We’ve had several community members asking about how the shiny new NVIDIA RTX 6000 Ada Generation Graphics card performs with Depthkit.
We’ve tested this new card at Scatter and have some initial results!
This card boasts a new level of video encoding capabilities as it’s the only RTX/Quadro card with 3 unrestricted NVENC chips (see the NVIDIA Support Matrix). This encoding chipset is the hardware that Depthkit Studio relies on for fast and efficient video encoding of multiple streams at high resolution, without dropping frames. As a reminder, you can see relevant part of the support matrix in our hardware guide.
The Good News: The NVIDIA RTX 6000 Ada Generation card unlocks 10x1440p capture with Depthkit Studio, which is a step up from the 10x1080p capture currently available in our prebuilt hardware packages
The Bad News Based on some isolated system benchmarking, our team believes that up to 10x2160p should be possible with this card, but we are seeing CPU bottlenecks when running the full end to end practical workload that limit this resolution in the current Depthkit Studio system. Briefly, the limitation is in the stage where Depthkit receives color frames from the Azure Kinect SDK.
Fixing these bottlenecks is currently a lower product priority than a few other key projects we are working on, so there’s no timeline on when we may unlock the full performance of the card. We’ll keep this thread updated as we learn more!
Thank you James for this initial feedback! So, a RTX 4090 with 2 NVENIC chips would be much better when using only 5 Kinect DKs (e.g, for filming the upper body only or because of just owning 5 of them…) instead of a a5000 with only 1 NVENIC?
Thank you for your feedback in advance!
Best,
Wolfgang
Hi, @WolfgangSchafer - I believe your assumption is correct that the RTX 4090 will perform better than the Ampere-generation A5000 for up to 5-sensor capture, but I’ll check with our technical team and report back.
The RTX 4090 has 1 NVENC chip containing 2 encoders, and a limit of 5 simultaneous encode sessions.
The Ampere-generation RTX A5000 has 1 NVENC chip containing 1 encoder and unrestricted encode sessions.
The Ada-generation has reintroduced multiple encoders per chip back into their professional line of GPU’s, and I believe the NVENC chip with 2 encoders in the RTX 4090 is the same one found in the Ada-generation RTX 4000 ADA & RTX 5000 ADA, which both have unrestricted encoding sessions. The RTX 6000 ADA has an even more capable NVENC chip containing 3 encoders.
@WolfgangSchafer - A quick update regarding capture performance: We have just completed some initial testing with the RTX 4000 ADA Generation GPU (1x NVENC chip containing 2x encoders), and when paired with an Intel Core i9-14900K CPU, it is able to capture 10 sensors each set to 1440p color, making that GPU a cost-effective high performance option.
Will it be possible to mix azure kinect dk with orbec Femto bolt ? … for example using the bolts for front or a bit closer to face as kind of frontal direction and the rest is handled by azure ??? Would help to not need to buy at the beginning so many bolts …
Interesting question, it’s something we discussed in the development of the Femto Bolt support.
Short answer is no, you’ll not be able to mix and match hardware. The hard constraint is that the Femto Bolt and Azure Kinect have completely different sync systems, so there will not be a supported way to bridge the sensors into the same array.
Beyond that, the color optics are so different that the images would not blend well even if the sync did work. Also, the depth quality is not different so there is really no benefit to using them together to make it worth the complexity.
so then I have to ask you something else … did your R&D Team test as well the Femto mega with the in build gpu into the sensor and the RJ45 connection.
As this could be interesting to staying with A4000 (not Ada) and using a high class depthKit azure System the Femto mega solution for a higher Resolution result and avoid buying a new Ada gpu and a new cpu system
And an additional question in the past you developed a sync system with the external sync deliverer with the ambient “clock it” for use cases like having 10 paired rgb cinema cameras with the azures … if the bolt system goes a complete own road how this pairing with external cinema camera will work out in the future?
I am aware that the bolt and the azure have different connectors (mini jack and the orbec Sync hub system) but the sync hub is some kind very basic system and normally the company ambient and the company tentacle are the global leaders in syncing … ambient teamed years ago up with Panavision to be even build in the latest high class Panavision cinema cameras … so I am super (negative) surprised that it looks like that they are not possible to sync …
Ok the difference in color I believe right away as each cinema camera has its own color management, but off course for that are normally LUT workflows with ACES and datacolor checker developed.
It’s a pity that these kind of standards haven’t arrived in the volumetric world.
@MartnD The next version of Depthkit will have experimental support for the Femto Mega for Depthkit Core (single-sensor) only. This is primarily because our highest priority for this release is support for the Femto Bolt as it is the most comparable in price, features, and performance to the now-discontinued Azure Kinect. Once that transition is complete, we will evaluate the Femto Mega for full Depthkit Studio support.
Today, it’s possible to get the additional performance 10x1440p capture by just upgrading the GPU to an Ada-generation card like the RTX 4000 ADA or higher. This is more cost effective than the additional cost incurred by purchasing 6 or more Femto Mega sensors over the equivalent number of Femto Bolt sensors.
Regarding sync and color workflows for different sensors, please start separate threads on those topics to discuss further.
Thx @CoryAllen!
The argument with price difference of 8-10 femto mega in compare to femto bolt plus one A4000 Ada is absolutely reasonable.
I will next days move my sync issue to different threat … at the moment we have a sick child at home (tiny one) and this takes over full attention … second time … I wanted to try the laustest release in December but there as well sick child and Christmas … and now I switched another time from german home to polish home and child is sick again so I guess I will find in February or March time to finally test your latest release …
Thx for keeping me in the loop I read from time to time in the evenings the forum as way of keeping me up to date … so thx for coming back to me …
We’ve tested this new card at Scatter and have some initial results!
…
… and now I switched another time from german home to polish home and child is sick again so I guess I will find in February or March time to finally test your latest release …