A block-based language expresses computation over whole tensor tiles instead of individual scalar GPU threads. The compiler maps these tile operations onto hardware, combining Python-level productivity with performance close to hand-tuned CUDA. section 1.1.1
The model changes GPU implementation in several practical ways:
- Operate on blocked, multidimensional arrays rather than interleaved scalar-thread instructions. section 1.1.1
- Let the compiler map tile operations to GPU hardware and generate target-specific low-level representations. section 1.1.1
- Use one kernel source across NVIDIA, AMD through ROCm, and Intel through the XPU backend. section 1.1.1
