{"slug": "show-hn-1-499x-speedup-and-28-5-dynamic-energy-drop-on-a40-via-bf16-tiling", "title": "Show HN: 1.499x speedup and 28.5% dynamic energy drop on A40 via BF16 tiling", "summary": "A BF16-tiled kernel on the NVIDIA A40 delivered a 1.499x speedup over FP32 and cut dynamic energy by 28.5%, according to benchmark results published by Kristian Lindholm and the AI Orchestra (Coldwater Tractors Inc) on November 16, 2026. On a 4608 × 4608 lattice (21.23M sites), the BF16 kernel ran a median 10.204 ms versus 15.291 ms for FP32, while dynamic energy fell from 261.164 mJ per sweep to 186.735 mJ per sweep; on a 9216 × 9216 lattice (84.93M sites) the speedup was 1.477x and the energy saving 27.1%. The authors locked correctness gates before execution and reported zero deviation from an FP64 reference across randomized and uniform initial states at L=384 and L=768.", "body_md": "Raw GPU benchmark outputs and dense context logic boot files for multi-agent execution loops.\n\n`sm_86`): `nsys`) kernel execution and CUDA API timing summary on NVIDIA A40 (2026-11-16).\n**Authors:** Kristian Lindholm and the AI Orchestra (Coldwater Tractors Inc)\n\n**Date:** November 16, 2026  |  **Architecture:** NVIDIA A40 (Ampere `sm_86`, 48 GB GDDR6, 300 W TDP)\n\n**Use Restriction:** Receipts and methodology open for civilian, academic, and commercial research. Strictly prohibited for kinetic, military, weapons, or targeting applications.\n      \n\nPrior to compiling or renting GPU compute, every architectural and numerical claim was subjected to a multi-model adversarial falsification loop (\"Red-Team Audit\"):\n\n`L ∈ {384, 768, 4608, 9216}`; updated from `4096/8192` via Patch 5 so `L` is divisible by `3 × TILE = 96` for seamless 3-color torus wrapping), physics parameters (2D 7-point hexagonal Ising lattice, `J = +1`, `T = 3.0`, stateless PCG32 PRNG), and pass/fail gates were locked before first execution.`FP64 std::exp`) reference across randomized (±1) and uniform (` s`<sub>i</sub> = +1) initial states.\n      `nvcc -O3 -arch=sm_86 --ptxas-options=-v`)\n| Kernel Variant | Registers / Thread | Shared Memory | Spills | Gate | \n|---|---|---|---|---|\n| `skew_tiled_phase<bf16_t, 38>` | **15 registers** | **2,584 bytes** | 0 bytes | PASS | \n| `skew_tiled_phase<float, 37>` | 36 registers | 5,032 bytes | 0 bytes | PASS | \n| `skew_tiled_phase<int8_t, 37>` | 36 registers | 1,258 bytes | 0 bytes | PASS | \n\n| Grid & Initialization | `L2(h)` /`L2(dE)` | `gpu_f32_diff` | `dec_f64_rate` | `\\|de\\|/\\|e\\|` &`\\|dm\\|` | Gate | \n|---|---|---|---|---|---|\n| `L = 384` (Random ±1) | `0.0e+00` /`0.0e+00` | `0` | `0.00e+00` | `0.00e+00` /`0.00e+00` | PASS | \n| `L = 384` (Uniform +1) | `0.0e+00` /`0.0e+00` | `0` | `0.00e+00` | `0.00e+00` /`0.00e+00` | PASS | \n| `L = 768` (Random ±1) | `0.0e+00` /`0.0e+00` | `0` | `0.00e+00` | `0.00e+00` /`0.00e+00` | PASS | \n| `L = 768` (Uniform +1) | `0.0e+00` /`0.0e+00` | `0` | `0.00e+00` | `0.00e+00` /`0.00e+00` | PASS | \n\n`cudaEvent` Throughput (30 Interleaved Trials, 10 Sweeps)\n| Lattice Size (L × L) | `bf16_t` Median (IQR) | `fp32` Median (IQR) | `int8_t` Scalar Ref | Speedup ( `fp32 / bf16` ) | \n|---|---|---|---|---|\n| **4608 × 4608** (21.23M sites) | **10.204 ms** (0.005 ms) | 15.291 ms (0.013 ms) | 13.356 ms | **1.499x** | \n| **9216 × 9216** (84.93M sites) | **40.679 ms** (0.015 ms) | 60.064 ms (0.020 ms) | 53.072 ms | **1.477x** | \n\n`P_idle = 73.02 W`)\n| Lattice Size (L × L) | Sweeps / Window | `E_dyn` (`fp32` ) | `E_dyn` (`bf16` ) | Dynamic Energy Saved | Total Board Saved | \n|---|---|---|---|---|---|\n| **4608 × 4608** | 4,500 (~4.6–6.9 s) | 261.164 mJ/sw | **186.735 mJ/sw** | **28.5%** (1.399x) | ~29.9% | \n| **9216 × 9216** | 1,200 (~4.9–7.2 s) | 1083.034 mJ/sw | **788.288 mJ/sw** | **27.1%** (1.371x) | ~28.8% | \n\n`ncu`)` ERR_NVGPUCTRPERM`):` bf16_t`) vs. 4-byte (` float`) representation and corroborated by the measured \n\n```\n[Gates 1-3 (rand_init, 2-sw) @ L=384] L2(h)=0.0e+00 L2(dE)=0.0e+00 gpu_f32_diff=0 dec_f64_rate=0.00e+00 |de|/|e|=0.00e+00 |dm|=0.00e+00 (|m|=0.0043)\n[Gates 1-3 (unif_init, 2-sw) @ L=384] L2(h)=0.0e+00 L2(dE)=0.0e+00 gpu_f32_diff=0 dec_f64_rate=0.00e+00 |de|/|e|=0.00e+00 |dm|=0.00e+00 (|m|=0.9400)\n[Gates 1-3 (rand_init, 2-sw) @ L=768] L2(h)=0.0e+00 L2(dE)=0.0e+00 gpu_f32_diff=0 dec_f64_rate=0.00e+00 |de|/|e|=0.00e+00 |dm|=0.00e+00 (|m|=0.0064)\n[Gates 1-3 (unif_init, 2-sw) @ L=768] L2(h)=0.0e+00 L2(dE)=0.0e+00 gpu_f32_diff=0 dec_f64_rate=0.00e+00 |de|/|e|=0.00e+00 |dm|=0.00e+00 (|m|=0.9418)\nL=4608 sweeps=10 | bf16 med=10.204ms (IQR 0.005) | fp32 med=15.291ms (IQR 0.013) | int8 med=13.356ms | speedup(fp32/bf16)=1.499x\nL=9216 sweeps=10 | bf16 med=40.679ms (IQR 0.015) | fp32 med=60.064ms (IQR 0.020) | int8 med=53.072ms | speedup(fp32/bf16)=1.477x\nGPU: NVIDIA A40\nBaseline P_idle = 73.02 W\n[NVML L=4608 sweeps=4500] E_dyn_fp32=261.164 mJ/sw | E_dyn_bf16=186.735 mJ/sw | Energy ratio=1.399x (28.5% reduction)\n[NVML L=9216 sweeps=1200] E_dyn_fp32=1083.034 mJ/sw | E_dyn_bf16=788.288 mJ/sw | Energy ratio=1.371x (27.1% reduction)\n```\n\n**Status:** VERIFIED PASS (3/3 large sizes with statistical separation)\n\n**Target Architecture:** NVIDIA A40 (48GB GDDR6, PCIe)\n\n**Performance Gain:** ~1.05x at L=2048, 1.119x at L=8192 over optimal control.\n\nFor shop inquiries, technical documentation packets, or custom fabrication specs:", "url": "https://wpnews.pro/news/show-hn-1-499x-speedup-and-28-5-dynamic-energy-drop-on-a40-via-bf16-tiling", "canonical_source": "https://cwtractors.com", "published_at": "2026-11-17 01:31:35+00:00", "updated_at": "2026-11-19 21:43:01.978559+00:00", "lang": "en", "topics": ["ai-infrastructure", "machine-learning", "ai-research"], "entities": ["NVIDIA A40", "Kristian Lindholm", "AI Orchestra", "Coldwater Tractors Inc", "NVIDIA", "CUDA", "BF16"], "also_reported_by": [], "takeaway": "Storing lattice spins as BF16 instead of FP32 cut register use to 15 per thread and 2,584 bytes of shared memory, yielding a 1.499x speedup and 28.5% lower dynamic energy at L=4608 with zero deviation from the FP64 reference.", "recommendation": {"advice": "Switch memory-bound kernel state from FP32 to BF16 tiling when a correctness gate against an FP64 reference passes; the source measured 1.499x speedup and 28.5% dynamic energy savings on an A40.", "playbook": "inference-serving", "playbook_url": "https://wpnews.pro/playbooks/inference-serving"}, "alternates": {"html": "https://wpnews.pro/news/show-hn-1-499x-speedup-and-28-5-dynamic-energy-drop-on-a40-via-bf16-tiling", "markdown": "https://wpnews.pro/news/show-hn-1-499x-speedup-and-28-5-dynamic-energy-drop-on-a40-via-bf16-tiling.md", "text": "https://wpnews.pro/news/show-hn-1-499x-speedup-and-28-5-dynamic-energy-drop-on-a40-via-bf16-tiling.txt", "jsonld": "https://wpnews.pro/news/show-hn-1-499x-speedup-and-28-5-dynamic-energy-drop-on-a40-via-bf16-tiling.jsonld"}}