Sunday, January 12th 2025

AMD Ryzen AI Max 395+ Mini PC: GMK Announces Strix Halo-Powered Compact System

At CES, AMD unleashed the much awaited Ryzen AI Max "Strix Halo" APUs with mammoth iGPUs, up to a whopping 40 CUs for the Radeon 8060S. These chips are powerful enough to not require discrete graphics at all, making them ideal for mini PCs, which lack the physical room for dedicated graphics. GMK appears to be among the first to announce a mini PC with the top-end Ryzen AI Max+ 395 APU, although any further details are under wraps as of now.

Unlike the Strix Point parts, Strix Halo abandons the smaller and more efficient Zen 5c cores for a Zen 5-only setup, with up to 16 Zen 5 cores for the highest-end Ryzen AI Max+ 395 SKU. This allows for some serious performance potential, with AMD promising substantially better performance than both Intel's Lunar Lake and Apple's M4 Pro, although it would be much fairer to compare Strix Halo to Apple's M4 Max, and Intel's Arrow Lake-H/X instead. Regardless, there is no denying Strix Halo APUs open up new doors in terms of performance for compact systems, the rest remains to be seen as and when the products reach reviewers.
Source: Notebookcheck
Add your own comment

43 Comments on AMD Ryzen AI Max 395+ Mini PC: GMK Announces Strix Halo-Powered Compact System

#26
Punkenjoy
EsliteMoby40 RDNA 4 CUs look like it's gonna outperform the PS5 36 RDNA 2 CUs. But will it bottleneck by DDR5 bandwidth?
That is a good question

Those are RDNA 3.5 and they are designed to work with LPDDR5 memory. It have 256 GB/s of bandwidth instead of the 448 GB/s of the PS5 but it will have 32 MB of MALL cache (infinity cache) just for the GPU.

It's unclear what is the bandwidth of that cache so it would be hard to calculate effective bandwidth of this GPU at different resolution. But as a reminder, this is what AMD published for cache hit ratio per resolution for RNDA2

Posted on Reply
#27
dragontamer5788
PunkenjoyIt's unclear what is the bandwidth of that cache so it would be hard to calculate effective bandwidth of this GPU at different resolution. But as a reminder, this is what AMD published for cache hit ratio per resolution for RNDA2
LLMs, especially the 80B parameter ones these machines are designed for, cannot reliably cache anything.

It requires you to run through all 80B parameters to compute even just one token. That's usually 80 GBs of weights. There's no caching strategy available here. Generative AI / Painting AIs are convolutional in nature and can benefit from caches. But as you see... it very much depends "which" AI you're trying to run on these machines.

Caches for RDNA / Video Games exist because the Framebuffer (aka: the screen's pixels) is very commonly used in video games. Not just for output, but also as inputs (TAA, and other anti-aliasing effects... blur effects and more). Storing the entire framebuffer/screen in Cache makes a big difference for video games.
Posted on Reply
#28
Gooigi's Ex
HOkayOh wow, that truly is hideous. I can't really explain why, it's just...all ugly.
Yeah I’m gonna lie, it does look fugly. Would definitely put a wrap around it lol
Posted on Reply
#29
A&P211
Sound_CardStrix Halo is worrying the hell out of Intel and Nvidia. This is a chip that is going to take a sizable bite out of everyone's lunch so long as TSMC can supply it, and it quite honestly the chip I care most about as I am migrating over to mini PC's. I have a Phoenix point Mini PC right now. Zen 3 with 12 RDNA 2 CU's. The damn tiny thing is awesome.
My current laptop has a 8845hs, I see myself using the AMD 780m gpu more than the 4070 mobile. I usually play more games with the 780m, fits just perfect.
Posted on Reply
#30
Shou Miko
TropickLooks like partners can opt to use LPDDR5X as well. HP is launching the Z2 Mini G1a which is a similar form factor but with soldered 8000MT/s memory and a MAX+ 395 PRO. That'd be LPDDR5X right?

Edit: not to mention this thing looks way nicer than the GMK box

www.hp.com/us-en/workstations/z2-mini-a.html

I cannot remember if it was Acer that showed off a AMD Ryzen AI Max+ 395 with without soldered memory which I am intrigued to see the performance between DDR5 vs LPDDR5X ram but I wish there would be a CAMM2 model instead of soldered memory so when there is a memory error you can change the model out and performance between CAMM2 and LPDDR5x should be on pair or am I far off?

I really hope the AMD Ryzen AI Max+ 395 based system won't be hella expensive because I personally want one and it can replace my current system if gaming performance is good this is why I would love to also see non-soldered vs soldered performance.

It would be awesome if AMD would release a Ryzen Max+ 395 without AI so it could be cheaper for normal people and the gamers that could need this one.
ChaitanyaProbably due to fact that CAMM2 and LPCAMM dont support 256bit quad channel memory which is what these AI Max series of CPUs are designed around.
Yeah this apu got quad channel memory support this was something I wasn't even thinking about. But this I would love to see what happens when there is a non-soldered version to see the performance difference.
ChaitanyaNot a fan of mini DP ports.
Same, but there is a place for it somewhere but again it's I am happy HP didn't go with mini-hdmi.
Posted on Reply
#31
Unregistered
JasBCYou can't install Asahi Linux on the Mac Mini?
You can, but Apple is not helping that team with documentation or drivers, so theres a lot of the hardware not being properly utilized.
EsliteMobyBut will it bottleneck by DDR5 bandwidth?
Its quad channel.
ymdhisIt is using 256bit DDR5, so it's still bottlenecked but not as much as say a 8700G.
But in quad channel.

Per this TPU article : www.techpowerup.com/330548/amd-debuts-ryzen-ai-max-series-strix-halo-soc-up-to-16-zen-5-cores-massive-igpu
To keep the up to 16 "Zen 5" CPU cores, the large iGPU with 40 CU, and the 50 TOPS NPU fed with sufficient memory bandwidth, AMD has given the "Strix Halo" a quad-channel (256-bit) LPDDR5X-6400 memory interface. This should give the processor a memory bandwidth of 256-bit. The iGPU talks to the system over the core switching fabric (Infinity Fabric) of the SoC die.
Sound_CardStrix Halo is worrying the hell out of Intel and Nvidia.
The scary part is, this APU was supposed to be released 2 years ago.

Not sure why the delay.
#32
kondamin
Over priced short run experimental product for both amd as the little box makers.
Posted on Reply
#33
Sandbo
dragontamer578840 CUs? The Rx 580x was 36 CUs.

That's a mighty amount of GPU-compute.
But it was 7 years ago. Not to say it wasn't a good start but it's more of a finally.
Posted on Reply
#35
R0H1T
Sound_CardStrix Halo is worrying the hell out of Intel and Nvidia.
Their biggest target is Apple, especially MacBook pros!
Posted on Reply
#36
3valatzy
R0H1TTheir biggest target is Apple, especially MacBook pros!
Can't be. That is a completely different market segment, RISC, non-x86 and non-Windows ecosystem.
AMD is a mess now.
Posted on Reply
#37
kondamin
R0H1TTheir biggest target is Apple, especially MacBook pros!
nah, it’s a 45w product the m4 isn’t even 10w.

its a pipe cleaner product to see if there is a real market for them.
if there is a big demand for it they will quickly replace it with an RDNA 4 gpu if there isn’t it’s going to be a one off.

given the state of the economy and the slump the pc market is in, it’s going to be a one off
Posted on Reply
#38
R0H1T
I'm talking about the (Halo) chips in general, not just this product as such.
3valatzydifferent market segment, RISC, non-x86 and non-Windows ecosystem.
How does that matter, unless you were talking about something else?
kondaminnah, it’s a 45w product the m4 isn’t even 10w.
You're kidding right o_O

Posted on Reply
#39
kondamin
R0H1TI'm talking about the (Halo) chips in general, not just this product as such.

How does that matter, unless you were talking about something else?


You're kidding right o_O

No I’m not kidding, the m4 and the m4 pro are vastly more energy efficient chips.
this thing only starts at 45w and has the option to go all the way op to 120w
Probably even more when stressed.
Posted on Reply
#40
R0H1T
Yes but they're not ~10W chips as you said, also the top m4(Pro/Max) chip probably peaks closer to ~80W in MT tasks. The 12c M4 pro peaks to 73.2W in CB15 multi, it could likely consume a lot more in the latest benches. Granted this doesn't exclude other components & the fan but still, these things aren't that efficient as they're portrayed!
www.notebookcheck.net/Apple-MacBook-Pro-14-2024-M4-Pro-entry-laptop-review-Who-needs-the-M4-Pro.932642.0.html

Apple's biggest advantage is the Mac OS & of course their entire well oiled & controlled/closed ecosystem
Posted on Reply
#41
Mack4285
Sound_CardStrix Halo is worrying the hell out of Intel and Nvidia. This is a chip that is going to take a sizable bite out of everyone's lunch so long as TSMC can supply it, and it quite honestly the chip I care most about as I am migrating over to mini PC's. I have a Phoenix point Mini PC right now. Zen 3 with 12 RDNA 2 CU's. The damn tiny thing is awesome.
I think AMD has supply problem from TSMC, not getting enough wafers allocated due to often being outbid by e.g. Apple, Nvidia and Intel. So that's one reoccurring problem.

Second, AMD has produced better CPU:s than Intel for a long time now, yet we still see Intel systems, motherboards and laptops dominate. Intel is clearly bribing manufacturers to use their stuff. Nvidia as well. They might even have exclusivity deals, and deals that say they own a certain laptop design from a company. E.g. Lenovo and ThinkPad X1 Carbon, Dell XPS, etc. Ever wondering why we don't see AMD chips in them? Because Intel owns the designs.

In other words, even if AMD has a very appealing product, as in this Strix Halo, it is difficult for them to penetrate the market. If AMD was in a dominant position themselves, they would act the same as Intel/Nvidia. These giant coorporations really don't care about anything other than your money. We should not feel sorry for AMD. Just hope that these more powerful APU:s are becoming more standard from now on.
Posted on Reply
#42
Makaveli
I'm so tempted to grab one of these for a Dedicated machine to run LLM's.

Currently running them on my 7900XTX which is good with 24GB of VRAM for models in the 30B range, however this where you have access to 96GB of unified VRAM would allow me to run much larger models 70B easily.

However price will be high and this is just a nice to have currently. Can wait on this abit to see what prices look like in a year.
Posted on Reply
#43
Kiddles
Mack4285Could be the future since big bulky performance PC:s are getting tiresome, at least for me. But it has to be executed right. Aesthetics, price, BIOS, I/O and a good quality cooling solution that doesn't sound like a jet engine. Would have better longevity than a Mac mini since Linux could be installed.
I'm in a similar boat. I'm kinda just done with building a PC every god damn time, and I don't play enough AAA, or even AA games, to warrant the need for a high end GPU anymore. I just want something power efficient that has a small spatial footprint at this point, but is still enough to play the games that I already have in my Steam library at 1080p/1440p, 60 FPS. Plus, 40 CUs boosted up to 2900 MHz (assuming it clocks that high) would offer around 14.8 TFLOPS, which is plenty for getting 120 FPS in CS:GO and other competitive shooters that aren't trying to go too hard on visuals.

That said, the BIOS for GMKTech computers isn't the most feature-rich, but it's generally enough for what most people are using these things for. And don't expect to get under 84°C at 100% load with the limited cooling most MiniPCs have.
Posted on Reply
Add your own comment
Nov 16th, 2026 04:50 CST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts