An uncomfortable truth: our CUDA based AI stacks were built by people who optimize equations, not hardware. Data scientists know the math; systems engineers know the memory. When I instrument real workloads, about ~55% of wall-clock isn’t “doing AI” at all—it’s vanishing into avoidable data motion and sync stalls. AI isn’t slow because models are big; it’s slow because memory was treated like an afterthought.
Sporkle is my fourth attempt at solving the problem: treat the whole thing as a device-management problem first, and a compute problem second. Orchestrate memory, lifetimes, and transfers cleanly across devices with one API; let the math ride on top of a sane pipeline. When you center memory, the stack stops fighting you and performance shows up everywhere.
Why Fortran
I wrote it in a new, pretty Fortran dialect because array semantics and explicit shapes let the compiler generate ruthless code while keeping the memory model obvious. Think do concurrent, pure procedures, explicit-shape arrays, predictable strides, and ISO C interop when needed. It reads like math, behaves like a system, and deletes the pointer soup that makes “fast” code fragile and opaque.
The measurements
Receipts—before NVIDIA support even lands: production CPU paths hit ~90–160 GFLOPS; a straightforward OpenGL compute backend clears 400+ GFLOPS; the triple-buffered async pipeline (fence-gated) aggregates ~3,630 GFLOPS. Big kernels see ~6.5× latency cuts inside the pipeline; small, memory-bound kernels still gain ~2.3× throughput by keeping data hot and launches sane. Under load, the thread-safe program cache even shows “negative overhead,” hitting ~124.5% parallel efficiency thanks to lock-free reads and shared artifacts.
🧩 Today it runs on AMD, issuing real work without vendor SDK fairy dust; NVIDIA and Intel are next. The thesis is device-agnostic: fix data motion and the wins show up on anything with RAM and lanes.
🖥️ And because this matters: every number above was produced on my three-year-old Ryzen 7, all-AMD gaming rig. Not a datacenter box. No exotic hardware. Just disciplined systems engineering.
⚡ Why pay attention to this? Lower minimum footprint for useful AI. Speed without bigger boxes. Fewer watts per token/frame/sample. Reproducible harnesses, real traces, research-grade code. Fix the device layer and the math flies.
Currently in Alpha2