OpenSolvers benchmarks and tunes open-source scientific software and AI inference engines on real RISC-V hardware — OpenBLAS, FlexiBLAS, EESSI, HPL, Quantum ESPRESSO, llama.cpp, and friends. The results and fixes stay public.
Today we work on what we can buy and keep on the desk: SiFive U74 (VisionFive 2), SpacemiT X60 / K1 (Orange Pi RV2, Banana Pi F3), SpacemiT K3 (Banana Pi SM10), and T-Head C910 (BeagleV-Ahead). That already uncovered real bugs (X60 gemv_n NaNs) and usable wins (U74 DGEMM, IME kernels, RVV backends, C910 xtheadvector GEMM, K3 IME2 int8 GEMM 103 GOP/s, K3 Q4_0 llama.cpp prefill 526 t/s). The same model at Q3_K_M is 499 t/s prefill. TinyLlama, whose matmuls are mostly Q3_K, is 254 t/s prefill. Larger llama models and the BXM GPU on the K3 are next.
Sponsorship is how we add platforms and finish the K3 story. Boards, shipping, power, and CI time are the bottleneck — not interest. Priorities:
- More K3 RAM / time — IME2 int8 GEMM, one Q4_0 llama.cpp model, the same model at Q3_K_M, and a TinyLlama Q3_K prefill are published; larger GGUFs and the BXM-4-64 GPU are still open. Extra RAM SKUs help large LLM / scientific runs: 16 GB ~$380–$400, 32 GB ~$575–$640 (Milk-V Jupiter 2 $575; Sipeed up to $639).
- T-Head C910 — the BeagleV-Ahead is on the desk. Draft RVV 0.7.1 (
xtheadvector) means the RVV 1.0 modules from the other boards do not apply. First local OpenBLAS numbers are up (DGEMM 7.42 GFLOP/s, 4 cores); small real/double GEMM shapes still fail CBLAS.
Your support buys those boards and the time to run the probes — then publish the notes here and in opensolvers/benchmarks.
Near-term Sponsors goal: fund K3 benchmark time + stretch toward a 16 GB (~$400) or 32 GB (~$600) SKU for larger AI inference runs.
For public questions, open an issue on opensolvers/benchmarks. Sponsors can also reach us privately via GitHub Sponsors Contact.
Tiers
| Monthly | Tier | What you get |
|---|---|---|
| $5 | Supporter | Sponsor badge on your GitHub profile |
| $25 | Backer | Badge + name listed below + priority issue triage on benchmarks |
| $100 | Sponsor | Logo + link below; listed in the benchmarks README |
| $250 | Gold | Homepage logo; mention in relevant YouTube descriptions; direct contact on what we benchmark next |
One-time options: $10 thank-you tip, $100 one-time sponsor (name or logo listed here for three months).
Gold sponsors
None yet — be the first.
Sponsors
None yet — be the first.
Backers
None yet — be the first.