AMD reveals Ryzen 5000 G-Series desktop APUs, FidelityFX Super Resolution and more

DLSS is an important killer feature! Well, it just got killed.

AMD fan here. Yet, IMHO it turned out as expected.
You can't match DLSS without an AI/ML approach.
And it unfortunately shows in these videos.
The parts with FSR simply don't look as good.

This doesn't mean it's not good for anything.
In particular, as it is available across manufacturers
and a large variety of GPU models.

But it did not kill DLSS imho.

Nvidia has two game changers for gaming platforms on their side.
ARM (see Apple's M1 performance and efficiency) and DLSS.

In the long run it might get much harder again for AMD
when it comes to gaming and consoles.
Good for AMD that the current gen platforms have just been released.
I'm 100 % sure, were they still in the concept phase,
it would've been an instant design win for Nvidia.

I don't get it why not on ALL polaris cards? after all a RX 480 is the same architecture as a RX 580. I suppose if they release the code this has great chances of being backported to older gpus.

Well, Joshua Ashton has just implemented Vulkan ray tracing on cards that don't have the hardware for it, so who can say?

Get itch. Scratch itch.

Ah, Liam's got an article up about that, now.


Last edited by CatKiller on 1 June 2021 at 10:44 am UTC

I think for an anti-feature, it's good enough to end DLSS. Because it works everywhere and will also get better over time. The other side of it - it's general purpose. While AI/ML is more limited to specific use cases. So it is better in some cases and worse in others. Everything is a trade off.


Last edited by Shmerl on 1 June 2021 at 4:21 pm UTC

Nvidia has two game changers for gaming platforms on their side.
ARM (see Apple's M1 performance and efficiency) and DLSS.

DLSS sure but I don't see how ARM is going to help them with PC gaming performance

Nvidia has two game changers for gaming platforms on their side.
ARM (see Apple's M1 performance and efficiency) and DLSS.

Don't get it. Why would a game console developer NOW pick Nvidia? because they own ARM? Why they haven't picked an ARM earlier? Mali was always an option or Tegra with or without Nvidia owning the "specification" comapny. With a licence everbody can design an ARM. If Nvidia changes this they can throw away ARM again everybody will switch to RISC V.

DLSS but could be the same as raytracing. If some vendor asks I have no doubt any gpu manufacturer will deliver no matter if AMD, Qualcomm, Apple or Intel.

Nvidia has two game changers for gaming platforms on their side.
ARM (see Apple's M1 performance and efficiency) and DLSS.

Don't get it. Why would a game console developer NOW pick Nvidia? because they own ARM? Why they haven't picked an ARM earlier? Mali was always an option or Tegra with or without Nvidia owning the "specification" comapny. With a licence everbody can design an ARM. If Nvidia changes this they can throw away ARM again everybody will switch to RISC V.

DLSS but could be the same as raytracing. If some vendor asks I have no doubt any gpu manufacturer will deliver no matter if AMD, Qualcomm, Apple or Intel.

As for ARM. Indeed, they could've licensed it before.
But the point is, that with implementations like M1 more than competitive horse power is now available at high efficiency.

More important DLSS looks like a real game changer to me.
I think you're wrong thinking this is easy to copy by competitors.
Nvidia has a aggregated a lot of know how in their software department plus they have silicons out, that provide hardware acceleration for AI operations (Tensor Cores).
Also I don't know if and how much of the DLSS stuff is patented - I guess it's a lot.

Stuff like upscaling is actually an ideal category for AI/DL,
that cannot simply be matched with a classical scaling and filtering algorithm.
It shows. FSR doesn't look that good in comparison. Plus it's slower.

There are ingame videos from games with DLSS upscaled from 1080p to 2160p running almost twice as fast and looking absolutely credible like they were rendered in the higher resolution.

I bet AMD wanted to go the same way but could not (lacking hardware support and probably patenting).

Rumor has it Nintendo will update their Switch (Switch Pro) to a version leveraging DLSS.
It's clever. Exactly what I would Nintendo expecting to do.
And you can get it from just one vendor: Nvidia.
Same like the AMD situation with the current (and previous) SONY and Microsoft consoles.

I believe FSR will do to DLSS the same thing Freesync did to G-Sync. It will become the industry standard and eventually Nvidia will be forced to officially support it too.

It will be "good enough" in up-scaling the image and most people will not see the differences, like MP3 and FLAC. The difference is there, but most gamers will not care.

I think DLSS is also a kind of a trade off idea. It's sold as some kind of improvement but in fact it also reduces quality in general like any upscaling does. I'd personally prefer lower resolution with higher quality to higher resolution with lower one. So I don't see a big appeal in upscaling just to put a checkmark that "I have higher resolution". GPUs are improving without upscaling tricks anyway, so what's the rush? For higher resolution there will be more powerful GPUs.


Last edited by Shmerl on 1 June 2021 at 8:04 pm UTC

It will be interesting to see how FSR fares in situations where DLSS (and temporal AA methods) struggle. According to the GPUOpen announcement, FSR is said not to require motion vectors, but maybe it requires the scene to have been antialiased with a temporal method beforehand to look good (which means it needs motion vectors indirectly).


Last edited by Calinou on 1 June 2021 at 9:40 pm UTC

I've watch Linus Tech Tips do a video on Cyberpunk featuring DLSS and it was underwhelming. When you see it in action side-by-side with the original image, it always suffers in the quality department.

But that's not how you game. You typically only have your one gaming PC and that's what you use. If you can play about with DLSS on/off and 4K vs 1080p, then you'll find your sweet spot.

So FSR is huge news for me. It's DLSS-like enough that I'm excited to try it. Because I went AMD last year, and I ain't going back to Nvidia any time soon. FSR is going to be open source for goodness sake. Absolutely love what AMD is doing right now.

If only GPUs were actually available...

If only GPUs were actually available...

Supposedly the situation should get better later this year. Still waiting to get Sapphire Pulse RX 6800 XT myself.

But the point is, that with implementations like M1 more than competitive horse power is now available at high efficiency.

Still don't get it why does this is a win for nvidia? The Apple M1 is not a Cortex design. Thus nvidia has no properties there and would still need to develop something. The M1 is ARMv8 instruction compatible but is otherwise Apple own design. It is like saying Intel is the go to CPU supplier because AMD Ryzen shown good performance. You could do a fast general purpose CPU with every modern instruction set.

I think for an anti-feature, it's good enough to end DLSS. Because it works everywhere and will also get better over time. The other side of it - it's general purpose. While AI/ML is more limited to specific use cases. So it is better in some cases and worse in others. Everything is a trade off.

This sounds like it was written by someone who doesn't understand either of these features. Hardware-assisted AI/ML being "limited" to specific use cases while an upscaling tech being "general purpose"? Give me a break...

Honestly, you sound like you don't know how Machine Learning and Artificial Intelligence works. It's typically incredibly focused. Using ML/AI in an implementation of anything doesn't magically make it versatile or flexible.

However, that said, the point of this technology is to enable fast frame rates of complex scenes (potentially with ray tracing thrown in) at high resolutions. That's a pretty focused target. So I'm not really sure why an upscaler like FSR is somehow more "general purpose" than DLSS.

But I agree with Shmerl that FSR can be used anywhere - mobiles, consoles, Windows, Linux, on AMD, Nvidia, ARM, Intel, you name it. So I sincerely hope it does well. DLSS can do well too... for all I care. But I don't care, since I doubt I'll ever buy Nvidia again, unless they start competing with AMD on the open source front. And DLSS doing well only benefits Nvidia.

As someone else has already pointed out, it's G-Sync vs Freesync all over again.

(as an aside, I actively avoided buying a monitor advertising g-sync support recently - it looked like a nice piece of hardware, but I knew that if I bought it, I'd be paying for a feature that would forever be locked from me. This sums up my view on Nvidia right now. Hopefully they change that view by following AMD's example in future)

However, that said, the point of this technology is to enable fast frame rates of complex scenes (potentially with ray tracing thrown in) at high resolutions. That's a pretty focused target. So I'm not really sure why an upscaler like FSR is somehow more "general purpose" than DLSS

I think if I understood correctly FSR makes generic assumptions, while DLSS has to be trained on specific input (unless I misunderstood the idea). Generic assumption sounds like a broader approach to me. Trained neural network can produce better result for what it's trained for, but will be close to useless for other cases. Something that's more generic is probably in between. So both are a trade off I think.


Last edited by Shmerl on 2 June 2021 at 9:30 pm UTC

However, that said, the point of this technology is to enable fast frame rates of complex scenes (potentially with ray tracing thrown in) at high resolutions. That's a pretty focused target. So I'm not really sure why an upscaler like FSR is somehow more "general purpose" than DLSS

I think if I understood correctly FSR makes generic assumptions, while DLSS has to be trained on specific input (unless I misunderstood the idea). Generic assumption sounds like a broader approach to me. Trained neural network can produce better result for what it's trained for, but will be close to useless for other cases. Something that's more generic is probably in between. So both are a trade off I think.

Yeah, but I think in this case, DLSS is trained on the textures of the game itself, so in terms of what it does, I don't think it's necessarily any less useful than FSR? I might be wrong.

Just being Nvidia-only is good enough for me to fully get behind FSR. And the AMD showcase video for it was pretty impressive given how young the technology is (dunno what user "sub" was talking about above, claiming that FSR doesn't look as good - not only is there barely any difference, the whole point of these technologies is that they won't look as good, but you'll get 100%+ FPS out of them at high-res, and if you can only tell the difference in a side-by-side video, then that's clear "good enough").

Yeah, but I think in this case, DLSS is trained on the textures of the game itself, so in terms of what it does, I don't think it's necessarily any less useful than FSR? I might be wrong.

What I mean is, what if you have a game or any kind of use case on which it wasn't trained? Will it still fare better than FSR?


Last edited by Shmerl on 2 June 2021 at 9:47 pm UTC

Yeah, but I think in this case, DLSS is trained on the textures of the game itself, so in terms of what it does, I don't think it's necessarily any less useful than FSR? I might be wrong.

What I mean is, what if you have a game or any kind of use case on which it wasn't trained? Will it still fare better than FSR?

Apparently the current iteration of DLSS (2.0) uses a generic algorithm and doesn't require per-game training anymore. So I guess any game could implement support, but it would still only work on a GPU with Nvidia's tensor cores or whatever they call them.

Apparently the current iteration of DLSS (2.0) uses a generic algorithm and doesn't require per-game training anymore. So I guess any game could implement support, but it would still only work on a GPU with Nvidia's tensor cores or whatever they call them.

Why do they need tensor cores if it's a generic algorithm and not some trained neural network? I thought those are focused on AI application.

In that sense I don't see it being too different from FSR. They'll just keep competing on who nails the better algorithm. However from cross GPU usage perspective DLSS is already DOA like another CUDA, so it's not even a choice and I agree with what @scaine said above about it.


Last edited by Shmerl on 2 June 2021 at 10:41 pm UTC

I think for an anti-feature, it's good enough to end DLSS. Because it works everywhere and will also get better over time. The other side of it - it's general purpose. While AI/ML is more limited to specific use cases. So it is better in some cases and worse in others. Everything is a trade off.

This sounds like it was written by someone who doesn't understand either of these features. Hardware-assisted AI/ML being "limited" to specific use cases while an upscaling tech being "general purpose"? Give me a break...

Honestly, you sound like you don't know how Machine Learning and Artificial Intelligence works. It's typically incredibly focused.

Personally, I think a more honest name for it would be Artificial Instinct. 'Cause like, instinct is this stuff animals do without actually being smart or thinking or figuring it out, where through evolution they became really good at some particular task they need to be good at to survive. Machine Learning AI seems to be this forced evolution thing where you let the algorithms survive that are good at some specific task, until you've evolved a black box that can do that specialized thing really well but has no general reasoning ability. So it's instinct, but for some (marketing) reason we call it "intelligence".