AMD Posts Linux Driver Patches For Video Core Next 5 "VCN 5.0"

[jaxa]

If VCN 5.0 is for RDNA4, is there something in between 4.0 and 5.0 for RDNA 3+ (3.5) i.e. Strix Point, Strix Halo, and Kraken Point?

Edit: Seems like it's VCN 4.0.6 for at least one of those.

[ezst036]

Maybe you should address the broader issue yourself then instead of talking nonsense about not being able to stream because of missing 444 support?


The nvidia encoder has a slight edge on encoding efficiency (as long as you don't use 4:4:4) because AMDs hardware doesn't support B frames (at least with older hardware). But you could stream with AMD just fine.
The best quality would be with software encoding but that might not be practical for streaming.

I am discussing the broader issue. The broader issue is that streamers know full well to stay the heck away from AMD cards because their encoder outright sucks, at least previous versions of it. The internet is littered with such articles, I only showed you two of them and more recent. Avoid AMD for streaming has been well established.

(B frames was what I was grasping for, but the specific feature/features are irrelevant)

[M.Bahr]

I think it is you who missed my point… Customisable FPGAs are significantly more expensive than hardwired chips. Certainly at the moment, if not by definition….

As usual with chips they are becoming cheaper thanks to improved mass production and with quantity. And unlike asics where the engineering costs are tremendous and need way more development time and debugging because hardware bugs cannot be reversed later on, this specific factor is way less potentially harmful to fpgas. Those are like a blanc paper in a more broader sense. The paper's initial content can be changed, corrected and improved even after months. And this can save a company lots of money while just one essential but overlooked product defect in an asic can cause a lot of damage to a brand's reputation.

[OneTimeShot]

And this is where you seem to miss my point. I was not proposing to completely replace asics with fpgas but to mix them. This is what i was talking about the whole time if you want to read again especially my last sentence. Amd's U30 card exactly did this.
.

I think it is you who missed my point… Customisable FPGAs are significantly more expensive than hardwired chips. Certainly at the moment, if not by definition….

[JEBjames]

Michael

Typo

"VCN5" should be "VCN 5.0".

"and signs pointing that way" should be "point"

[Anux]

Look, if you don't want to address the broader issue that is fine.

Maybe you should address the broader issue yourself then instead of talking nonsense about not being able to stream because of missing 444 support?

To my knowledge, the best way to get good streaming quality is to swap out an AMD card for an Nvidia card, and it's all about the NVenc.

The nvidia encoder has a slight edge on encoding efficiency (as long as you don't use 4:4:4) because AMDs hardware doesn't support B frames (at least with older hardware). But you could stream with AMD just fine.
The best quality would be with software encoding but that might not be practical for streaming.

[ezst036]

What does 4:4:4 have to do with streaming?

Look, if you don't want to address the broader issue that is fine. To my knowledge, the best way to get good streaming quality is to swap out an AMD card for an Nvidia card, and it's all about the NVenc.

Full disclosure: I do not own an AMD 7x000 (RDNA3) video card. So maybe AMD already fixed it in the latest and I just haven't gotten it yet.

https://www.xsplit.com/blog/why-is-nvenc-so-important-for-live-streaming

https://www.digitaltrends.com/computing/amd-gpus-are-finally-good-enough-for-streaming/

[M.Bahr]

FPGAs are great for small volume runs, but after you ship ~10,000 of the things, hardwired chips are much cheaper. Also the maximum speed of an FPGA is much lower than hardwired chips.

It would be interesting to have an FPGA in a PC, but I'm not sure if it has a mass-market use case. The video encoder/decoder block is basically a small CPU with a custom instruction set anyway. Only "the expensive bits" of the process are done in tailor-made hardware.

Don't worry. I already considered this. And for the purpose of video decoding fpgas are more than sufficient. Note amd's previous dedicated video decoding cards benefited greatly from fpgas and it worked great. The current lack of higher frequencies can be compensated very well by parallel scaling. Especially fpgas are very good for this. Also nothing prevents a vendor from mixing asics with fpgas. And this is where you seem to miss my point. I was not proposing to completely replace asics with fpgas but to mix them. This is what i was talking about the whole time if you want to read again especially my last sentence. Amd's U30 card exactly did this.

However my main point is. Hardware is getting more and more complex. And this automatically increases failure rates in hardware which are irreversible in asics. They have to be mitigated by software and this costs performance similar to spectre workarounds in intel cpus. We also see this degradation of quality very well in modern gpus like for example with the 1082p output bug in rdna3, instead of delivering 1080p lines. https://github.com/GPUOpen-Libraries...AMF/issues/423
If we had more fpgas in gpus we could completely and cleanly resolve those kind of issues at the hardware level and without performance degradations.

[Serafean]

maybe they will add h.266/vvc decoding support.

Hopefully they don't. Leave h.266 to die under a mountain of patent trolls.
Unfortunately DVB-T2 has already been extended to support it.

Companies known not to be a part of the Access Advance or Via-LA patent pools as of November 2023 are: Apple, Canon, Ericsson, Fraunhofer, Google, Huawei, Humax, Intel, LG, Interdigital, Maxell, Microsoft, Oppo, Qualcomm, Samsung, Sharp and Sony.

Big names not interested in VVC. I'm still kind of hopeful.

[hajj_3]

maybe they will add h.266/vvc decoding support.