do not use wifi connection for servers since this will lead to wifi card overheating. always use ethernet
Server grade CPU like AMD EPYC is capable of doing AI inference, and has a massive amount of RAM.
Keep your machine plugged, 24/7, even if it is just a Raspberry Pi.
Many times you face some problem like connection error, program unexpected exits, system freeze etc. Your purpose is to add more features to the system and never let it powered down.
If you cannot afford the power consumption, start with some small server, ramp it up for a long time.
Celeron MiniPC is very cheap. However SSD and large RAM are required for quick response.
You need OCuLink for eGPU connection.
When using smartphones and laptops as server, if no software level charging limit is supported, you need hardware switch controller, controlled by USB or HomeAssistant compatible protocols.
Read More
says that the cost of building some mini-itx with enormous ram costs a lot (more than te mac studio), and with a lot of headaches with custom designing, cnc and stl. worth it?
also the power consumption will be higher than mac studio.
assemble the machine with proper thermo, wires and some open space, then enclose the thing with design.
the weight:
mac studio (m1 ultra) for 3.6kg
mini-itx machine: 6.3 kg - 10 kg (won’t lose weight)
graphic cards
a100 80g
a16 64g (with self-designed thermal solution)
mother boards
Asrock X299e-itx using 4 mini ram slots
ASROCK X570D4I-2T
AsRock Rack E3C246D4I-2T Mini-ITX Server Motherboard Intel LGA 1151 C246, (max ram: 128gb)
ROMED4ID-2T for amd epyc 7002
asrock C422 WSI/IPMI
thermo solutions
generally you need to design case, thermo yourself, to make both side covered
use metal first. do not use plastic in case of overheating and burning
cpu
corsair h5 sf (hydra)
gpu
a100 thermo mod for hydra
Read More