QuoteAccording to early leaked benchmarks, it appears that the Gorgon Halo APU is roughly 5% faster than its predecessor.
This is because Gorgon Halo is (just) a software-patched Strix Halo to support more than up to 128 GB RAM. All Ryzen AI 400 are a rename/refresh of the Ryzen AI 300 series.
Wish AMD would allow for 256 GB RAM straight away and not artificially limit the maximum RAM config..
On desktop B850 motherboards, up to 256 GB RAM are possible, and this is on a 128-bit memory interface (aka dual-channel). Strix/Gorgon Halo use a 256-bit memory interface (and faster RAM, too, btw.), so by this argument, up to 512 GB RAM should be possible on Strix/Gorgon Halo, and even more, because the memory speed is up to 2x faster vs desktop's 5600 MT/s, too, so, like easily 768 GB RAM then.
But, 256-384 GB LPDDR6 RAM is where Medusa Halo comes in later and its speed / memory bandwidth, and therefore also probably compute, increases are very welcomed.
That said, 192 GB RAM is nice, because it allows to fully fit huggingface.co/unsloth/DeepSeek-V4-Flash-0731-GGUF at its native weights (162 GB) + a decent amount of context on Linux, because on Windows on Strix Halo, last time I looked, less memory can be allocated than on Linux (maybe it's been fixed on Windows by now). So, on Windows it is: 192 GB-32 GB = 160 GB and on Linux it is 192 GB-8 GB = 184 GB. Or any other 4-bit quant of a ~300B parameters model, or a lower-end 4-bit quant of huggingface.co/unsloth/GLM-5.3-Flash-GGUF (IQ4_XS) (157 GB), tho, this is where 256 GB RAM would be nice, to be able to run UD-Q4_K_XL (200 GB).
QuoteGiven that the iGPU can access a whopping 160 GB of the system's total memory
I guess this confirms that it has not been fixed. So, know that on Linux you can allocate additional ~24 GB RAM. More infos here: reddit.com/r/LocalLLaMa and a general search would be something like "strix halo gtt linux". Why AMD has not fixed this on Windows, I don't know, because Windows doesn't need 32 GB RAM to run, 16 or even 8 GB RAM would have been fine. The additional 24 GB are so crucial to be able to run these ~300B parameters AI models at good 4-bit quants and context.
English
Español
Français
Italiano
Nederlands
Polski
Português
Русский
Türkçe
Svenska
Chinese
Magyar