AMD could introduce a product line based on a one-size-fits-all chiplet approach featuring a multi-use Graphics Memory Die (GMD). A few days ago, new rumors surfaced by leaker Kepler_L2 that AMD will be using names from the classic kids show, Transformers, for its upcoming RDNA 5 lineup. However, among those details, 3DCenter picked up on another possibility and has run with an interesting concept whereby AMD might be working on an almost modular approach, allowing one particular chiplet to be used in multiple applications.
Per 3DCenter (machine translated):
“In addition to the code names alone, which of course do not change the technology, further discussion on Twitter/X has also led to other interesting statements from ‘Kepler_L2’. At least AT2 is said to be implemented as a ” GMD chiplet ,” a “Graphics Memory Die” with arithmetic units, cache, and memory interface, but without a video/media engine and without a PCIe interface.”
Kepler responded on social media, clarifying that the GMD lacks certain GPU functions such as display, media, and PCIe, and as far as they know, the above approach is not planned for RDNA 5. However, both parties seem to agree that AMD might have other plans in store for the future.
Indeed. This was included this as a hypothetical possibility. It may be usable in the future, but it's obviously not planned for RDNA5 (otherwise they wouldn't need so many AT chips).
— 3DCenter.org (@3DCenter_org) September 5, 2025
While it sounds like using the same GMD for various AT chip applications across the board isn’t in the cards for AMD’s next GPU series, the idea of it seems plausible as a cost-saving strategy. From PC to console SoCs, APUs, and more, there are benefits to be found. Obviously, this is all pure speculation, but AMD has been prominent in exploring new package designs for multiple usage scenarios, whether it be for GPUs, CPUs, or APUs.
“But even if this option isn’t yet used for the RDNA5/UDNA generation, the technical foundation for future multi-chip solutions would already be created – which could then be activated in subsequent UDNA generations.”
– 3DCenter (machine translated)
Discussion (7 replies)
Join Discussion →Very interesting. Here's hoping for real competition.
It appears that Microsoft is going with chiplet models for both handheld & console (as opposed to traditional monolithic approach by Sony)
Both will have 2 chiplets
In an interesting (clever?/desperate??) approach it seems the cpu + i/o chiplet will be shared between both handheld & console options
The GPU chiplets would be shared with AMD's traditional discreet GPUs
As you can see above the console GPU ~ 3x handheld GPU
This is similar to the current ratio between series S & series X
Effectively
handheld = series S (priced between $800 to $1000)
console = series X (priced approximately as 5080 Street price by next year end?)
Xbox handheld (medusa premium) vs console (magnus)
I wonder if their ai MAX cpus have given this creedence.
They aint doing nothing. No game changer anything, graphics die my butt.
Thanks for that scintillating rebuke and sharing your opinion as well informed as it was...
Hehe, its just AMD has frustrated me with their lame APU strategies for so long, i ended up with Intel mini PCs and I blame AMD.
Fine fine my Intel mini pcs are quite decent.
I am aware of Ryzen mini PCs, fairly decent yes, I understand the price point to be absurd considering how much less hardware it is vs a laptop, and how much less of everything thermal you have to do then doing a small box vs a laptop.
Idk, it might be an everybody else problem that is not worth it for AMd to do better APUs, meaning they woulnt sell and will always be niche i guess, chicken and egg and what not.
And thats why my insightful commentary :) .
LTT just did a review of one of their mini systems with the AIMAX+ setup... seems quite proficient for 1080p performance. And even better for AI.