the only company I've had a decent experience with on a warranty, but that's personal opinion.