Nvidia shelves 2026 new card releases

wsterdan · September 25

Cam Fox said:

wsterdan said:

Cam Fox said:

I'm still rockin a 2012 macbook pro as one of my daily drivers. No, it can't run Daz Studio. :) It can't even run OSX, so I put Linux on it. Unironically, cloud ai increases the value of such an old laptop since I can still get real work done.

That's odd, I have a 21-inch 2011 iMac and a 2012 27-inch iMac, both running DAZ Studio with OSX; I use them for doing any DAZ Studio lip sync. Your 2012 MacBook Pro can't run the latest OSX, but I've been running OSX since 2001 so it should come with OSX installed when you bought it and I believe it can run up to Catalina (10.15).

As to Cloud-based AI, heck ya, I use it to make images and animations on my very low-end iPad.

Yeah, mybad, I meant latest/modern OSX. The last version it could run is 7 versions behind the current. It used to have Yosemite before I put linux on it, that's 12 versions back. I could probably put Catalina on it to get Daz Studio running, but it wouldn't be much use for iray stuff due to system specs. I'd eventually like Daz Studio running in Linux but I'd use a newer workstation for that. (I read some folks have got it working, at least!)

Thanks for taking the time to clarify, much appreciated.

Padone · September 26

update. GTX 1060 VIDEO.

Well, after palying with Anima I decided to give a look at video generation. By the way I advise Anima a lot, the Lumina 2 AI chat is extremely powerful and can also turn Anima into a semi-photoreal style, despite everyone keeps telling this is not what the model is for.

https://huggingface.co/circlestone-labs/Anima/discussions/236

To make video generation on a 1060 I struggled a little, because vram is 6gb. But finally I managed to find the right guys for the job. The key is always to use quantized versions, plus there's a distilled quantized version for LTXV 0.9.6, that means the video generation is fast, same as Anima Turbo with images.

https://huggingface.co/city96/LTX-Video-0.9.6-distilled-gguf

https://huggingface.co/city96/t5-v1_1-xxl-encoder-gguf/tree/main

With these guys and the powerful memory management of ComfyUI, I can do 768x512 animations with no limits for the frame numbers, it takes 4.5 vram max and 9 gb ram max. The quality is good enough, it preserves the semirealistic anima style and does animation on a preview image. Here it computed 57 frames in about 90 seconds, thanks to the distilled version.

edit. Below as example there's Chun-Li waving at the viewer. Note that LTXV is trained with real photos, so it tends to change anime characters into real people. Another great feature of Anima is, I didn't use any LoRA, as Anima is heavily trained on anime and games so it knows perfectly how Chun-Li is made, just mention her in the prompt and Lumina brings her out for you. I apologize for the not excellent quality I'm just started with LTXV so can't hadle the parameters well yet.

For the mp4 just remove the zip extension as the forum doesn't allow mp4.

omvendt · September 26

garrett_3d said:

As long as people continue to use AI, in whatever form it takes, then this situation will continue. The only solution, bar an entire collapse of the internet, is to ban AI completely worldwide.

AI is not the answer - it is only as intelligent as the muffins that program it. Humans aren't that intelligent. We're screwed.

I am sympathetic to the sentiment, Garrett, but the Chinese won't ban AI so if we do that, we jeopardize our future. Better to control AI with sensible legislation and also prepare for potential mass unemployment. In the US, that means Medicare for all, or something like that.

omvendt · September 26

I read that AI centers replace their chips every two years. Should that not mean that there will soon be a flood of used GPUs for us to buy?

It is weird, I had an old PC with two 1080Titan GPUs for sale for a few hundred dollars, and there was never any interest. They have 12 GB RAM each, I think. Maybe I should take out the RAM and the GPUs and sell them separately.

jmucchiello · September 26

omvendt said:

I read that AI centers replace their chips every two years. Should that not mean that there will soon be a flood of used GPUs for us to buy?

It is weird, I had an old PC with two 1080Titan GPUs for sale for a few hundred dollars, and there was never any interest. They have 12 GB RAM each, I think. Maybe I should take out the RAM and the GPUs and sell them separately.

No AI data center has been online for 2 years. The CEO of NVIDIA suggests upgrading data centers every year. And the kicker, it takes more than a year to bring a data center online. Many data center projects are on pause because there's a lack of demand.

IceCrMn · September 26

datacenter hardware and consumer hardware are not compatible at all really.

https://servermall.com/blog/server-gpu-vs-consumer-gpu-overview/

https://etcjournal.com/2025/09/12/desktop-gpu-vs-data-center-gpu/

this is what an HBM memory module looks like

https://www.knowres.com/product/krm-9vh1x82/

heres a common desktop DDR5 module

https://www.walmart.com/ip/Kingston-FURY-Beast-16GB-288-Pin-PC-RAM-DDR5-5200-PC5-41600-Desktop-Memory-Model-KF552C40BB-16/782945716

Yes you could use those datacenter server parts if you have a server rack at home.

Which you can. It's not cheap.

None of it fits in an average consumer motherboard or works with consumer grade power supplies.

We all might be building a home server rack the way it looks on the consumer hardware market.

Used datacenter hardware is only really useful to another datacenter, or someone with a server rack at home. They use HBM memory and other server specific stuff.

Richard Haseltine · September 26

jmucchiello said:

omvendt said:

I read that AI centers replace their chips every two years. Should that not mean that there will soon be a flood of used GPUs for us to buy?

It is weird, I had an old PC with two 1080Titan GPUs for sale for a few hundred dollars, and there was never any interest. They have 12 GB RAM each, I think. Maybe I should take out the RAM and the GPUs and sell them separately.

No AI data center has been online for 2 years. The CEO of NVIDIA suggests upgrading data centers every year. And the kicker, it takes more than a year to bring a data center online. Many data center projects are on pause because there's a lack of demand.

The datacentres may be competing for GPU chips, but that doesn't mean they are actually using regular GPUs - and they are also, I blieve, often using AI-specific chips which are competing for wafers and production time. Memory may be more standard. In any event, at least some of that replacement cycle is going to be drivenm by the heavy load the parts work under which would probably translate into shortened remaining usable life.

jmucchiello · September 27

Richard Haseltine said:

jmucchiello said:

omvendt said:

I read that AI centers replace their chips every two years. Should that not mean that there will soon be a flood of used GPUs for us to buy?

It is weird, I had an old PC with two 1080Titan GPUs for sale for a few hundred dollars, and there was never any interest. They have 12 GB RAM each, I think. Maybe I should take out the RAM and the GPUs and sell them separately.

No AI data center has been online for 2 years. The CEO of NVIDIA suggests upgrading data centers every year. And the kicker, it takes more than a year to bring a data center online. Many data center projects are on pause because there's a lack of demand.

The datacentres may be competing for GPU chips, but that doesn't mean they are actually using regular GPUs - and they are also, I blieve, often using AI-specific chips which are competing for wafers and production time. Memory may be more standard. In any event, at least some of that replacement cycle is going to be drivenm by the heavy load the parts work under which would probably translate into shortened remaining usable life.

There's a glut of Blackwell chipsets NVIDIA has allocated to the various big movers (some say through 2030, some through 2028, they're private contracts so it's hard to know). Their allocation of Blackwell chipsets to consumer markets is nil by comparison. Yes, they aren't actual 5090s in the data centers. But the third party GPU card manufacturers (ASUS, MSI, Gigabyte, PNY, etc) are paying a lot more and competing for a much smaller pool of chipsets with which to create consumer GPUs. And DDR5 VRAM is equally scarse because RAM makers (Micron, Samsung, SK Hynix) have decided to make big profit from scarsity. They could ramp up production. But that would require spending. And actually spending on new fabrication facilities has long term costs whenever the demand for this memory returns to normal levels.

And this isn't "competing for GPU chips", NVIDIA is locking the data centers builders (MS, Amazon, etc) into long term contracts and saying if you don't contract with us now, you might be waiting 3-4 years before NVIDIA has capacity to sell them chips. So a lot of the Blackwell scarsity is FOMO, too.

Richard Haseltine · September 27

The lead time for new capacity is pretty long, and the cost high - it isn't surprising if there is a certain amount of wariness, and even if they do comit it stil won't help for a while. The lines for making the actual high-end GPU (and CPU) chips are even more expensive and time-consuming, which is why there are so few places that can handle the work.

Torquinox · September 27

With respect to gpus, even old gpus, it seems to be a sellers market.

https://www.engadget.com/2267877/your-old-gpu-worth-more-than-you-think/

jmucchiello · September 27

Torquinox said:

With respect to gpus, even old gpus, it seems to be a sellers market.

https://www.engadget.com/2267877/your-old-gpu-worth-more-than-you-think/

That's called demand.

omvendt · September 27

IceCrMn said:

datacenter hardware and consumer hardware are not compatible at all really.

https://servermall.com/blog/server-gpu-vs-consumer-gpu-overview/

https://etcjournal.com/2025/09/12/desktop-gpu-vs-data-center-gpu/

this is what an HBM memory module looks like

https://www.knowres.com/product/krm-9vh1x82/

heres a common desktop DDR5 module

https://www.walmart.com/ip/Kingston-FURY-Beast-16GB-288-Pin-PC-RAM-DDR5-5200-PC5-41600-Desktop-Memory-Model-KF552C40BB-16/782945716

Yes you could use those datacenter server parts if you have a server rack at home.

Which you can. It's not cheap.

None of it fits in an average consumer motherboard or works with consumer grade power supplies.

We all might be building a home server rack the way it looks on the consumer hardware market.

Used datacenter hardware is only really useful to another datacenter, or someone with a server rack at home. They use HBM memory and other server specific stuff.

Thanks for your thorough and documented advice. Another one of my great ideas bites the dust...
(O:

Padone · September 29

update. Krea-2 on GTX 1060.

So, well, after trying anything and see it works fine, I wanted to try the impossible. Fit the 12-billion parameters Krea-2 into my tiny 1060 with 16 gb ram. Guess what, it worked fine.

This time I struggled a lot. The key is always the GGUF Turbo version of course, otherwise forget it. But there's another complication with ComfyUI as the standard GGUF doesn't recognize Krea nor Qwen. While the GGUF provided by Rebel only recognizes the official Qwen version, leaving out customizations. Luckily there's a good guy around who made a working GGUF, much better than the officials.

https://github.com/molbal/ComfyUI-GGUF

https://huggingface.co/realrebelai/KREA-2_GGUFs

With this combination, using Krea Turbo Q3 with Qwen Q4, I get it running on my hardware. Rendering time is bad though, about 3 minutes for a 1024 image, that's horribly slow for a Turbo version. While rendering it takes about 5gb vram and 13gb ram, using the ComfyUI smart memory management.

Bendinggrass · September 29

kyoto kid said:

...one of the reasons I dropped Poser (12, 13, and 14 do so), Daz3D only does so if you opt into it for services like Daz Connect or are have a Daz Premier subscription (because of the exclusive content).

Yeah, I am really hesitant about getting anything that calls home.

omvendt · September 29

Padone said:

update. Krea-2 on GTX 1060.

So, well, after trying anything and see it works fine, I wanted to try the impossible. Fit the 12-billion parameters Krea-2 into my tiny 1060 with 16 gb ram. Guess what, it worked fine.

This time I struggled a lot. The key is always the GGUF Turbo version of course, otherwise forget it. But there's another complication with ComfyUI as the standard GGUF doesn't recognize Krea nor Qwen. While the GGUF provided by Rebel only recognizes the official Qwen version, leaving out customizations. Luckily there's a good guy around who made a working GGUF, much better than the officials.

https://github.com/molbal/ComfyUI-GGUF

https://huggingface.co/realrebelai/KREA-2_GGUFs

With this combination, using Krea Turbo Q3 with Qwen Q4, I get it running on my hardware. Rendering time is bad though, about 3 minutes for a 1024 image, that's horribly slow for a Turbo version. While rendering it takes about 5gb vram and 13gb ram, using the ComfyUI smart memory management.

Thanks for sharing the link. I should be getting my RTX-5090 on Friday. What model would you suggest I install to run locally?

Padone · September 30

@omvendt Just pick a Turbo which fits your vram, the link above provides various quantizations available. By the way I'm a total newbie in AI so probably not the best guy to ask for advice.

Notifications

Nvidia shelves 2026 new card releases

Comments

Adding to Cart…