The material documents a BF16 tiled CUDA kernel for updating a 2D, seven-point hexagonal Ising lattice on an NVIDIA A40. It presents a pre-run specification, independent adversarial reviews of implementation risks, compiler and register data, and deterministic checks against FP32 and double-precision references. Across two sweeps and six sub-lattice phases, tests at lattice sizes 384 and 768 report no site disagreements or measured differences for randomized and uniform initial states. The kernel uses 15 registers per thread in its listed BF16 configuration, with no register spills.
Performance tests report that BF16 tiling cuts sweep time by 1.499× at 4608² sites and 1.477× at 9216², compared with FP32. NVML measurements show dynamic energy reductions of 28.5% and 27.1%, respectively; total board energy savings are estimated at about 30% and 29%. The explanation is that BF16 halves the data representation from four bytes to two, reducing expected memory traffic, but container permissions blocked direct DRAM-counter measurements, so that reduction is inferred rather than directly confirmed. The site also includes related A40 GEMV benchmarks, environment and profiling receipts, and files describing a broader “dense logic” framework.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.