It's bad luck on your end. I have 3 AMD-based "desktops", never had a single problem with any of them. I just throw whatever memory in and it just works. These are being used more as servers than desktops, with large RAID arrays, HBA cards, tape drives, etc. They're consumer-level systems - Ryzen 7 5700G, Ryzen 5 2600, Ryzen 7 7800X3D.
Likes it
Using ASRock X570 PG 4S + Ryzen 5 2600 + Kingston 32GB 2666 ECC(cpu/mem support list for this mobo says the ECC works in this config). Dmidecode still reports 128bit data width instead of 72 however it also reports multi-bit correction instead of single-bit so ... ;). I'm kind of used to 72bit on Intel boards with UDIMM for example Supermicro+Xeon that I have. I think that a combination of memory controller plus mobo reporting has an effect on that info instead of actual hardware support - but still EDAC works, registers proper driver and I get the warnings once in a while from EDAC/RAS that a correctable error has been indeed corrected. Now if I should question whether it actually corrects something then I should also question the Supermicro+Xeon config - I don't have any means to check that - however if it's not working then I should see that on the ZFS dataset every month or so during scrubbing and I don't see anything there. So for me it is settled.
Likes it
Nice, main.exe seems to work just fine with the 7B quantized model - generates a token every 400ms on an AMD Ryzen 5 2600! But, quantize.exe doesn't seem to work - any valid command (such as below) pauses for a split second, then returns with no output? $ quantize.exe ggml-model-f16.bin ggml-model-q4_0.bin 2
Likes it
I have a Ryzen 2600 (CPU not APU) and ECC works. I wonder what generation ECC fuses became a thing?
Likes it
Many consumer Ryzen motherboards (and all CPU) support unbuffered ECC dram which is a great thing to have in a server. I use a Ryzen 2600 in a consumer tower. It's quiet, low power, and like 1/10 the cost of real server hardware.