LLM

Quotelow iGPU performance

This means slow LLM prompt processing.

Quote17.1 pt

Strix Halo is 97 pt, or 467%, or 5.67 times, faster. The iGPU in this PC is extremely slow: If your LLM input is large, you will have to wait a veeeery long.

Up to 256 GB RAM is nice, it allows to run some nice (quants of) open-weight LLMs, like:

• huggingface.co/unsloth/GLM-4.7-GGUF
• 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
• huggingface.co/unsloth/DeepSeek-V3.1-GGUF

Quote67904 MB/s

So it's a 128-bit (2 * 64-bit per channel, aka dual-channel) system. Strix Halo is 256-bit / quad-channel, but only goes up to 128 GB RAM. The iGPU is Strix Halo is still much faster.

Looks promising.
I'd like to have details about pcie lanes because afaik the 285hx has 24 but with all the ports included it's not sufficient.
How many lanes per m.2 slot, pcie port,...?
Is bifurcation supported and on which port ?
What choices made minisforum ?
 

I don't know why nbc doesn't do the consumer a favour and just state in every article that the fast memory of kits like the 395+ provide a compelling advantage. AIDA (memory copy speed), "NPU" (the AMD has none but is still much faster than the NPU enabled ms-02), power usage anything with a 395+ is much better than this kit with its slow, slow RAM.