Thursday, November 12th 2026

AMD "Siren Halo" APU Rumored to Pair Dedicated GDDR7 With LPDDR Memory

AMD's next big Halo APU might ship with two separate memory pools, if a new leak from Gotou_3rd on X is accurate. In a short post on November 12, the leaker wrote that "Siren Halo has dedicated GD7 VRAM." When asked whether that meant unified GDDR memory, like on consoles, or separate pools, Gotou clarified that LPDDR system memory would still be there. He also claimed that Siren Halo follows Medusa Halo, that Siren Point is basically a Medusa refresh, and that the new name exists because AMD won't have a fully new Halo design in 2027. There's no word on capacities, bus widths, or launch timing. AMD already added GDDR7 support to its Linux driver in September, a change widely tied to RDNA 5.

This would be quite a change for the lineup. Strix Halo and its Gorgon Halo refresh use a single LPDDR5X pool shared by the CPU and GPU, with the Ryzen AI Max+ PRO 495 supporting up to 192 GB, of which up to 160 GB can go to graphics. That shared pool is a big reason these chips are popular for running LLMs locally. Dedicated GDDR7 should give the iGPU far more bandwidth for gaming, though it's unclear how much memory AI workloads would get. The leaker also suggested Siren Halo could be related to Magnus, the AMD chip linked to Microsoft's Project Helix, though that is speculation at this point. AMD hasn't confirmed Siren Halo at all, so it's best treated as a rumor for now.
Sources: Gotou_3rd on X, via Videocardz
Add your own comment

5 Comments on AMD "Siren Halo" APU Rumored to Pair Dedicated GDDR7 With LPDDR Memory

#1
ymdhis
This could coincide with the rumour that they plan on using the AT3/4 chips for both Halo APUs and dGPUs. It would certainly make sense to make the chip support both LPDDR and GDDR in that case.
Posted on Reply
#2
Athena
DrUsmanSaleemAMD's next big Halo APU might ship with two separate memory pools, if a new leak from Gotou_3rd on X is accurate. In a short post on November 12, the leaker wrote that "Siren Halo has dedicated GD7 VRAM." When asked whether that meant unified GDDR memory, like on consoles, or separate pools, Gotou clarified that LPDDR system memory would still be there. He also claimed that Siren Halo follows Medusa Halo, that Siren Point is basically a Medusa refresh, and that the new name exists because AMD won't have a fully new Halo design in 2027. There's no word on capacities, bus widths, or launch timing. AMD already added GDDR7 support to its Linux driver in September, a change widely tied to RDNA 5.
Looks like AMD reps will be talking about Halo in a livestream from MicroCenter, perhaps some hints at what's coming up?






www.microcenter.com/site/mc-news/article/micro-center-austin-grand-opening.aspx
Posted on Reply
#4
londiste
I do not think it makes sense to have both. Support for both LPDDR and GDDR7, sure. But not both at the same time. That would invalidate the unified memory feature that is especially important in the AI age.
Posted on Reply
#5
outlw
IF AMD can manage to keep the memory pool unified (probably need something like Intel's Thread Director, to sort out which data is stored in which RAM), I could see this being useful.

GDDR7 would give a huge boost to bandwidth to any data that manages to fit within it, but it is also power hungry, and demanding on the PCB.
LPDDR is slower, but also much lower power, and not as difficult to run traces for.

So, you could (potentially) get the speedup of using GDDR7, while keeping the power effeciency and capaciaty of LPDDR.

I suppose they could also treat the GDDR7 as a massive L4 cache 🤔

IF they can get this working propperly, I would love to see it trickel down to the rest of us!
Being able to slap some SODIMMs directly onto your GPU could be a great way to make local AI workloads much more accessable, without having to design another AI-Centric SKU.
Posted on Reply
Nov 13th, 2026 07:15 PST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts