Heads-up: links to Amazon on this page are affiliate links. If you buy through them we may earn a commission, at no extra cost to you. How we make money.
These are two different answers to the same problem. A big graphics card gives you a moderate amount of very fast memory. A Ryzen AI Max+ 395 mini PC gives you a lot of slower memory. Which is better depends entirely on the size of the models you want to run.
The numbers
RTX 5090
Radeon AI PRO R9700
Ryzen AI Max+ 395 mini PC
Memory for models
32 GB
32 GB
up to 96 GB of 128 GB
Memory bandwidth
1,792 GB/s
640 GB/s
256 GB/s
What it is
A card for a desktop PC
A card for a desktop PC
A complete small computer
Software
CUDA — everything works
ROCm / Vulkan
ROCm / Vulkan
Power
575 W card, 1,000 W PSU (NVIDIA)
300 W card
See each maker’s listing
Why memory speed matters so much for chat models
To write each word, a chat model reads through its whole set of weights. So the speed at which the chip reads memory sets a ceiling on how fast words come out. With the same model, a chip with seven times the memory bandwidth has a far higher ceiling. That’s the 5090’s advantage.
The mini PC’s advantage is size: with 96 GB for graphics, it can load models that simply do not fit in 32 GB. AMD’s own example is a 109B-parameter model (17B active) at 4-bit, which needs about 96 GB. On a 32 GB card that model doesn’t run at all.
“17B active” is the key detail. In mixture-of-experts models only part of the model is read for each word, so they are much kinder to slower memory. That makes them the natural fit for this kind of machine. Big dense models, where every weight is read every time, will fit but answer slowly.
Who should get which
You also want images and voice: the NVIDIA card. Most image and speech tools are built for CUDA first — faster-whisper’s GPU mode, for example, requires NVIDIA’s libraries.
You want the biggest chat models and can live with slower answers: the mini PC.
You want 32 GB for chat models and already have a desktop: the R9700 is the other route to 32 GB.
You don’t have a desktop PC at all: the mini PC is a whole computer; a card needs a PC with a big enough power supply around it.
At a glance — maker specs; tap a name for the full pick
96 GB — off this scale128 GB shared; AMD lets you assign up to 96 GB to graphics
Graphics memory
Up to 96 GB of the 128 GB LPDDR5X (AMD)
Memory bandwidth
256 GB/s
AI software
ROCm / Vulkan (AMD): fine for LLMs, rougher for image and voice tools
Chip
AMD Ryzen AI Max+ 395 · Radeon 8060S graphics
Memory
128 GB LPDDR5X, shared by CPU and graphics
Storage
2 TB NVMe
Runs models far beyond any gaming card, slowly for dense models and reasonably for mixture-of-experts ones. Best on Linux, where Ollama supports this chip through ROCm.
