⚙️ 3D Operator Fusion and the Roofline Model: Memory Traffic, Arithmetic Intensity and Kernel Launch Overhead
A Blender-built accelerator bench where a chain of neural-network operators moves tensors between a DRAM bank and a compute die: fuse or split kernels, change tensor size and precision, and watch the memory traffic, arithmetic intensity, launch overhead and roofline position change, verified by exhaustive plan enumeration.
⚙ Under the hood
A Blender-built accelerator bench where a chain of neural-network operators moves tensors between a DRAM bank and a compute die: fuse or split kernels, change tensor size and precision, and watch the memory traffic, arithmetic intensity, launch overhead and roofline position change, verified by exhaustive plan enumeration.
Operator FusionRoofline ModelArithmetic IntensityMemory BandwidthKernel Launch OverheadDeep Learning CompilersGPU PerformanceInference Optimisation
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install