Support NeoGAF

[Buggy Loop]

But how else would you get all these clicks on your YouTube videos?

Slow news when something so obvious needs a video

They kept the mother of all bombs for slower times :

FSR4 is even better at 8K!

 

[viveks86]

• Squiggly lines on zebra crossings in cyberpunk.
• Shimmering artifacts on godrays through trees in kcd2.
• Similar issues with distant foliage in virtually every open world game with foliage i tried though i did not try it in ac shadows since its already so heavy.
• Blocky artifacts in Space Marine 2 when alpha effects are in play which is all the time since the game is so effects driven. especially in the distance.

this was back at launch a couple of months ago, so maybe things have improved since but im barely able to do dlss quality in these games. makes no sense to take a big performance hit and drop down to dlss 4 balanced or performance anyway.

Interesting! Hadn't heard about any of these

 

[SolidQ]

Tensor Cores in RDNA4 are not as powerful

There no Tensor's in RDNA4. Matrix cores in RDNA5 will fix it

 

[Lysandros]

Are they called Tensor Cores? Thought this was patented by NVIDIA. Techpowerup calls them Tensor Cores, but no documentation from AMD calls them that.

I am perplexed as well, first time seeing such a statement.

 

[JohnnyFootball]

On topic, it sounds like fsr4 quality is somewhere between dlsse and dlss4.

I've not been using dlss4 because of the performance hit and several issues with artifacts and am wondering if i would be able to use fsr4 instead of dlss3 going forward.

No FSR4 is exclusive to AMD and will likely be that way permanently.

 

[JohnnyFootball]

If AMD starts to be competent I might buy them my next graphics card

They're already there. Love my 9070XT in my home theater PC.

If only HDMI 2.1 support existed in Linux….

 

[Gaiff]

I am perplexed as well, first time seeing such a statement.

Seems it has just become a generic name for matrix multiplication accelerators. Tensor Cores is definitely an NVIDIA-only thing, but AMD has similar hardware, just with a different name. It's a bit like people calling ray tracing RTX when RTX is specifically an NVIDIA thing.

 

[Black_Stride]

• Squiggly lines on zebra crossings in cyberpunk.
• Shimmering artifacts on godrays through trees in kcd2.
• Similar issues with distant foliage in virtually every open world game with foliage i tried though i did not try it in ac shadows since its already so heavy.
• Blocky artifacts in Space Marine 2 when alpha effects are in play which is all the time since the game is so effects driven. especially in the distance.

this was back at launch a couple of months ago, so maybe things have improved since but im barely able to do dlss quality in these games. makes no sense to take a big performance hit and drop down to dlss 4 balanced or performance anyway.

How are/were you injecting DLSS4 into Space Marine 2?
If you havent updated the game use the NVApp to inject DLSS4

DLSS4 is a tier or more above DLSS3 in every mode.
So "going down" to DLSS4 Balanced or Performance would offer you better visuals than DLSS3 Quality.
DLSS3 has no advantages in any game ive tested.
Even forced Ultra Perf can look better than DLSS3 Balanced.

 

[Loxus]

Are they called Tensor Cores? Thought this was patented by NVIDIA. Techpowerup calls them Tensor Cores, but no documentation from AMD calls them that.

They aren't called Tensor Cores, but it clear the AI Accelerators are equivalent to Nvidia's Tensor Cores. The AI Accelerators also support similar formats to the Matrix Cores in CDNA.

On a said note, AMD does mention the word Tensor though.

16bit / 8bit / 4bit Tensor Input Types

 

[nowhat]

Cerny is correct. But Amethyst started at a later date.
Just consider that the Radeon 9070 and FSR4 were released over a month ago and there is no FSR4 on the Pro.
The collaboration probably started when Cerny made that interview. So it will take some time until we see the results of such collaboration.
And it will be beneficial for AMD and Sony.

I was going to post something cheeky... but nah, I'm hungover and just want to rest. I guess I wouldn't be that funny anyway.

But as you seem to have some insight into this stuff, has there been any "brain drain", so to speak, from Sony to AMD? At all? They clearly have been collaborating up to a point with each other, for some years. So, Z-buffer? Anything?

 

[Gaiff]

They aren't called Tensor Cores, but it clear the AI Accelerators are equivalent to Nvidia's Tensor Cores. The AI Accelerators also support similar formats to the Matrix Cores in CDNA.

On a said note, AMD does mention the word Tensor though.

16bit / 8bit / 4bit Tensor Input Types

Tensor isn't surprising. Tensor is a type of operation predating Tensor Cores. Even Google has a Tensor Processing Unit. I was just curious about Tensor Cores specifically because Techpowerup on the specs sheet also mentions them for the 9070 XT.

It says it had 128 cores compared to 280 for the 5070 Ti. I didn't check, but this disparity is strange, unless of course, they're not 1 to 1 which I suspect they aren't, hence Tensor Cores is a bit of a misnomer.

 

[winjer]

I was going to post something cheeky... but nah, I'm hungover and just want to rest. I guess I wouldn't be that funny anyway.

But as you seem to have some insight into this stuff, has there been any "brain drain", so to speak, from Sony to AMD? At all? They clearly have been collaborating up to a point with each other, for some years. So, Z-buffer? Anything?

Sony just sets the hardware specs they want and AMD design the chips. Then TSMC produces them.

 

[nowhat]

Sony just sets the hardware specs they want and AMD design the chips. Then TSMC produces them.

Makes sense. But just regarding the Z-buffer - it was just a stopgap measure really, to make checkerboading easier on a PS4 Pro (and yes, it doesn't mean that there's really hardware support for it, just hardware to make the process more efficient). And, that kind of seems like a feature that could be beneficial in general, given API support. But that never made its way to AMD GPUs?

 

[Javi97]

PSSR was almost ready before AMD and Sony partnered

FSR4 was created based on the research and progress made by AMD and Sony

Now Cerny is looking to make PSSR more like FSR4

 

[winjer]

Makes sense. But just regarding the Z-buffer - it was just a stopgap measure really, to make checkerboading easier on a PS4 Pro (and yes, it doesn't mean that there's really hardware support for it, just hardware to make the process more efficient). And, that kind of seems like a feature that could be beneficial in general, given API support. But that never made its way to AMD GPUs?

Z-buffer is a technique that has several decades. Many generations of GPUs and consoles have it.
It's just a buffer where the depth values are sorted and stored. And it's the basis for a ton graphical techniques and effects.

 

[nowhat]

Z-buffer is a technique that has several decades. Many generations of GPUs and consoles have it.
It's just a buffer where the depth values are sorted and stored. And it's the basis for a ton graphical techniques and effects.

I know, but there was some particular HW on a Pro that was for that purpose.

But anyway. Why do I even care. I need some ibuprofen. Or a beer. Probably beer it is.

 

[Magic Carpet]

Is there anyway to test FSR4 using my Nvidia RTX 4070? I can enable FSR in many games but I hear FSR4 only works on AMD cards.

 

[winjer]

Is there anyway to test FSR4 using my Nvidia RTX 4070? I can enable FSR in many games but I hear FSR4 only works on AMD cards.

So far, it only works on RDNA4. Will work on RDNA3 soon. But that's it.
Truth be told, with a 4070, you don't need FSR4. Just use DLSS4, as it's better and optimized for your GPU.

 

[Magic Carpet]

So far, it only works on RDNA4. Will work on RDNA3 soon. But that's it.
Truth be told, with a 4070, you don't need FSR4. Just use DLSS4, as it's better and optimized for your GPU.

I just want to see how it compares with DLSS in person.

 

[Loxus]

Tensor isn't surprising. Tensor is a type of operation predating Tensor Cores. Even Google has a Tensor Processing Unit. I was just curious about Tensor Cores specifically because Techpowerup on the specs sheet also mentions them for the 9070 XT.

It says it had 128 cores compared to 280 for the 5070 Ti. I didn't check, but this disparity is strange, unless of course, they're not 1 to 1 which I suspect they aren't, hence Tensor Cores is a bit of a misnomer.

Techpowerup may just have a standard labeling format in their specs sheet table.

To understand the 128 and 280 numbers, you have to look at these two images.

WGP = SM (with each having 4 × Tensor Cores/AI Accelerators)

The 5070TI has 70 SM, so 70 × 4 = 280 Tensor Cores.

The 9070XT has 32 WGP, so 32 × 4 = 128 AI Accelerators.

So the 9070XT has almost half the Tensor/AI Cores as the 5070TI.

A big question is, does this mean AMD's AI Accelerators are more powerful than Nvidia's Tensor Cores based on the graph in the video?

 

[Bojji]

So DLSS4 > FSR4 > DLSS3 > PSSR = XeSS > FSR 3

This is the correct order at the moment. Pssr has potential to catch up to dlss3 next year.

 

[Lysandros]

Seems it has just become a generic name for matrix multiplication accelerators. Tensor Cores is definitely an NVIDIA-only thing, but AMD has similar hardware, just with a different name. It's a bit like people calling ray tracing RTX when RTX is specifically an NVIDIA thing.

I'll stick with the already existent terminology to avoid confusion. Besides, i think the difference in architecture of those units and approaches is enough to warrant it.

 

[Buggy Loop]

So the 9070XT has almost half the Tensor/AI Cores as the 5070TI.

A big question is, does this mean AMD's AI Accelerators are more powerful than Nvidia's Tensor Cores based on the graph in the video?

uh?

None of the solutions take up all the tensor cores

There's only so many TOPs you need to make the inference for DLSS. It runs from 5090 all the way down to a 2050 laptop chipset. Throwing more at it wouldn't necessarily lower the latency. Its already quite efficient.

Nvidia has a lot of TOPs leeway for neural denoising, neural path tracing, neural texture, neural this neural that, etc.

 

[Lysandros]

But as you seem to have some insight into this stuff, has there been any "brain drain", so to speak, from Sony to AMD? At all? They clearly have been collaborating up to a point with each other, for some years. So, Z-buffer? Anything?

Are you talking about the PS4 PRO's custom ID Buffer hardware? That is a Sony idea unique to PS4 PRO. It was designed to offload some of the processing cost of reconstruction algorithms such as checkerboard rendering. It can also help with other forms of reconstruction and anti-aliasing. We know it's also present in PS5 and plausibly in PS5 PRO even. It has no equivalent in any AMD desktop hardware. Cache Scrubbers are also unique to Playstation hardware like Cerny stated plenly in Road to PS5. Same is true for the entire custom I/O complex of course. PS4 had also custom/unique bits before like the volatile bit and Onion+ bus. Sony and AMD have a pretty long history of collaborating in such customizations.

 

[Physiognomonics]

But it's not ready on the PS5 Pro, while it's ready on RDNA4 GPUs, for over a month.
So it's clear that AMD started work on FSR4 much sooner than Sony.
Chances are that Sony saw that FSR4 was superior to PSSR and decided to ditch it and join AMD in developing FSR4 for the Pro as well.

You get it wrong. PSSR is the prototype of FSR4. Cerny's team went to AMD and made with them an AI upscaling tailored to one RDNA4 GPU using their experience with PSSR (which is why they did it in 1 year, because they already worked on PSSR for years beforehand). They had to do this way because machine learning hardware are very different. FSR4 is both owned by AMD and Sony. Now they have to make it work on a very different hardware, PS5 Pro which has a similar TOPs power than a 9700 albeit with different instructions.

 

[mckmas8808]

But it's not ready on the PS5 Pro, while it's ready on RDNA4 GPUs, for over a month.
So it's clear that AMD started work on FSR4 much sooner than Sony.
Chances are that Sony saw that FSR4 was superior to PSSR and decided to ditch it and join AMD in developing FSR4 for the Pro as well.

But......Sony is still updating PSSR today. So, your take doesn't seem to align with reality.

 

[winjer]

But......Sony is still updating PSSR today. So, your take doesn't seem to align with reality.

Of course Sony is updating PSSR. So is AMD updating FSR4. And Nvidia updating DLSS.
But it's going to take time to put FSR4 into the Pro.