Available for pre-order for $300 and up, the computer consists of a 100 x 86mm (3.9″ x 3.4″) board that comes with an aluminum enclosure and built-in fan. The board supports up to 32GB of onboard LPDDR5-6400 memory, and supports up to 256GB of UFS storage as well as an M.2 2280 slot that can be used for a PCIE 3.0 x4 NVMe SSD, among other things.
Meh. Wake me up when they can beat a Raspberry Pi of a similar price level. By the way the $300 version has 8GB of ram. The 32GB version is $575.
I wonder what’s the point of this. They brag about AI performance, but then the thing has only 32GB of what seems to be non-expandable RAM. At this point you might as well run your AI on your regular PC.
This (or frankly any other single-board computer releasing this decade) absolutely isn’t something that’d make sense to use as an inference server that you access from a desktop that already has a GPU and plenty of memory.
It’s not even remotely in the same product category as GB10. More akin to a Jetson NX: a low-power device that can do a little local processing for contexts where it doesn’t have a reliable connection to a more powerful server, though even for that purpose it has lower performance than Orin. Probably would’ve been substantially cheaper than Orin if not for the RAM crisis.
I’d guess most of the actual users of this thing probably will ignore the AI cores entirely and just use it as a RISC-V development board. It’s the first hardware that supports RVA23, which is the set of ISA extensions that most distros will treat as the baseline requirement for RISC-V. Also the fastest RISC-V CPU to date, meaning faster code compile (if you can’t cross-compile) and automated tests than any other RISC-V system.
That’s why I don’t get the AI marketing angle here at all. Any somewhat knowledgeable user will not use it for AI, and people with no clue who might fall for that kind of marketing usually don’t host their own local models.
And still AI marketing for underpowered devices is everywhere. Even the ESP32-S3 is marketed with AI features. You know, the microcontroller that tops out at 16MB RAM.
A lot of the AI marketing you see for various devices might be meant to appease and interest investors rather than potential users. Or the result of some marketing person or executive being convinced that absolutely every product needs to have an AI angle.
Even so, not every “AI” workload needs to be running the biggest local LLMs you can at the highest speeds possible.
Robotics often need low-power chips that can handle computer vision, for example, and that’s a use-case that Nvidia Jetson seems to focus on, to great success, with only 8-16GB RAM. SpacemiT’s engineers seem to be aiming at that market considering they put K3 on modules that are pin-compatible with Jetson NX carrier boards, and gave it dedicated cores for matrix and vector processing.
Regardless whether they manage to succeed at all for AI uses, K3 is designed for that. Maybe it makes more sense for that domestically in China, where they would want drop-in substitutes for less reliance on Nvidia imports.
For anyone else, it’s a RISC-V devboard that manages to be faster than any others despite the low power budget and having so much of the silicon wasted on AI stuff instead of more general-purpose cores. It’s not really designed for general dev use but the alternatives just happen to be so much worse for now.
As with anything RISC V at the moment, it’s mostly for shits and giggles. Once it matures and hopefully gets actually performant it might be a nice alternative to ARM/x86. But for now it’s mostly a curiosity.
The fact that this has an sfp+ port makes it instantly magically interesting to me. Probably not gonna be practical, as risc-v generally isn’t yet, but I do love where this is going.
SpacemiT has only just started mainlining K3 features to Linux. This only makes sense to kernel hackers wanting to get a jump on the K3 at this point. There’s still plenty of kernel hacking to keep one busy with the K1 and K1X though.
deleted by creator
