QuoteFramework Desktop Ryzen AI Max
AMD Ryzen AI Max+ 395, Radeon 8060S
123082 MB/s
Geekom A9 Max, AI 9 HX 370
AMD Ryzen AI 9 HX 370, Radeon 890M
86541 MB/s
HX 370: 128-bit * 5600 MT/s / 1000 / 8 = 89.6 GB/s (does check out with the measurement)
AI Max+ 395: 256-bit * 8000 MT/s / 1000 / 8 = 256 GB/s (does
NOT check out with the measurement)
Can you measure with the newest version 8.0?
QuoteUnfortunately, the system is not compatible with the recent AMD FSR4 update as the GPU is based on the same RDNA 3.5 architecture as the Radeon 890M and not the Radeon RX 9000 RDNA4 series.
Imagine paying 1600 bucks for last gen GPU architecture.
LLMsThe "Strix Halo" APU is a 256-bit chip with a theoretical memory bandwidth of 256 GB/s (256-bit * 8000 MT/s / 1000 / 8) (and ~210 GB/s practically (expected)), comparable to an entry level quad-channel (4 * 64-bit) workstation' memory bandwidth. A normal desktop PC is dual-channel at best. AMD specifically advertises "Strix Halo" for running/inferencing LLMs. You can run the same LLMs on any PC, if you have at least the same amount of RAM (well, running off of a SSD will also work, but the speed will be super slow), ATX sized or not, dual-channel RAM or not, the differences are:
- The size: This is 4.5 Liters (with integrated power supply).
- The RAM speed at which any LLM will be running at: Strix Halo is a quad-channel chip at 8000 MT/s vs a normal PC, which is dual-channel at 5600 MT/s to 6200 MT/s (2*64-bit*6200/1000/8 = 99,2 GB/s)). A (mini-)PC based on the "Strix Halo" APU will run a LLM about 2.5 times faster: 256 GB/s / 99,2 GB/s = ~2.58.
- The RAM upgradability: The LPDDR5X RAM in "Strix Halo"-based PCs is not upgradable, maybe because it runs at 8000 MT/s vs 5600 MT/s to 6200 MT/s typically seen in DDR5 UDIMMs. A DDR5 UDIMM version with upgradable RAM may appear later, but it's not going to be 8000 MT/s, like the soldered ones.
Using the relatively expensive Strix Halo APU/chip, but giving it only 64 GB RAM is wasting expensive silicon, because it's simply not enough for many LLMs (btw: the memory bandwidth will be the same, they are just using less dense RAM chips): Give it at least 96 GB RAM or just the full 128 GB, because on Windows, only 75 % can be allocated to a LLM (as far as I know, maybe it has changed) (that's 96 GB out of 128 GB RAM total).
Questions to ask yourself:
- Is the LLM speed difference of 2.5 times (150 %) and the price worth it vs simply getting 2x48GB RAM sticks or 2x64GB RAM sticks for a fraction of the price and having then more RAM (although, yes, 2x slower) vs paying 1600 bucks and being stuck with the hardware and no upgrade path?
- And, if the size matters, you can still get a mini-ITX case, AM5 mini-ITX motherboard and build a PC of the same size (or get a pre-built mini-ITX PC), with the possibility to:
- Upgrade the RAM.
- Having a dedicated GPU. There are not many choices for 4.0 - 4.5 Liter mini-ITX builds, mostly low profile RTX 4060 or RTX 5060), but this is still better and faster (and harder, stronger, hehe) than the built-in iGPU in Strix Halo:
- You get the ability to upgrade the GPU later, like when/if in 2026 the Refresh GPUs come out, using 3GB, instead of 2GB, GDDR7 chips and you get 50 % more VRAM in the same size.
- A dedicated GPU (4060 / 5060) will also have faster prompt processing (pp).
- The ability to partially or fully offload to the fast VRAM of the GPU (5060: 448 GB/s).
- A dedicated GPU adds additional capacity to the RAM.
- And, not LLM related, but: You can also game with higher FPS if you add a GPU that is faster than Strix Halo's iGPU (between RTX 4060 Laptop (=RTX 4050 desktop, which doesn't even exist, this is how bad it would be (en.wikipedia.org/wiki/GeForce_RTX_40_series)) and RTX 4070 Laptop (=RTX 4060 desktop)).
- And, if looks matter, there are many arguably better looking mini-ITX cases, too.
Love how silent it is (especially with the Noctua printed mod), yet powerful enough for stuff that my previous PC couldn't handle.
It's perfect for my HTPC setup. Really happy camper.
So odd to have a box this relatively big for nothing but a hefty cooler and PSU - there's no room for extra storage or cards, you get two m.2 slots and that's it.
Definitely not a great approach, despite Apple doing this all-soldered thing for years now.
Julian M, any Strix Halo implementations that do have the extra storage? Otherwise, this is maybe due to Strix Halo only having 16 PCIe gen 4 lanes, while in my mentioned comment above, building on a mini-ITX motherboard, would have 24 PCIe gen 4 lanes, which is an another advantage.
Doesn't it bother you that you installed the cooler BACKWARDS? THIS STICKER SHOULD BE ON THE BOTTOM, - THE FAN SHOULD PUSH AIR IN, NOT SUCK IT OUT
I don't underatand cons.
Cons
- smaller and lighter alternatives are available with the same CPU
- RZ717 WLAN does not support 320 MHz
- RAM not user-upgradeable
- no 10 Gbps RJ-45 option
Smaller and lighter means louder and hotter. I have one miniPC and it is too small and noisy. 2-3 cm bigger and the problems go away and it will still be a miniPC.
Does a miniPC really need such a wifi card? Wifi card should be optional.
RAM is not user upgradeable but it certainly has its advantages. I would consider it a disadvantage on a regular PC but not on a PC like this.
it has 5 Gbps. Is the chance that someone will need faster LAN connection?
Ok, but why/who would need 64 GB or 128 GB RAM all of a sudden?
The latest hype is AI/running LLMs. Download e.g. LM Studio (wrapper for llama.cpp) and try it out.
The difference to ChatGPT is that it is private (and, to be fair, not as good as the best that proprietary AI has to offer, but it's catching up quick (and if the open-weight LLMs are always ~1 year behind, does it matter? (the difference is actually getting closer: as per artificialanalysis.ai/?intelligence-tab=openWeights report)).
You can also download llama.cpp directly and the LLMs can be downloaded from huggingface.co.
Current SOTA model that fit into 64 GB RAM at q8 / 8-bit quant:
huggingface.co/unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF
huggingface.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF
Current SOTA models that fit into 128 GB RAM (or fit using a decent quant, not the full FP16) are e.g.:
huggingface.co/unsloth/GLM-4.5-Air-GGUF
huggingface.co/unsloth/gpt-oss-120b-GGUF (from the makers of ChatGPT)
Qwen3-Next-80B-A3B (github.com/ggml-org/llama.cpp issue: "Feature Request: Qwen3-Next support")
and many more, including thinking variants, like:
huggingface.co/Qwen/Qwen3-30B-A3B-Thinking-2507.
For open weights LLMs evaluations/benchmarks, see e.g. artificialanalysis.ai/?intelligence-tab=openWeights.
NOTEBOOKCHECK, maybe write an article to explain to your readers about the AI LLM self-hosting ability or why 64 GB RAM or even 128 GB RAM are a thing all of a sudden now, because normally, 32 GB RAM are enough for most people.
@Running AI LLMs locally:
Have you found a real use case for LLMs yet though? Besides basic chatgpt stuff like entering words in a prompt to summarize articles?
Ed Zitron, define "real use case". I don't summarize articles, but I found many use cases(, but nothing that either needs Strix Halo's 256 GB/s (practical 210 GB/s) aka the 2.5x token per second speed improvement vs just upgrading RAM in one's PC that one already has, which is much cheaper, or nothing, where Strix Halo's 128 GB RAM are enough, that's why I posted the (DIY) mini-ITX build suggestion above). A week ago OPENAI posted a LLM usage breakdown (image), see what people use LLMs for.
Quote from: Ed Zitron on September 22, 2025, 19:38:41Have you found a real use case for LLMs yet though? Besides basic chatgpt stuff like entering words in a prompt to summarize articles?
Running visual LLM locally to fully replace Google Photos with your own personal and private cloud.
Quotebuilding on a mini-ITX motherboard, would have 24 PCIe gen 4 lanes, which is an another advantage.
Just know that some non-normal desktop CPUs have only 16 PCIe lanes (like the 8400F), instead of the full 24.
Wonder if Notebookcheck team will re-review the temperature and noise parts, taking into consideration that they put the CPU fan upside down 🤦🏻 (a corrected picture would also be nice, so as not to confuse potential buyers).
It's not as if it's extremely clear on their instructions how it should be put together: https://guides.frame.work/Guide/CPU+Fan/489?lang=en
heffeque, I asked AIs (through lmarena.ai) and they say the temp increase is going to be from 5 to 20 °C.
Don't expect a fixed retest:
QuoteThe test sample was provided to the author as a loan by the manufacturer or retailer for the purpose of this review.