Friday, August 7th 2026
AMD RDNA 4m iGPU Support Arrives in Open-Source Mesa GPU Driver
AMD's mysterious RDNA 4m integrated graphics have reappeared after a few months of silence, as the GPU has officially been merged into the Mesa 26.3 open-source driver stack. Running under the GFX1171 GPU target, this RDNA 4m is neither a true RDNA 4 design nor a simple RDNA 3 GPU IP. It sits just above RDNA 3.5, with supposed capabilities to support INT8 data types and native FSR 4 technology. However, since AMD announced that older RDNA generations are now getting FSR 4 upscaling, the true purpose of RDNA 4m remains unclear.
As readers may recall, AMD plans to keep its RDNA 3.5 iGPU in use until 2029, meaning most consumers buying laptops in the coming years will experience the same integrated graphics capabilities as those found in today's Ryzen AI 300 and 400 series processors. This suggests that the RDNA 4m, despite its limitations, will be a welcome addition to integrated graphics. "Redstone" features such as ML-based upscaling and rendering technologies rely on the RDNA 4 IP to function effectively. With RDNA 4m support now merged into the Mesa open-source graphics driver, the RADV and RadeonSI drivers are being updated for RDNA 4m.We can expect the first appearance of RDNA 4m in "Medusa Point," which will introduce a desktop-exclusive RDNA 4 with the new RDNA 4m variant. In contrast, "Medusa Halo" will utilize AMD's next-generation RDNA 5/UDNA GPU microarchitecture. Although initial comparisons between the two are lacking, some instruction set extensions, such as WMMA and SWMMAC instructions, indicate support in the new "GFX1170" GPU, which should be associated with the GFX11 generation, also known as RDNA 3. Currently, this is believed to be an upgraded RDNA 3 with many RDNA 4 modules, enabling FSR 4 support even on the less powerful "Medusa Point" APU.
Source:
Phoronix
As readers may recall, AMD plans to keep its RDNA 3.5 iGPU in use until 2029, meaning most consumers buying laptops in the coming years will experience the same integrated graphics capabilities as those found in today's Ryzen AI 300 and 400 series processors. This suggests that the RDNA 4m, despite its limitations, will be a welcome addition to integrated graphics. "Redstone" features such as ML-based upscaling and rendering technologies rely on the RDNA 4 IP to function effectively. With RDNA 4m support now merged into the Mesa open-source graphics driver, the RADV and RadeonSI drivers are being updated for RDNA 4m.We can expect the first appearance of RDNA 4m in "Medusa Point," which will introduce a desktop-exclusive RDNA 4 with the new RDNA 4m variant. In contrast, "Medusa Halo" will utilize AMD's next-generation RDNA 5/UDNA GPU microarchitecture. Although initial comparisons between the two are lacking, some instruction set extensions, such as WMMA and SWMMAC instructions, indicate support in the new "GFX1170" GPU, which should be associated with the GFX11 generation, also known as RDNA 3. Currently, this is believed to be an upgraded RDNA 3 with many RDNA 4 modules, enabling FSR 4 support even on the less powerful "Medusa Point" APU.

32 Comments on AMD RDNA 4m iGPU Support Arrives in Open-Source Mesa GPU Driver
AI has bought all the factory time for next two years, this is vaporware announcement
my point is though they didn't even try. one model is not enough.
Decent device with performance comparable to the laptop 3070 / 3080 (lower TGP) and an overall decent laptop.
AMD then followed this up with..... er.... um..... well there were a few Asus TUF, Zephyrus and HP Omen* models with the 6600/7600 and 6700/6800(S) models - all basically variants of the 6600/7600 desktop parts.
Rumour was there was meant to have been a 7900M XT (likely a binned 7900 GRE or 7800XT desktop die) but that seemingly never happened.
Notably, both the Desktop 7900GRE and Mobile 7900M use Navi 32-sized packaging while retaining Navi31's MCM structure (though, shader-cut and gimped to 256-bit).
It is my sincere and deeply-held belief that the lack of interest in 7900M, is what facilitated 7900 GRE's global launch.
IMO, 7900M-destined silicon, got turned into Navi 31 XL 7900GREs
There's a reason they had so many mobile 7600M units that found their way to the Steam Machine.
Lastly, I think AMD still considers themselves to be a hardware company first and foremost while Nvidia considers itself an ecosystem of sorts. The latter offers a more complete solution. They have made some changes though due to the AI boom and have started to put effort into ROCm.
These were discontinued after a generation or two.
While Nvidia is talking about N1 and N1X series, AMD's Strix Halo can only be found in a laptop or two.
They aren't consistent with their vision and execution.
But since it have to make compromise on memory (using lpddr6 instead of GDDR7 by example), the bandwidth will still be less than a comparable GPU. The Radeon 7700 is a 40 compute unit (RDNA3) like the 8060s in the top strix halo (but on RDNA 3.5). the 7700 have about 640 GB/s Bandwidth where the 8060s is capped at 256 GB/s. The difference is quite significant and explain why the dedicated GPU is much more powerful dans the large IGPU of similar size.
The main advantages of Strix Halo is how much memory it can support. For gaming, that is pretty much useless as of now (vs comparable GPU with 16GB, it is indeed way better than a 4 GB GPU) but for running local llm that run in VRAM/RAM, it do wonder.
This is why strix halo sell quite well, but people that are buying it aren't playing games. They run local llm to code with privacy and without explosive token cost.
A lot of people see the writing on the wall that we are entering in a enshitification phase from the large LLM where they try to jack up the price to start making money. People just try to become autonomous and limit cost.
For hardware, this mean the pressure on hardware will remain high for quite some time. I think that it's in the interest of the large ai shop that Ram stay as expensive as it is. Local LLM aren't as good as the largest frontier model like Anthropic Fable, but they are very usable and getting better and better every year.
The scenario where an individual can purchase a 256GB large iGPU like one of the future AMD halo product for a reasonable price is not something that the large ai company want. People will be able to get a lot of stuff done with free open weight model that are very good. Maybe not as good as the largest model, but good enough.
This can't last forever.
People complaining that iGPU using RDNA4 is "old tech".
Make up your mind. We're talking about iGPU's here not mGPU's.
As I see It, AMD's fault is their stupid rebrandings along with naming and that they didn't update their APUs for some time. CPU performance is Ok, they are using their latest cores, what is worse is that IGP was not updated to RDNA4, which is much better than RDNA3/3.5. Then battery life is not bad but Intel's latest is better.
For AMD the good news is that Panther Lake with B390 IPU is expensive and the rest is using only the much weaker 4EU variant.
What I want to know is what iGP architecture will Medusa Point use, Medusa Halo will use RDNA5.
It would be interesting to see if AMD has considered using its 3D cache tech for future APUs, making the main igpu chip smaller and slapping a big cache chip underneath seems like a reasonable path to take.
AMD is taking a different route here. Even Nvidia is not able to ignore it, as they too are about to deploy N1X with integrated GPU. Strix Halo has been a pilot, experimental product exploring markets it could serve. More of gaming adoption is out of question during RAM prices crisis, but this can change easily in future, if or when RAM prices come down. Last year, before the crisis, Strix Halo mini-PC with 96GB was ~$1,600
For Medusa Halo, 384-bus with LPDDR6 is rumoured, so we will see how this works out. If they manage to utilise LP6 chips at 12.8 or 14.4 Gbps, that already going to be miles better. Bandwidth of mega iGPU is smaller, of course, but this is well-compensated by being more energy efficient and not requiring gigantic footprint in laptop case. It's been only one generation, so next iterations will only get better.Not really. Their x86 laptop share has grown from 13% in the beginning of 2019 to ~29% in 2026. Check out Mercury Research data. "Negative brand perception" is directly contradicted by those numbers. Essentially, the idea of "negative brand perception" is BS, unless you could convince us that increasing number of buyers of AMD laptops are idiots who buy and contribute to market share while creating "negative brand perception". Good luck with that one.Medusa Point will use GFX 1171/1172, which is RDNA4m, a hybrid of RDNA3.5 with some instructions from RDNA4.
There are plenty of asian ODMs that were not impenetrable that could have made their Radeon Books.Well, they ARE wasting time and silicon on that market, because that market is HUGE, so why not put a little effort in?
But keep in mind, this is AMD, the same AMD that tried to force ODMs to write their own drivers for Ryzen APUs then pikachu faced when nobody used their chips.This is true. AMD did stick with rDNA 3.5 in most of their products and refuses to support it like their competitionThis is ALSO true. rDNA5 is coming and will be AMD's main focus, and with 3.5 we've already seen how AMD treats older oddball architectures. By 2029, rDNA4 will indeed be very old.More then 1 thing can be true at a time.There is nothing stopping an iGPU from using the newest uArch. See also: Intel.
What is four years old today? Oh wait - Ada Lovelace, that launched back in 2022.
Is someone calling 40 series very old today? I doubt it.
And i doubt anyone will call RDNA4 very old in 2029 when only RDNA5 is newer and RDNA6 is yet to launch.
Four years was old back when each new generation brought significant changes and performance uplifts. These days people can easily run 5+ year old hardware and it doesn't feel old by any means because the gains from newer generations have slowed down so much.
Also, AMD is the only company that broke the monopoly of HDMI Forum and created support for HDMI 2.1 48 Gbps and associated features for open source driver on Linux. Now, Intel is trying to catch up and they are doing the same, while Nvidia is dragging their feet.Irrelevant. For example, majority of users use processors from 5-6 years ago, and they still work. If you are in a tiny group that upgrades regularly, you will be able to do this anyway.As you said, more than one thing can be true at a time.
Intel has different cadence and different business priorities. They did a good job with Xe3 iGPU, but unfortunately, those top iGPU laptops and handhelds are very expensive due to memory price crisis. However, their discrete GPU business is done, plus Nvidia will enter into some iGPU deals. So, it's give and take. None of the three vendors is competitive or present in all possible segments, and that's fine.
But indeed, FP8 was added to rdna 4m