NVIDIA discussion [2024]

[Newguy]

Nvidia have increased their driver level limit for encoding video simultaneously to 8 (used to be 2, then 3 in 2020, then 5 in march last year) on all consumer GPUs that support Nvenc back to Maxwell

NVENC Update: All Nvidia GeForce Cards Quietly Updated to 8 Encoding Sessions - Stream Guides

Nvidia released Game Ready Driver 551.23 on January 24 to support the new RTX 4070ti Super graphics card and add new features like the RTX Video HDR enhancement. What wasn't reported in the changelog for the driver update was an increase to the number of possible NVENC encoder sessions, bringing...

streamguides.gg

Video Encode and Decode GPU Support Matrix

Find the related video encoding and decoding support for all NVIDIA GPU products.

developer.nvidia.com

Should be fine for 1080p but this doesn't mean it'll work well if you try to crank the resolution, bit depth and/or fps per encode especially on older GPUs, but it's possible. Nvidia's FPS results in a table here for 1080p, 4:2:0 8 bit for reference (table 3)

NVENC Application Note

Archive version of Video Codec SDK

docs.nvidia.com

 

[pharma]

Nvidia simplified the board design of its RTX 4080 Super Founders Edition

Lower MSRP requires lower costs.

www.tomshardware.com

Nvidia's GeForce RTX 4080 Founders Edition graphics card has a voltage regulating module featuring 13 power phases for its AD103 graphics processing unit, with three power phases for its 16GB of GDDR6X memory. By contrast, the new GeForce RTX 4080 Super Founders Edition board has an 11-phase VRM for the GPU and a 2-phase power supply circuitry for the memory, as noticed by reviewer Geekerwan (via @9550pro).

The slightly simplified printed circuit board (PCB) design does not really come as a surprise. On the one hand, by now Nvidia and TSMC have probably reduced performance variability of the AD103 graphics processor to a minimum and so Nvidia doesn't need to ensure the best possible power supply for the GPU — at least with 10,240 CUDA cores operating at up to 2,550 MHz — as virtually all chips will perform as advertised even with an 11+2 phase VRM.

Meanwhile, simplifying the GeForce RTX 4080 Super Founders Edition makes the product slightly cheaper to produce, so Nvidia can sell one of the best graphics cards on the market for $999 instead of $1,199 and retain its profit margin. The slight design optimization should not tangibly affect reliability of the product.

 

[DavidGraham]

https://twitter.com/i/web/status/1753027683590811918

 

[pjbliverpool]

No discussion for the new Super series or have I just missed it?



4070S looks great. 4070Ti Super is a nice little boost with a welcome VRAM upgrade, and the 4080S barely moves the needle but brings a very welcome and significant price drop.

I'm hoping the drop to $999 means the 5080 will also launch at that price. Cheaper would of course be far more welcome but I'm not that naive.

 

[arandomguy]

No discussion for the new Super series or have I just missed it?

It was in a thread in the Architecture and Products subforum -

Nvidia GeForce RTX 40x0 rumors and speculation

Bios was not optimized ... BIOS Update for Enhanced Performance MSI GeForce RTX 4070 Ti SUPER Series Graphics Cards Hmm I wonder what they changed.

forum.beyond3d.com

4070S looks great. 4070Ti Super is a nice little boost with a welcome VRAM upgrade, and the 4080S barely moves the needle but brings a very welcome and significant price drop.

I'm hoping the drop to $999 means the 5080 will also launch at that price. Cheaper would of course be far more welcome but I'm not that naive.

In terms of actual discussion I wonder if going forward we shouldn't necessarily expect perf/$ increases with new architectures, given the lack of transistor cost scaling at the onset, higher fixed costs, and consumer/market dynamics.

Newer generations will primarily push new levels of performance, features/capability, and efficiency to target early adopters. We'll go back to refreshes being the norm which will have a greater focus on pushing perf/$.

 

[RedVi]

I'm hoping the drop to $999 means the 5080 will also launch at that price. Cheaper would of course be far more welcome but I'm not that naive.

If 5080 ends up being more like the 3080, ie larger chip with a wider bus and 20GB VRAM I think $999 would be a fair price in a years time or less. If it doesn't move the needle on those metrics, then it's personally unappealing to me unless priced a fair bit lower.

 

[DegustatoR]

What if it will be a larger chip with a wider bus and 20GB VRAM at $999 with the same performance as 4080?

Before you'll start saying how impossible that is I urge you to study Turing.

 

[techuse]

What if it will be a larger chip with a wider bus and 20GB VRAM at $999 with the same performance as 4080?

Before you'll start saying how impossible that is I urge you to study Turing.

What new rendering feature would be a big enough sell?

 

[Malo]

What new rendering feature would be a big enough sell?

Well Nvidia tried to use frame generation as the basis for the 4000 series pricing perf/$, with the marketing from Nvidia and folks here pushing it as such. I'm sure they'll come up with something new again.

RTX 5000 series, now with AI dynamic texture creation! Games no longer need to ship textures, their AI knows exactly what is needed based on a single paragraph of text describing the game! No need for those pesky human artists, make so much more profit and speed up time-to-market!

 

[pharma]

What new rendering feature would be a big enough sell?

Would it really be required seeing the breathtaking pace of novel and innovative gaming features marketed by competitors over the past few years?
It basically been a coattail ride with no end in sight ...

 

[techuse]

Would it really be required seeing the breathtaking pace of novel and innovative gaming features marketed by competitors over the past few years?
It basically been a coattail ride with no end in sight ...

Well if they were to release a new chip at the same price that’s larger with a bigger bus and offers no performance improvement, I would assume there would be some groundbreaking new rendering feature.

 

[pharma]

and the 4080S barely moves the needle but brings a very welcome and significant price drop.

I'm hoping the drop to $999 means the 5080 will also launch at that price. Cheaper would of course be far more welcome but I'm not that naive.

Agreed, while the 4080S performance barely registers against it's predessor I think this was done as a strategic move rather performance upgrade.
If Nvidia can keep a similar 5080 price/feature ratio compared to the 4080S it should satiate many who will upgrade.

 

[trinibwoy]

What new rendering feature would be a big enough sell?

Most of Nvidia’s proprietary stuff is about performance hacks and not necessarily rendering features. Reflex, DLSS, frame gen. Rendering features like OMM and DMM don’t have api support. I think OMM made an appearance in CP2077 but that’s it. “New” features like VRS and sampler feedback have api support but are a bust.

Not sure why we would want Nvidia to push the envelope on rendering tech. It’s just going to invite the same hate hurled at them for doing so in the past. I think the better question is what new features can we expect from the main graphics apis.

 

[techuse]

Most of Nvidia’s proprietary stuff is about performance hacks and not necessarily rendering features. Reflex, DLSS, frame gen. Rendering features like OMM and DMM don’t have api support. I think OMM made an appearance in CP2077 but that’s it. “New” features like VRS and sampler feedback have api support but are a bust.

Not sure why we would want Nvidia to push the envelope on rendering tech. It’s just going to invite the same hate hurled at them for doing so in the past. I think the better question is what new features can we expect from the main graphics apis.

It was in response to @DegustatoR suggesting the possibility of a 5080 offering no improvement in performance despite being a bigger chip with a wider bus. That would necessitate some drastic new paradigm like with Turing.

 

[trinibwoy]

It was in response to @DegustatoR suggesting the possibility of a 5080 offering no improvement in performance despite being a bigger chip with a wider bus. That would necessitate some drastic new paradigm like with Turing.

Gotcha. Yeah that doesn’t seem plausible.

 

[Man from Atlantis]

I think there were rumours that blackwell will take some or most of the BVH work from CPU to GPU.

 

[DavidGraham]

I think OMM made an appearance in CP2077 but that’s it

Also in Portal RTX and Portal RTX Prelude. There is also SER in more games: Witcher 3, Sackboy, Portal RTX, Portal RTX Prelude, F1 2023, Cyberpunk and UE5 in general.

RTX Remix titles coulde be using SER and OMM automatically.

I think there were rumours that blackwell will take some or most of the BVH work from CPU to GPU.

Blackwell is rumored to have Ray Tracing Level 4 or Level 5.

There is also some speculation that NVIDIA will integrate generative AI for NPCs and characters in certain titles, and will rely on Blackwell to deliver responses with minimal latency.

 

[DegustatoR]

What new rendering feature would be a big enough sell?

I dunno.
Strictly speaking an increase in die area for the same performance can happen even without the addition of any new rendering features - this is almost what we've got with Lovelace where caches were added to maintain similar performance level (AD106 vs GA104 for example).
I'm just pointing out that the idea that a bigger die with wider bus would somehow guarantee a higher perf/price improvement is a flawed one to being with.

 

[techuse]

I dunno.
Strictly speaking an increase in die area for the same performance can happen even without the addition of any new rendering features - this is almost what we've got with Lovelace where caches were added to maintain similar performance level (AD106 vs GA104 for example).
I'm just pointing out that the idea that a bigger die with wider bus would somehow guarantee a higher perf/price improvement is a flawed one to being with.

AD106 has more transistors but is a much smaller chip.

 

[DegustatoR]

AD106 has more transistors but is a much smaller chip.

Ampere to Ada was a huge scaling step which is unlikely to repeat any time soon if ever.