Nvidia 4000 series anounced - w/Da Vinci Resolve info

https://blogs.nvidia.com/blog/2022/09/2 ... 40-series/

studio-itns-racerx-ada-wk23-encoders-1280w.png (374.43 KiB) Viewed 4374 times

As a side note the AMD 7000 series will be presented at begin of November.

Hopefully the new Optical Flow Accelerator can be used by Resolve as well. My GPUs have done well for seven years but I think this generation is the one where I upgrade.

Hopefully the new Optical Flow Accelerator can be used by Resolve as well. My GPUs have done well for seven years but I think this generation is the one where I upgrade.

I think Resolve already has Optical Flow. There is a Optical Flow SDK from nvidia and its not just for ADA. It started from Turing and its the stuff they use for DLSS.

The Accelerator sounds like buzzwords to me, meaning they will use motion interpolation for DLSS3 games starting from DLSS3 where before they didn't.

Doubt there will be any difference in Resolve to anything that won't be explained by the increase of CUDA cores on some models. (4080 12GB looks very bad in this aspect - looks more like an extremely overpriced 70 card). 70% more CUDA cores for 70% more performance.

Hope I'm wrong...

4080 12GB looks very bad in this aspect - looks more like an extremely overpriced 70 card

It's actually not the same GPU as the real 4080, so you are correct.

Any info on codec supported for hardware acceleration?

Only Intel 11/12th gen and Mac M1/2 have support for H265 4:2:2 today

https://www.pugetsystems.com/labs/artic ... udio-2122/


M1 kicks butt in many productivity benchmarks simply from its amazing hardware acceleration. Nvidia could eat some of their lunch if they catch up.

Probably unreasonable to expect ProRes support like M1, I doubt Mac would sell their competition a license. But I want H265 422. No doubt Google would be happy to see VP9 added like Intel has.

At some point are we going to need hardware acceleration for Dnxhr? Mac thought ProRes needed it.

At some point are we going to need hardware acceleration for Dnxhr? Mac thought ProRes needed it.

I never understood the obsession with ProRes Apple people have.(since the m1). I could edit heavy prores files on an FX8350 (still do when all our other machines are busy rendering other stuff). And that CPU is extremely old by today's standards. Its an extremely light codec for CPUs ...don't really need an asic for it. Same for dnxhr.

150fps vs 900 fps for final rendering? Who really cares...

H265 is another story...there I get it. AV1 too.

My 2 cents.

Hopefully the new Optical Flow Accelerator can be used by Resolve as well. My GPUs have done well for seven years but I think this generation is the one where I upgrade.

Highly unlikely I think. The tech is meant to be used in conjunction with DLSS3.0 which uses motion vectors from the game engine + AI to predict the between frames of the games rendered frames and also needs nVidia Reflex to be implemented.

I doubt they'll make it's compute power available for different tasks. But who knows... maybe it could trickle into 3D renderers at some point which could also use motion data to drive between frames. If it's stable enough you could potentially render twice as many frames. Maybe not perfect for final render but for iterations that would be quite nice.

I think Resolve already has Optical Flow.

Yes it does, with or without tensor cores. But the question is wether this new functionality is "plug and play" or if they have to actively use another API and if it's exposed for "dumb" frames at all.

Highly unlikely I think. The tech is meant to be used in conjunction with DLSS3.0 which uses motion vectors from the game engine + AI to predict the between frames of the games rendered frames and also needs nVidia Reflex to be implemented.

The way I understood it is they don't use vector data for the Optical Flow Accelerator. Just frames. They say this:

The Optical Flow Accelerator is able to capture pixel-level information such as particles, reflections, shadows, and lighting, which are not included in game engine motion vector calculations.

But we'll find out eventually.

There's a quote from BMD in this nVidia article.

“The new GeForce RTX 40 Series GPUs are going to supercharge the speed at which our users are able to produce video through the power of AI and dual encoding — completing their work in a fraction of the time,” said Rohit Gupta, director of software development at Blackmagic Design.

AV1 and VP9 acceleration for decode is already in the current gen, but now it's getting AV1 encoding.

The Video Encode and Decode GPU Support Matrix hasn't been updated for ADA yet. But this SDK page has ADA listed with some decode capabilities (expand the "Supported Format Details" header), nothing mentioned about 4:2:2 yet unfortunately. Hopefully it's just an incomplete table...

No doubt Google would be happy to see VP9 added like Intel has.

I thought the latest chips dropped VP9 and replaced it with AV1.

Do you think the RTX4000 series will be supported by DaVinci 17.X ? I mean, they might make a last small update for this.

Do you think the RTX4000 series will be supported by DaVinci 17.X ? I mean, they might make a last small update for this.

I haven't heard anyone claim V16 (apr-2019) does not work with 30 series (Sept-2020).

I expect 40 series will still support all the API's V17 is using today.

Do you think the RTX4000 series will be supported by DaVinci 17.X ? I mean, they might make a last small update for this.

I haven't heard anyone claim V16 (apr-2019) does not work with 30 series (Sept-2020).

I expect 40 series will still support all the API's V17 is using today.

There's new hardware encoding export such as AV1 codec. I suppose it requires an update anyway.

Well this is disappointing. The Video Encode and Decode GPU Support Matrix is updated now. Only the encoder is new, the decoder is the same generation as on the 3000/Ampere series.

Likely no 4:2:2 support at all, unless they didn't feel it was necessary to show it.

Well this is disappointing. The Video Encode and Decode GPU Support Matrix is updated now. Only the encoder is new, the decoder is the same generation as on the 3000/Ampere series.

Likely no 4:2:2 support at all, unless they didn't feel it was necessary to show it.

Thanks, yes it seems not 4:2:2 unless they forgot, but since that seems an high level of detail Nvidia page i think the odds is that 4000 series do not come with HEVC 4:2:2.
Do you think that 4:2:2 have future?

All the cameras with good H.264/H.265 are 4:2:2, so this is a bit of a let-down. For Windows users that want decode/encode acceleration that means Intel is still the only choice with their recent CPUs.

All the cameras with good H.264/H.265 are 4:2:2, so this is a bit of a let-down. For Windows users that want decode/encode acceleration that means Intel is still the only choice with their recent CPUs.

Or a nice Threadripper and you can convert to ProRes. H265 is horrible for editing.

Well, with a Threadripper you rather convert to DNxHR or Cineform.

Well that's the point, Threadrippers don't have hardware accelerated decoding of 4:2:2 material either.

That's whats nice about the Intel CPUs and Apple Silicon. No need for workarounds.

NVLink is confirmed to not be in 4xxx series.

Nvidia confirms there is no HEVC 4:2:2 decode in 4xxx series, it still uses 3xxx decoder.
That appear specially dumb, meaning giving resources to their new GPU competitor Intel, since it forces anyone in PC to buy an Intel CPU, since AMD also seems to be sleeping.

https://www.nvidia.com/en-us/geforce/ne ... munity-qa/

NVIDIA Broadcast, NVIDIA Studio & NVENC
Q: What would be the benefits of using Nvidia NVENC over the more traditional x264?

Link

x264 is a software encoder that works on CPU, whereas NVENC is a hardware encoder that uses dedicated hardware on NVIDIA GPUs. x264 will utilize part of your CPU, leaving less power to run your games or other apps. NVENC operates on an independent part of the GPU, leaving the CPU and GPU to render the game and apps. Hence, NVENC allows you to maximize the use of your hardware and get more FPS.

On top of that, next gen codecs like AV1 consume a lot of resources and cannot run on a typical CPU. But with NVENC in GeForce RTX 40 Series you can encode AV1 up to 8K60 seamlessly.

Q: Any updates to NVDEC? Still reliant on either M1 or Quicksync for 10 bit 422 decode.

Link

GeForce RTX 40 Series use the same NVIDIA Decoder as RTX 30 Series - 5th generation NVDEC. There is no support for 10-bit 4:2:2 decode.

Q: Will ShadowPlay benefit from the NVENC improvements and the addition of AV1 encoding?

Link

ShadowPlay has been updated to make use of GeForce RTX 40 Series dual encoders, enabling ShadowPlay to record up to 8K60 HDR in HEVC. AV1 is not currently supported in ShadowPlay.

Well that's the point, Threadrippers don't have hardware accelerated decoding of 4:2:2 material either.

That's whats nice about the Intel CPUs and Apple Silicon. No need for workarounds.

Threadrippers don't need asics for decoding. That's my point. They can brute force it. There's no point on gimping yourself on CPU for every other tasks just to decode 4:2:2 H265. But that's just me...I also find Apple Silicon hilariously overpriced for what they offer...

Nvidia confirms there is no HEVC 4:2:2 decode in 4xxx series, it still uses 3xxx decoder.
That appear specially dumb, meaning giving resources to their new GPU competitor Intel, since it forces anyone in PC to buy an Intel CPU, since AMD also seems to be sleeping.

Sad for sure. :'(


Agreed, unless Rdna3 really picks up the ball on decode (they aren't even in the stadium today), Intel 12th gen is the only choice.

Or threadripper pro if you need more PCIe

Nvidia confirms there is no HEVC 4:2:2 decode in 4xxx series, it still uses 3xxx decoder.
That appear specially dumb, meaning giving resources to their new GPU competitor Intel, since it forces anyone in PC to buy an Intel CPU, since AMD also seems to be sleeping.

Sad for sure. :'(


Agreed, unless Rdna3 really picks up the ball on decode (they aren't even in the stadium today), Intel 12th gen is the only choice.

Or threadripper pro if you need more PCIe

The interest for decoding 4:2:2 on the consumer side is pretty low hence why neither AMD nor Nvidia will add this. I think Intel adds it because Intel cannot compete in anything else with their iGPUs.

Maybe if 4:2:2 was more wide spread on stuff like Netflix and such there would be more cards decoding it.

Nvidia confirms there is no HEVC 4:2:2 decode in 4xxx series, it still uses 3xxx decoder.
That appear specially dumb, meaning giving resources to their new GPU competitor Intel, since it forces anyone in PC to buy an Intel CPU, since AMD also seems to be sleeping.

Sad for sure. :'(


Agreed, unless Rdna3 really picks up the ball on decode (they aren't even in the stadium today), Intel 12th gen is the only choice.

Or threadripper pro if you need more PCIe

The interest for decoding 4:2:2 on the consumer side is pretty low hence why neither AMD nor Nvidia will add this. I think Intel adds it because Intel cannot compete in anything else with their iGPUs.

Maybe if 4:2:2 was more wide spread on stuff like Netflix and such there would be more cards decoding it.

Nvidia promotes their "gaming" GPU's for Resolve, Premiere, 3D rendering in Blender , Redshift, Vray. They have even their "Studio" certification for laptops and desktops for artists. There are 4K cameras now in drones, GoPros and competitors and everyone can be a content creator. I think it is a serious overlook.

Nvidia promotes their "gaming" GPU's for Resolve, Premiere, 3D rendering in Blender , Redshift, Vray. They have even their "Studio" certification for laptops and desktops for artists. There are 4K cameras now in drones, GoPros and competitors and everyone can be a content creator. I think it is a serious overlook.

I still think there are too few devices using 4:2:2 H265 in the grand scheme of things for nvidia and amd.
Maybe when they'll run out of gimmicks like DLSS3 they'll focus on stuff that actually matters.Wishful thinking.

They didn't even update display port to 2.0 on the new 4000 cards...That's how much they care about color . 4:2:0 forever baby!

If Intel got their ARC together they might interest a lot of creators who want better codec support, especially if it cooperated with other (NVIDIA/AMD) cards in the same machine.

The winning combo might be a threadripper (lots of CPU and PCIE lanes) plus an RTX4090 for the heavy GPU lifting and an ARC card for enhanced hardware-based codec support.

If Intel got their ARC together they might interest a lot of creators who want better codec support, especially if it cooperated with other (NVIDIA/AMD) cards in the same machine.

The winning combo might be a threadripper (lots of CPU and PCIE lanes) plus an RTX4090 for the heavy GPU lifting and an ARC card for enhanced hardware-based codec support.

There are rumors Arc GPUS will be beasts with Ray Tracing. (aka real ray tracing in software like Blender and Houdini). The state Arc GPUs are in right now though...

The winning combo moving forward I think would be Epyc + Whatever flagship GPU. Threadrippers have become really expensive. Lack of competition from Intel in the HEDT market...

But that's just me...I also find Apple Silicon hilariously overpriced for what they offer...

Well, maybe in the US. Not so much here, it may be levelled out to some degree by your next electricity bill vs a Threadripper and a fast Nvidia GPU. OTOH, with those you won’t need an electric heater when the gas is running low…

But that's just me...I also find Apple Silicon hilariously overpriced for what they offer...

Well, maybe in the US. Not so much here, it may be levelled out to some degree by your next electricity bill vs a Threadripper and a fast Nvidia GPU. OTOH, with those you won’t need an electric heater when the gas is running low…

That the big draw with Apple Silicon which I have been entertaining. However GPU prices are dropping to pre-pandemic prices which makes it more of a decision issue. Do I sprint for a $4K USD M1 Ultra or upgrade my existing workstation with a $980 USD RTX 3090 GPU or a AMD RX 6950X for $690 (recent prices on NewEgg) and just deal with the electric bill?

But that's just me...I also find Apple Silicon hilariously overpriced for what they offer...

Well, maybe in the US. Not so much here, it may be levelled out to some degree by your next electricity bill vs a Threadripper and a fast Nvidia GPU. OTOH, with those you won’t need an electric heater when the gas is running low…

Is the Apple as fast as the Nvidia? example 10 seconds at 200w is the same energy as 200sec at 10 w. At least in Blender Cycles a Nvidia is much more efficient than an Apple M1.

I wasn't referring to Blender (or C4D for that matter), only DR. When you are modelling with those, Gpu load is pretty low most of the time. It goes up on final render. DR is generating quite some GPU load when you are just editing and grading.
We have one room full of PCs and another with Macs. If a whole class is doing DR in either room (still Intel Macs) they can turn down the heating. Not only when rendering.

And then, I'd never buy a top-level model from Apple. They are not offering a good price/performance level, they are for showing off. One level below those is usually pretty good. But I think that energy efficiency is a very important consideration these days, prices will only rise. At least until someone is really building a workable fusion reactor (and solving all the problems with nuclear waste).

Apple's ARM cpus are ok in some tasks that are single threaded. And very efficient in those tasks.

However when you need real computing power they fall flat. They aren't a match for Threadripper / Ryzen 5000 or Intel 12gen in any way or match amd and nvidia gpus.

Based on transistor count alone and the bleeding edge node they use you can see how hilariously bad they are vs their x86/gpu counterparts. And x86 isn't that bad at power consumption when doing editing and such while being much much better at multithreading. How fast does the M1 choke when you have multiple software open at the same time? Like the famous Adobe suites for example, Resolve, 40 chrome tabs open. And since these function like an APU regarding memory they constantly swap to nvme which is soldered, how long till that dies (being used in a professional matter daily)?

And the problem with ARM is that they can't really scale up because then you lose that power efficiency and end up with something much worse than x86.

My 2 cents.

There must be a gpu pluggable on laptop, so we can change GPU without buy NEW LAPTOP.

There must be a gpu pluggable on laptop, so we can change GPU without buy NEW LAPTOP.

There are external GPUs. Blackmagic sells some I think. They plug in via Thunderbolt / USB.

There are, but they are outdated. And the better ones are slower than they are inside a proper desktop machine. IMHO, eGPUs are an endangered species. BTW, mobile GPUs insides laptops are often crippled compared to their desktop counterparts.
If everything would be possible with a laptop, who'd still want a desktop?

...I never understood the obsession with ProRes Apple people have.(since the m1). I could edit heavy prores files on an FX8350 (still do when all our other machines are busy rendering other stuff). And that CPU is extremely old by today's standards. Its an extremely light codec for CPUs ...don't really need an asic for it. Same for dnxhr.

150fps vs 900 fps for final rendering? Who really cares...

I formerly thought the same thing about the "Afterburner" card in the 2019 Mac Pro. That accelerated decoding ProRes and ProRes RAW. I thought ProRes IS optimized media, why have a special card (or a year later the M1) to accelerate that? Possible answers:

- Even though 4k ProRes 422, etc. is relatively lightweight, decoding ProRes RAW has more compute burden, which esp. manifests when handling 8k multicam. The Afterburner card and subsequent Apple Silicon ProRes accelerators were partially for ProRes RAW, not just ProRes.

- Even for regular ProRes, it's not always about the final export. For large doc or scripted productions, you are dealing with multiple multicam teams, offloading that and transcoding to some mezzanine codec. An M1 Ultra Mac Studio using Apple Compressor can transcode 4k ProRes to 1080p ProRes Proxy at about 800 frames/sec. However that task is parallelizable, so you could split the files across several slower Windows or Apple machines and achieve the same result.

Re nVidia 4000's dual accelerators doubling the rate of ProRes to H265 exports, M1 Max has two H264/H265 encoders and the M1 Ultra has four, yet no NLE I've tested can use those in parallel on a single stream. IOW the current single-stream scalability is zero. That raises the question of how did nVidia get the results shown in the original post, whereby the RTX 4090 encoded to H265 at 2x the rate of the RTX 3080 Ti.

If the H265 Long GOP format they used involved "open" aka independent GOPs, in theory they could split the input file, dispatch it to different parallel threads or processes, then concatenate the result. However if the H264 source media used "closed" aka dependent GOPs, I don't think that's possible because one GOP needs info from other GOPs.

Or they could just use multiple H265 input files and run those in parallel on different processes. Given two H265 encoders that would be about 2x faster.

Apple's Compressor already does that if you enable the advanced preference "Enable additional Compressor Instances". On M1 Ultra it can transcode 4k ProRes 422 source to 720p HEVC at about 1,000 frames/sec, aggregated across four input files and four output files. The scalability is not perfect but useful: 1 instance yields 418 frames/sec, 2 instances 790 frames/sec and 4 instances 960 frames/sec. It can do that for 8-bit 4:2:0 HEVC or 10-bit 4:2:2 HEVC at the same rate.

That is only possible because the M1 Ultra has four separate H264/HEVC accelerators, in addition to the four ProRes accelerators.

It would be nice to have effective decode & encode acceleration of single stream 10-bit 4:2:2 H264/H265. Even if the software could harness that, it would be limited to the M1 Max (which has one decoder but two encoders) and M1 Ultra (which has two decoders and four encoders), plus other non-Apple CPU/GPUs that may add multiple accelerators to their new designs.

Currently on the M1 Ultra the multiple accelerators help for H264/265 multicam (at least on FCP, I tested it on Resolve Studio and I think it also helped). But they don't help for smoother editing or faster export on single-stream H264/H265. It is already pretty quick but if multiple accelerators could be leveraged on single-stream Long GOP formats, that would be better.

I formerly thought the same thing about the "Afterburner" card in the 2019 Mac Pro. That accelerated decoding ProRes and ProRes RAW. I thought ProRes IS optimized media, why have a special card (or a year later the M1) to accelerate that? Possible answers:

- Even though 4k ProRes 422, etc. is relatively lightweight, decoding ProRes RAW has more compute burden, which esp. manifests when handling 8k multicam. The Afterburner card and subsequent Apple Silicon ProRes accelerators were partially for ProRes RAW, not just ProRes.

- Even for regular ProRes, it's not always about the final export. For large doc or scripted productions, you are dealing with multiple multicam teams, offloading that and transcoding to some mezzanine codec.

ProRes RAW is a different story. I agree its good for that.
But for regular ProRes...Not so much. Making proxies or final render, same stuff in the end.

Might not be as fast as an asic, but on a Threadripper and a good card, making proxies is blazing fast.
I never thought "gee, 300fps is slow...need 800"... Maybe 40TB of footage (max i've ever dealt with)
is not enough to make me want 800? I'm getting bottlenecked by hdd/ssd speed anyway.

My point is ...I think they should focus on something else...They keep adding asics that frankly no one asked for (whatever the first iteration was is good enough imho) and keep charging you for it. Focus on making a working GPU (transistor count is much higher than a 3090 while it works worse than a 3060...). I'm exaggerating but ...not that much.

It seems also upgrading to the 4000 series cards will also require upgrading to a gen 3, data controlled power supply. In which the graphics card and power supply communicate with each other directly to manage available current draw across power supply rails.
It seems that older power supplies with out the data connect will either result in the 4000 series card throttling. Or even risk over current draws, connector and cable meltdowns and fire.

Personally, if I was desperate for more compute performance, I would buy a second studio max (or m2 mini max, not the ultra), and begin developing a render farm. For the same price as a 4000 series plus gold rated 1200w power supply. Or equipping myself with the maxed out Mac studio Max, plus a 16” MacBook Pro. Both with a ridiculously overpriced SSD upgrade on board and an external thunderbolt4/USB4 NVMe so that renders were not bottlenecked by SSD speeds.
When I see Apple computers catching fire on the desktop, I might reconsider. After all a 4090 equiped pc, would not only be as loud as the deck of an aircraft carrier, it might genuinely be almost as dangerous to sit next to.
The better Windows PC GPU upgrade alternative, may actually be AMD, if Resolve compatibility was delivered.


Sent from my iPad using Tapatalk

It seems also upgrading to the 4000 series cards will also require upgrading to a gen 3, data controlled power supply. In which the graphics card and power supply communicate with each other directly to manage available current draw across power supply rails.
It seems that older power supplies with out the data connect will either result in the 4000 series card throttling. Or even risk over current draws, connector and cable meltdowns and fire.

Personally, if I was desperate for more compute performance, I would buy a second studio max (or m2 mini max, not the ultra), and begin developing a render farm. For the same price as a 4000 series plus gold rated 1200w power supply. Or equipping myself with the maxed out Mac studio Max, plus a 16” MacBook Pro. Both with a ridiculously overpriced SSD upgrade on board and an external thunderbolt4/USB4 NVMe so that renders were not bottlenecked by SSD speeds.
When I see Apple computers catching fire on the desktop, I might reconsider. After all a 4090 equiped pc, would not only be as loud as the deck of an aircraft carrier, it might genuinely be almost as dangerous to sit next to.
The better Windows PC GPU upgrade alternative, may actually be AMD, if Resolve compatibility was delivered.


Sent from my iPad using Tapatalk

A 4090 reportedly consuming 460w is not more than 55ºC.

When I see Apple computers catching fire on the desktop, I might reconsider.

Sent from my iPad using Tapatalk

I would be more afraid near one of those trashcan macs than on anything GPU related.

The new PSUs handle transient spikes better. This doesn't mean old ones can't. Just means old no name PSUs should not be used for powerful GPUs since the PSUs can be a fire hazard. But those were a fire hazard for anything anyway so it really doesn't change much.

As for more compute power...the GPUs in the M1/M2 are weak ...not sure 2 of them can even come close to a 3090.(if both acted like one unit).

DLSS3 Video frame creation. At 23min in video a comparison with Topaz for After Effects.

Some more information how DLSS3 frame creation works:, this would be good if available for Resolve

https://www.nvidia.com/en-us/geforce/ne ... novations/

First comparison results in Blender GPU render benchmark

image-11.jpg (49.32 KiB) Viewed 2552 times

https://opendata.blender.org

Small preview of benchmark:

https://benchlife.info/nvidia-geforce-r ... x-3090-ti/
https://www.hardwareluxx.de/index.php/a ... l?start=29
https://news.mynavi.jp/article/20221011-2478652/

Waiting for more...

I think we must wait for the new updated version of DaVinci to see the real performance gain of the dual encoders.

From Nvidia blog:

Blackmagic Design’s DaVinci Resolve, the popular Voukoder plugin for Adobe Premiere Pro, and Jianying — the top video editing app in China — are all enabling AV1 support, as well as a dual encoder through encode presets. Expect dual encoder and AV1 availability for these apps in October.

I hope it's a matter of update because it sucks at the moment

Not sure to place an order tomorrow...

Different hardware configuration to explain such differences? CPU/RAM/SSD limitation...