Show HN:通过 BF16 分块在 A40 上实现 1.499 倍加速和 28.5% 动态能耗降低
Show HN: 1.499x speedup and 28.5% dynamic energy drop on A40 via BF16 tiling

原始链接: https://cwtractors.com

Coldwater Tractors 报告在 NVIDIA A40(Ampere sm_86)上验证通过的 SKEW_TILED_BF16 协议。在 2D 7 点六边形 Ising 晶格基准测试中,L=4608 时加速比为 1.499 倍(10.204 毫秒 vs 15.291 毫秒),动态能耗降低 28.5%(186.735 vs 261.164 mJ/sweep),且六个子晶格相中站点不一致数为零。工作流采用冻结规格、多模型红队审计和双轨确定性验证;正确性门在 L=384 和 L=768 通过。bf16 内核使用每线程 15 个寄存器、2,584 字节共享内存,无溢出。L=9216 的更大运行显示 1.477 倍加速和 27.1% 能耗降低。硬件计数器被阻塞,因此 DRAM 降低为推断值。总体状态:验证通过。使用限制为民用、学术和商业研究;禁止军事或动能应用。

Hacker News 新 | 过去 | 评论 | 提问 | 展示 | 工作 | 提交 登录 展示 HN:通过 BF16 分块在 A40 上实现 1.499 倍加速与 28.5% 动态能耗下降 (cwtractors.com) 9 分,由 ktractor123 发布于 1 小时前 | 隐藏 | 过去 | 收藏 | 讨论 帮助 指南 | 常见问题 | 列表 | API | 安全 | 法律 | 申请 YC | 联系 搜索:
相关文章

原文
Coldwater Tractors Inc. | High-Performance Compute & Research

Raw GPU benchmark outputs and dense context logic boot files for multi-agent execution loops.

  • SKEW_TILED_BF16 — Verified NVIDIA A40 Receipts & Unified Logic Protocol NEW RECEIPT

    2D 7-point hexagonal Ising lattice (Ampere sm_86): 1.499x speedup at L=4608 (10.204 ms vs 15.291 ms), 28.5% dynamic energy reduction (186.735 vs 261.164 mJ/sweep), and 0 site disagreements across all 6 sub-lattice phases.

  • a40_environment_receipt.txt RECEIPT

    NVIDIA A40 hardware environment, driver/nvcc versions, ptxas register allocation (15 regs/thread for bf16_t), and SHA-256 verification receipt (2026-11-16).

  • a40_nsys_summary.txt PROFILE

    NVIDIA Nsight Systems (nsys) kernel execution and CUDA API timing summary on NVIDIA A40 (2026-11-16).

  • Quattra Kernel #1 Benchmark Receipt.md RECEIPT

    Official Empirical Performance & Energy Benchmark Receipt for skew_tiled vs Controls (NVIDIA A40, float32, 1.10x speedup at L=4096/8192).

  • tiled_results_2.txt BENCHMARK

    Quattra Tiled D4GEMV Kernel Benchmark Results (NVIDIA A40, L=16 to L=8192, 12.2% speedup verified pass).

  • tiled_results.txt BENCHMARK

    Quattra Tiled D4GEMV Kernel Benchmark Results (NVIDIA A40, L=16 to L=8192, initial run).

  • hexa_boot_1_physics.txt BOOT 1

    Hexa Dense Logic Boot - Physical Layer & Hardware Mapping.

  • hexa_boot_2_active.txt BOOT 2

    Hexa Dense Logic Boot - Active Execution Loop Packet.

  • hexa_boot_3_philosophy .txt BOOT 3

    Hexa Dense Logic Boot - Architectural Philosophy & Framework.

SKEW_TILED_BF16 — NVIDIA A40 Receipts & Unified Logic Protocol

Authors: Kristian Lindholm and the AI Orchestra (Coldwater Tractors Inc)
Date: November 16, 2026  |  Architecture: NVIDIA A40 (Ampere sm_86, 48 GB GDDR6, 300 W TDP)
Use Restriction: Receipts and methodology open for civilian, academic, and commercial research. Strictly prohibited for kinetic, military, weapons, or targeting applications.

Unified Logic: The Scientific Method for Intelligent Systems

Prior to compiling or renting GPU compute, every architectural and numerical claim was subjected to a multi-model adversarial falsification loop ("Red-Team Audit"):

1. Frozen Pre-Run Specification: Grid sizes (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.

2. Multi-Model Red-Team Audits: Independent model windows audited shared-memory bank conflicts, wrap-around alignment, multi-sweep pointer hazards, ULP tie drift, and observable bias—catching and resolving 15+ structural and verification edge cases prior to execution.

3. Dual-Track Deterministic Verification: Every run executes two full sweeps (6 sub-lattice color phases) against both an exact FP32 parity track and a double-precision (FP64 std::exp) reference across randomized (±1) and uniform (si = +1) initial states.

1. Compiler & Register Footprint (nvcc -O3 -arch=sm_86 --ptxas-options=-v)

Kernel Variant Registers / Thread Shared Memory Spills Gate
skew_tiled_phase<bf16_t, 38> 15 registers 2,584 bytes 0 bytes PASS
skew_tiled_phase<float, 37> 36 registers 5,032 bytes 0 bytes PASS
skew_tiled_phase<int8_t, 37> 36 registers 1,258 bytes 0 bytes PASS

2. Correctness & Detailed-Balance Gates (2 Sweeps = 6 Sub-Lattice Phases)

Grid & Initialization L2(h) / L2(dE) gpu_f32_diff dec_f64_rate |de|/|e| & |dm| Gate
L = 384 (Random ±1) 0.0e+00 / 0.0e+00 0 0.00e+00 0.00e+00 / 0.00e+00 PASS
L = 384 (Uniform +1) 0.0e+00 / 0.0e+00 0 0.00e+00 0.00e+00 / 0.00e+00 PASS
L = 768 (Random ±1) 0.0e+00 / 0.0e+00 0 0.00e+00 0.00e+00 / 0.00e+00 PASS
L = 768 (Uniform +1) 0.0e+00 / 0.0e+00 0 0.00e+00 0.00e+00 / 0.00e+00 PASS

3. Standalone cudaEvent Throughput (30 Interleaved Trials, 10 Sweeps)

Lattice Size (L × L) bf16_t Median (IQR) fp32 Median (IQR) int8_t Scalar Ref Speedup (fp32 / bf16)
4608 × 4608 (21.23M sites) 10.204 ms (0.005 ms) 15.291 ms (0.013 ms) 13.356 ms 1.499x
9216 × 9216 (84.93M sites) 40.679 ms (0.015 ms) 60.064 ms (0.020 ms) 53.072 ms 1.477x

4. Multi-Second Paired ABBA NVML Energy (16 Quads, P_idle = 73.02 W)

Lattice Size (L × L) Sweeps / Window E_dyn (fp32) E_dyn (bf16) Dynamic Energy Saved Total Board Saved
4608 × 4608 4,500 (~4.6–6.9 s) 261.164 mJ/sw 186.735 mJ/sw 28.5% (1.399x) ~29.9%
9216 × 9216 1,200 (~4.9–7.2 s) 1083.034 mJ/sw 788.288 mJ/sw 27.1% (1.371x) ~28.8%

5. Hardware Performance Counter Disclosure (ncu)

UNMETERED PMC   Container Hypervisor Restriction (ERR_NVGPUCTRPERM): Low-level hardware performance counters (ncu DRAM byte counters) were blocked by host container permissions on the benchmarked A40 instance. Accordingly, the expected ≥42% DRAM traffic reduction is architecturally inferred from the 2-byte (bf16_t) vs. 4-byte (float) representation and corroborated by the measured 1.499x wall-clock speedup and 28.5% NVML dynamic energy reduction, rather than directly metered via hardware PMC counters.

6. Raw Console Receipt

[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)
[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)
[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)
[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)
L=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
L=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
GPU: NVIDIA A40
Baseline P_idle = 73.02 W
[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)
[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)

NVIDIA A40 GemV Skewed-Tiling Verification

Status: VERIFIED PASS (3/3 large sizes with statistical separation)

Target Architecture: NVIDIA A40 (48GB GDDR6, PCIe)

Performance Gain: ~1.05x at L=2048, 1.119x at L=8192 over optimal control.

Audit Receipts & Files

// DIRECT CONTACT & INQUIRIES

For shop inquiries, technical documentation packets, or custom fabrication specs:

联系我们 contact @ memedata.com