logoalt Hacker News

20 points • by ktractor123 • last Tuesday at 1:31 AM • 0 comments • view on HN

Comments

ktractor123 • last Tuesday at 1:32 AM

I run a small custom tractor and loader fabrication shop in Washington, and lately I've been using a multi-model falsification loop ("Unified Logic" / AI Orchestra) where separate model windows adversarially red-team a CUDA design before spending any money on GPU compute.

For this test, we implemented a 2D 7-point hexagonal Ising lattice using BF16 skewed shared-memory tiling on an NVIDIA A40 (Ampere sm_86). Before compiling, separate model windows audited the spec for shared-memory bank conflicts, 3-color torus wrap alignment, multi-sweep pointer hazards, and ULP tie drift—catching 15+ structural edge cases before first silicon execution.

Verified A40 receipts (posted on the page with raw environment and nsys logs): - Register footprint: 15 regs/thread for bf16_t (0 spills) vs. 36 regs/thread for fp32 - Correctness: 0 site disagreements across all 6 sub-lattice phases against both FP32 parity and FP64 std::exp reference (random ±1 and uniform +1 ensembles at L=384 and L=768) - Throughput (30 interleaved trials): 10.204 ms vs. 15.291 ms at L=4608 (1.499x speedup); 40.679 ms vs. 60.064 ms at L=9216 (1.477x) - NVML dynamic energy (16 paired ABBA quads, P_idle = 73.02 W): 186.735 vs. 261.164 mJ/sweep at L=4608 (28.5% reduction); 788.288 vs. 1083.034 mJ/sweep at L=9216 (27.1% reduction)

One explicit caveat documented in the receipts: low-level hardware PMC counters (ncu) were restricted by the container hypervisor (ERR_NVGPUCTRPERM), so the DRAM byte reduction is inferred from the 2B vs. 4B slab representation and corroborated by the wall-time speedup and NVML power logs rather than metered directly via PMC.

Happy to answer questions about the skewed-tile shared-memory layout or the multi-model red-team protocol.

➕ show 1 reply