AMD Radeon RX 9070 XT Could Get a 32 GB GDDR6 Upgrade

[Bruno_O]

Gaming wise, there is little point moving above 12~16GB VRAM in 2025Q1.

unless you play at 4k native + ultra settings + RT + AA

which can and will use over 16GB easily in many games

 

[InVasMani]

Don't worry the 8GB is all you need crowd will be replacing it with a 12GB is all you need card soon enough.

 

[Makaveli]

AMD themselves only support 7900 series desktop GPUs.

Incorrect ROCm supports RDNA 3 & 2

System requirements (Windows) — HIP SDK installation (Windows)

Windows GPU and OS support

rocm.docs.amd.com

 

[Visible Noise]

HIP SDK isn’t ROCm.

System requirements (Linux) — ROCm installation (Linux)

System requirements for AMD ROCm

rocm.docs.amd.com

Edit: That’s hilarious, Auswolf gave you a thumbs up until I posted. Oopsies! lol.

 

[Makaveli]

HIP SDK isn’t ROCm.

System requirements (Linux) — ROCm installation (Linux)

System requirements for AMD ROCm

rocm.docs.amd.com

View attachment 384592

Edit: That’s hilarious, Auswolf gave you a thumbs up until I posted. Oopsies! lol.

I posted Windows and you are posting Linux

 

[Visible Noise]

I posted Windows and you are posting Linux

Christ. ROCm doesn’t run on Windows. A Software Development Kit isn’t a Platform. Pay attention please.

The HIP SDK for Windows brings a subset of the ROCm platform to Windows.

HIP SDK installation for Windows — HIP SDK installation (Windows)

HIP SDK installation for Windows

rocm.docs.amd.com

No debugger, no tools, no AI, no communications. Jesus, Make doesn’t even work on Windows. Tell me again about all the compute apps for AMD desktop GPUs?

Edit: the porting tools have been “Coming Soon” for a year. And AMD is all “We’re a software company now”. What a fucking joke.

Amatuers

 

[Dr. Dro]

What's the bloody point?

As follows:

1. It helps them mask the regression in VRAM capacity in their uppermost segment GPU when compared to the previous generation by introducing this variant
2. It helps them shift cards to AI inference customers and whatever cryptocurrency miners are left, as well as milking some extra profits from the one-off gamer that will purchase this
3. It helps them increase margins by pricing this model significantly higher, while quietly reducing the supply of the 16 GB model with far less backlash from the customer base
4. It helps them state "If Nvidia can build a 32 GB gaming card, so can we!" to shareholders
5. Loyal fanbase will not question any of the four above points and attempt to spin whatever is convenient (specifically, #2) as a positive

unless you play at 4k native + ultra settings + RT + AA

which can and will use over 16GB easily in many games

Not only it generally cannot (16 GB is perfectly adequate for this performance level), but the 9070 XT is not powerful enough to pull this off. I don't think the 5090 really is, depending on the game we're talking about.

 

[AusWolf]

As follows:

1. It helps them mask the regression in VRAM capacity in their uppermost segment GPU when compared to the previous generation by introducing this variant

That "uppermost" segment is $599 x70 level. Nvidia offers 12 GB here. So even AMD's "regression" is more. Hmm...

2. It helps them shift cards to AI inference customers and whatever cryptocurrency miners are left, as well as milking some extra profits from the one-off gamer that will purchase this

Definitely this.

3. It helps them increase margins by pricing this model significantly higher, while quietly reducing the supply of the 16 GB model with far less backlash from the customer base

It'll definitely be priced higher to milk AI customers, but reducing 16 GB supply is pure speculation with no basis in real life.

4. It helps them state "If Nvidia can build a 32 GB gaming card, so can we!" to shareholders

Yes. Not that it affects buyers with half a brain, though.

 

[Dr. Dro]

It'll definitely be priced higher to milk AI customers, but reducing 16 GB supply is pure speculation with no basis in real life.

Only one problem with that, there is no reason to prioritize an $599 16 GB version if they can sell the same chip on an $999 32 GB version that could arguably still be spun as a bargain to many

 

[AusWolf]

Only one problem with that, there is no reason to prioritize an $599 16 GB version if they can sell the same chip on an $999 32 GB version that could arguably still be spun as a bargain to many

There is: company reputation and market share in gaming GPUs.

 

[Jtuck9]

There is: company reputation and market share in gaming GPUs.

Maybe that will become more of a lucrative market segment going forward with a.i. supposedly becoming "more expensive to train and cheaper to run"

 

[AusWolf]

Maybe that will become more of a lucrative market segment going forward with a.i. supposedly becoming "more expensive to train and cheaper to run"

I'm not saying it's not more lucrative. But I can't see AMD, or even Nvidia abandoning the gaming GPU market as of now. Why have any midrange GPUs otherwise? Why not just sell all for AI people?

 

[Jtuck9]

I'm not saying it's not more lucrative. But I can't see AMD, or even Nvidia abandoning the gaming GPU market as of now. Why have any midrange GPUs otherwise? Why not just sell all for AI people?

It would be interesting to see the architectural modification . Maybe these are sort of poised as 9900X3D / 9950X3D equivalents.

 

[3valatzy]

But they are however stringy with their features compared to NVIDIA though. And that's where AMDs problem is, or most of it.

Utter waste of VRAM. No game uses that much VRAM, any VRAM not being used just sits idle doing nothing. All this does is artificially inflates the cost of the card with zero performance improvements. Maybe do better with faster memory than more of it.

Ok. And what about non-game applications ? Is that also an utter waste for them ?

The general Radeon performance for compute is extremely poor.

Modded Skyrim (esp. VR) takes >20GB VRAM, with Nvidia still limiting VRAM, I'd welcome it

This is not the card for you, then. Buy the RX 7900 XTX instead!

Lots of VRAM, not enough horsepower.. cool!

Not enough AMD software support, either.

 

[Jtuck9]

Not enough AMD software support, either.

This is what I'm wondering if they'll slowly integrate (as much as possible) as they move over to UDNA

 

[AusWolf]

It would be interesting to see the architectural modification . Maybe these are sort of poised as 9900X3D / 9950X3D equivalents.

We're talking about GPUs here.

 

[Jtuck9]

We're talking about GPUs here.

Equivalents of the GPU world, i.e. "best of both worlds"

 

[AusWolf]

Equivalents of the GPU world, i.e. "best of both worlds"

Ah, I get your point. Perhaps you're right. It probably won't make any sense purely for gaming, just like the 9950X3D doesn't.

 

[Jtuck9]

Ah, I get your point. Perhaps you're right. It probably won't make any sense purely for gaming, just like the 9950X3D doesn't.

My fault, I'm generally assuming and unclear.

True, but there seem to be a lot of people excited for both! Interested to see if we get a reason for dual V-Cache on Zen6

 

[AusWolf]

My fault, I'm generally assuming and unclear.

True, but there seem to be a lot of people excited for both! Interested to see if we get a reason for dual V-Cache on Zen6

Personally, I'd prefer a single 12-core CCD with V-cache (not that I need it, but that would feel more of an upgrade over my 7800X3D).

 

[Ultron1337]

Lots of VRAM, not enough horsepower.. cool!

It is so for gaming, where GPU has to work pretty hard for for each frame. In LLM world when running pre-trained model aka inferencing you will run into memory amount and them memory bandwidth bottleneck, lack of compute is not the issue. So 32GB 9070XT can run some LLM bigger models faster than 24GB 7900XTX or 4090.

 

[Nhonho]

Useless for gaming, kind of useful as a bone thrown to the AI crowd, but, as noted above, it would have to be priced very aggressively to compensate for the lack of NV amenities like CUDA.

AMD pays a very high price today for not having developed a wide range of software compatible with its GPUs, as Nvidia did.

AMD has always been the type of company that just developed the hardware and expected software developers to optimize their apps for AMD hardware on their own.

On the other hand, Nvidia has always had a very close relationship with software developers to optimize all types of software for its hardware.

 

[bitsandboots]

32gb would be a day 1 buy for me, with all of the AMD crap driver quality that entails.
Been itching to have something faster than a 3090 that doesnt skimp on vram, has displayport 2, and doesnt have power circuitry design flaws that pose a fire hazard.
I just want to VR and AI faster than I can today.

 

[10tothemin9volts]

With the 5090, NV has released a 32GB VRAM consumer GPU, ofc AMD is going to do the same (wasn't it the same with 24GB VRAM consumer GPUs?). The difference is the 9070 (XT) is based on a 256-bit chip using GDDR6 ~600 GB/s vs 512-bit GDDR7 1792 GB/s for the 5090. Still fast enough.

AFAIK, only modded games may require more than 24GB VRAM in 4K right now, but 32GB are nice for fully offloading/hosting big-ish LLMs locally.

Regarding CUDA/ML stack, indeed, I think of AMD GPUs only in terms of running/inferencing LLMs, not training/finetuning, but I read it's still possible and supposedly got easier over the last years, but CUDA is tier agnostic and supports consumer GPUs, workstation GPUs and enterprise cards. To improve this, UDNA (U for unified) will replace RDNA at some point.

5090' idle power consumption unfortunately increased to 30W (4090 22W), but it's still not too bad (it's more than linear in video playback: 54W 5090 vs 26W 4090) considering there are 16 2GB modules (linear increase: 22W[4090]/12[GDDR6X]*16[GDDR7] = 29.33W).

For me to consider this RDNA4 32GB GPU (in no particular order):

• DLSS 2-like upscaling quality improvement
• Fix HDMI 2.1 48GB/s, aka HDMI 2.1a on Linux
• Back to good power scaling like in RDNA2
• Low idle power consumption, linear increase with the amount of VRAM compared to the 16GB VRAM, at the worst
• Just like the 5090, 9070 (XT) 32GB also must be a consumer GPU, so that the price increase is minimal

So, AMD, it's 48GB VRAM consumer GPUs for the UDNA arch after RDNA4 then as well? Would allow to fully offload `Llama-3.3-70B-Instruct-Q4_K_M.gguf` (42.5GB) (by then we will have a different and more capable 70B LLM, ofc), or allow for much higher context.

Try to run a 30GB model on a fancy 24 or 16GB RTX with CUDA and compare the experience against a Radeon with 32GB VRAM.
VRAM is valuable real-estate!

Yes, 24GB VRAM can't fit a e.g 27GB `Qwen2.5-32B-Instruct-Q6_K.gguf` SOTA LLM, but the .gguf format allows to offload the rest of the LLM layers to RAM, but it will run much slower. The tokens per second speed increases exponentially the more layers are offloaded to the GPU, I did some testing:

 

[csendesmark]

Yes, 24GB VRAM can't fit a e.g 27GB `Qwen2.5-32B-Instruct-Q6_K.gguf` SOTA LLM, but the .gguf format allows to offload the rest of the LLM layers to RAM, but it will run much slower. The tokens per second speed increases exponentially the more layers are offloaded to the GPU, I did some testing:
View attachment 384684

Oh, please don't tell me, I am painfully aware how slow things can get when you run LLM-s outside of a GPU
May I ask you to post here?