Benchmarking llama.cpp on SpacemiT K3: RISC-V AI Cores vs Standard RVV (Part 4)
2026-03-12
Benchmarking llama.cpp on SpacemiT's K3 RISC-V chip with three models: X100 cores hit 76 t/s, A100 AI cores are 30x slower on pp until vendor IME2 code unlocks 111 t/s.
...
➦ First Words: LLM Inference on RISC-V
2026-03-11
Running LLM inference on RISC-V with the Python ML stack. Transformers generates text at 0.15-0.93 tok/s on a BananaPi F3, while llama.cpp hits 8.21 tok/s.
...
➦ The Dependency Rabbit Hole: Why 25 RISC-V Python Wheels Weren’t Enough
2026-03-11
A clean-machine pip install exposed stale riscv64 PyTorch wheels, missing system libraries like libopenblas, and transitive dependencies (numpy, grpcio) our PEP 503 wheel index missed.
...
➦ Building a Python Wheel Factory for RISC-V
2026-03-10
How I built a PEP 503 wheel index for riscv64 Python packages using two BananaPi F3 boards, GitHub Actions self-hosted runners, and 26 forked repos.
...
➦ Installing XCP-ng on Scaleway Elastic Metal: 30 Hours of Pain, 10 Minutes of Script
2026-03-05
Installing XCP-ng on Scaleway Elastic Metal the hard way: QEMU disk passthrough, Dell firmware bugs, Open vSwitch networking, and a script that does it all in 10 minutes. Part 2 of 3.
...
➦ Running 100 Wasm Containers on a Raspberry Pi (with 903 MB of RAM)
2026-03-01
100 WebAssembly containers on a Raspberry Pi 3B+ with 903 MB of RAM: 34 MB total, 164 KB per container. Measured every data point. Here's how Wasm compares to Docker on memory-constrained hardware.
...
➦ 