The other day while considering getting a Mini PC to be my primary computer, I did intensive research (hint: conversations with AI), to determine how it'll perform compared my 5-years' old primary laptop which is low-end, and my gaming device, the Steam Deck.
While I didn't have a concrete Mini PC in mind, I had a minimum specs and guidelines to help me choose:
- It must be AMD.
- It must have 32GB of RAM or more.
- It must have 1TB SSD storage or better.
- It must be Zen 4.
With these specs I guarantee that the device is able to improve on what I already had, for example the 32GB RAM + Zen 4 combination should improve my local AI options exponentially.
On the lower end of the Mini PCs that fulfills my requirement was Minisforum UM760 with its AMD 7640HS (with 760M iGPU) APU. It would be my option, if I wasn't able to get a deal on the better 7845HS or 8645HS.
So, I researched how it'd do in my purposes: AI and Gaming.
Since I wanted to play games on this I researched the gaming performance. All Zen 4 devices I mentioned above have RDNA 3 iGPUs, but they vary in raw power. I especially researched games that my Steam Deck couldn't handle, like Final Fantasy VII Rebirth. This game should run perfectly on a Ryzen 7 with 780M iGPU, but I wasn't sure how Minisforum UM760 would handle it.
According to my research, it's the RAM that would be the bottleneck for this particular game, and that's also the case for Hogwarts Legacy. They won't get massive performance increase like other games. If not, the Mini PC should handle most games I'm thinking of playing no problem...
@ahmadmanga/re-leothreads-2b6hkomnu
While there are a bunch of videos exploring gaming benchmarks and the 7640HS processor's stats on them, the best performance indicator I saw is this video about FFVII Rebirth on various desktop APUs. It should be somwhere between the 8500G (740M) and 7600G (760M) and it seems the 7640HS is only limited by its smaller TDP compared to the desktop APU.
As for AI usage... The device should be able to run any model of 12GB or less at a good speed, that includes models of 30B with heavy quantization. Theoretically, it should be able to run 70B models at 1 token/s, but that's not usable for me.
I'm banking on the hope than Small Language Models of 14B or less becoming much more robust than they are now. For example, Ministral 14B is supposedly better than last year's 70B models despite having 80% less parameters.
What do you think?
Of course, I also researched the performance of Ryzen 7 7845 with the better 780M GPU, and it's way more impressive than the 7640. But as it's usually $100 more expensive, I wanted to have a better understanding of the minimum I could get.
- Cover image is generated with Qwen-Image on Venice.ai.