• mlg@lemmy.world
    link
    fedilink
    English
    arrow-up
    4
    ·
    20 hours ago

    There’s the MTT S80 (First PCIe Gen 5 GPU lol) which is the consumer grade version of Moore Thread’s enterprise GPUs like S4000, but the problem is that they trade off super cheap VRAM and PCIe bandwidth for low compute power compared to even antiquated stuff from Intel, AMD, and Nvidia.

    They’re actually a great choice if you want to run AI/LLM stuff for really cheap, and Moore threads has their own CUDA knockoff called MUSA which iirc does have support in the various LLM backends available. Back when they released, it was going for something like $160 in China and ~$200-250 online. Could easily pool the VRAM, though finding a mobo+CPU combo with enough PCIe lanes to spare meant you’d most likely not be taking advantage of more than maybe 2 or 3 cards in one tensor parallel split.

    China’s domestic processor production is still catching up, so even though they have access to high speed RAM and all the latest standards, they don’t have the cores to match.

    Their last KX7000 x86 CPU was comparable to a skylake i5 or i7, but just with newer standards like DDR5 and PCIe gen 4. So they’re about 7 years behind based on that estimate.