Detailed AMD Medusa Halo and Medusa Halo Mini APUs leak claims up to 26x Zen 6 cores and next-gen RDNA 5 iGPUs

[Redaktion]

In a comprehensive leak covering AMD Zen 6 APUs, including Medusa Point, serial leaker Moore's Law Is Dead has revealed a ton of details regarding the Medusa Halo and the Medusa Halo Mini APUs. Both APUs feature a mix of Zen 6 CPU cores and an RDNA 5 iGPU.

https://www.notebookcheck.net/Detailed-AMD-Medusa-Halo-and-Medusa-Halo-Mini-APUs-leak-claims-up-to-26x-Zen-6-cores-and-next-gen-RDNA-5-iGPUs.1093912.0.html

[Papua]

People haven't even been able to get their hands on strix halo chips yet.

[Drakkan]

People haven't even been able to get their hands on strix halo chips yet.

Eh? I got one months ago with no effort.

[Hotz]

AMD: RDNA 4 iGPUs haven't even arrived in the bulk of mini-PCs and laptops (only RDNA 3 or 3.5 in the case of Strix Halo), and yet they wave the flag about RDNA 5....

INTEL: Xe2 iGPUs haven't even arrived in the bulk of mini-PCs and laptops (only as niche products, whereas it should have been in Arrow Lake or Refresh), and yet they wave the flag about Xe3....


It's disgusting how both of them are jumping from one iteration to the next within such a short time, without delivering in large quantities, and make you feel whatever you buy is old §hit, and cause the driver support being be stopped earlier, as only the new GPU iteration will be served.

[Alphonsokurukuchu]

ofcourse it's MILD

[RXAX]

AMD: RDNA 4 iGPUs haven't even arrived

I don't think rdna4 igpus are ever coming. All the leaked timelines for it suggest very late release dates, like 2027. By that time, they might as well just go rdna5/udna. It's not unusual for amd to skip generations. After all, didn't they skip rdna1 by going directly from vega to rdna2?

Also, wasn't there an interview from someone pretty high up at amd recently, where someone asked about amds plans for rdna4 outside of desktop? And the response was pretty much, rdna4 is mainly for desktop only, that there aren't any plans currently in the making and not to expect much else.

Months ago there were some leaks about rdna4 mobile dgpu being worked on but it wouldn't surprise me if they've been cancelled. Nvidia is just too dominant when it comes to laptop dgpus with OEMs, so it seems AMD is just doubling down on the APU/igpu strategy.

[256-bit vs 384-bit]

a 384-bit LPDDR6 or a 256-bit LPDDR5X memory controller.

(AMD's Strix Halo has a 256-bit wide memory interface and supports up to 128 GB RAM. Strix Halo: 8000 MT/s * 256-bit / 1000 / 8 = 256 GB/s.)

AMD specifically advertises Strix Halo for running AI / LLMs and if the memory density wouldn't change with LPDDR5X in 2027+, then the 256-bit rumor does only make sense, if the "Halo" word is defined to stay at 256-bit / at basically the same performance as Strix Halo (just like the "Point" word in e.g. Phoenix Point indicates the same tier). This [using LPDDR5X] would make Medusa Halo a refresh/rename of Strix Halo, just like Phoenix Point to Hawk Point. Hawk Point came one year after Phoenix Point, but Medusa Halo would come out 2+ years after Strix Halo and to be at the same tier / performance and memory size (let's say it will be 10000 MT/s vs 8000 MT/s, but still at the same up to 128 GB RAM), is almost unexpected?

In 2027+, what would really be needed, is at least a 384-bit wide chip (preferable a 512-bit one): All the big and smart generalizer MoE LLMs require at least a 384-bit wide memory interface and the 192 GB RAM that would come with it to fit into the RAM.

If DDR6 / LPDDR6 would (almost) double the memory bandwidth and double the memory density at the same bit width, then a 256-bit width would be enough:
14400 MT/s * 256-bit / 1000 / 8 = 460.8 GB/s and up to 256 GB RAM. Improvements in the MoE LLM architecture, would also reduce the requirement for the memory bandwidth: 8000 MT/s vs 14400 MT/s is not double, but then it wouldn't need to be either. But the doubling in memory size would be required, as this is how MoE scales (fewer active parameters (=less memory bandwidth required), but more in size to compensate for it).

[1]:

• huggingface.co/unsloth/GLM-4.7-GGUF (the Q3_K_M quant is 171 GB, so one would at least need 171 GB RAM + additional 10-20 GB for the OS and the context)
• huggingface.co/unsloth/MiniMax-M2.1-GGUF (at least a Q4_K_M quant (138 GB) is recommended according to the latest user comments)
• huggingface.co/unsloth/MiMo-V2-Flash-GGUF (Q3_K_M (147 GB) to Q4_K_M (187 GB))