two new mainboards for framework 13!
RISCV - https://community.frame.work/t/dc-roma-risc-v-mainboard-iii-announced-by-deepcomputing/80698/7
ARM - https://community.frame.work/t/arm-based-cpus/7203/194
This is interesting and the two options seem reasonably comparable.
In rough terms, for general use, I'd say the RISC-V board is like a late 2000s Core 2 Quad but with twice as many cores.
The Arm board uses a CIX CP8180 with 8 A720 cores and 4 A520. The A720s seem to be in pairs clocking at most at 2.6, 2.5, 2.4, 2.3 GHz. I'm not sure if that means if you're using all 8 then all will be at 2.3 GHz.
The four A520s are running at 1.8 GHz.
All cores have 128-bit SVE/NEON.
They claim a theoretical 45 TOPS when combining NPU + GPU + CPU.
The RISC-V board using the SpacemiT K3 has 8 OoO X100 cores running at 2.4 GHz with 256 bit vectors. There are also 8 in-order A100 cores (derived from the X60 in the K1 chip) with 1024 bit vectors which provide the "AI" capability with a claimed 60 TOPS. Both types of cores can run exactly the same program executables, but due to the difference vector register length you can't migrate once the program is active.
K3 machines have not shipped yet but I and a few others have has ssh access to a machine at SpacemiT in China for the last month.
SpacemiT have provided a copy of llama that has been modified to do general processing on the X100 cores and running the model on the A100 cores. They claim good performance on models with 30B - 80B parameters.
One of us has a video on youtube showing the performance running an local Qwen3 30B model. Seems quite impressive for a 15W SoC running the model on standard CPU cores, not a specialised NPU or GPU.
Note: I put a "&t=590s" (9m50s) on the URL but playing it in the thread doesn't respect that, so jump there manually.