cs.AR提交于 2026-10-09已译

Tensor Memory Equilibrium (TME) 模型缺失的第四项：残差解构代价

The Missing Fourth Term for the Emulation Tensor Memory Equilibrium (TME) Model: The Residue Deconstruction Cost

Harun Bayraktar · John Gunnels · Peter Caday

中文摘要

"FP8 is All You Need (Part 1)" 中的 Tensor-Memory Equilibrium (TME) 模型将 fp64 的 Ozaki Scheme II 模拟执行时间建模为张量核心项与高带宽内存 (HBM) 流量项的最大值，再加上每个输出的重构项。然而该模型遗漏了每个输入的解构代价：每个流式 fp64 操作数都必须在 SIMT 流水线上经过缩放、取整以及对 $r$ 个模数取模，才能发起矩阵乘法。本文补上这一第四项，依据 cuBLAS 模拟路径标定其常数，并推导出闭式运算强度阈值 $\mathrm{OI}^{*} = c_q r P_{\mathrm{fp64}}/(8P_{\mathrm{int}})$，低于该阈值时模拟性能无法匹配原生 fp64，无论张量核心吞吐多高。在 NVIDIA B300 GPU 上该阈值约为 $\mathrm{OI}^{*}\approx 0.56$ FLOP/B。结果显示，原始论文声称可加速的内存受限 kernel——GEMV、SpMV 和低 batch GEMV——仅能达到原生性能的 0.3-0.9 倍，7 点 stencil 仅为 1.8 倍而非声称的 3.1 倍。稠密 GEMM 不受影响，符合预期。还证明预计算并存储残差会把同一代价转移到带宽项中，并给出使该上界失效所需达到的指令数。

关键要点

01TME 模型遗漏了 fp64 模拟中每输入的解构代价（缩放、取整、模运算），需新增第四项
02依据 cuBLAS 模拟路径标定常数，推出运算强度阈值 OI* ≈ 0.56 FLOP/B（B300 GPU），低于此值无法匹配原生 fp64
03内存受限 kernel（GEMV、SpMV、低 batch GEMV）实际仅达原生 0.3-0.9 倍，7 点 stencil 仅为 1.8 倍而非声称的 3.1 倍
04稠密 GEMM 不受影响；预计算并存储残差可将解构代价转移至带宽项

解读

尚无解读。

原始英文摘要

arXiv:2610.10924v1 Announce Type: cross Abstract: The Tensor-Memory Equilibrium (TME) model of "FP8 is All You Need (Part 1)" calculates the execution time of Ozaki Scheme II emulation of fp64 as the maximum of a tensor-core term and a High-Bandwidth Memory (HBM) traffic term, plus a per-output reconstruction term. However, it omits the per-input deconstruction cost: every streamed fp64 operand must be scaled, rounded, and reduced modulo each of the $r$ moduli on SIMT pipes before any matrix multiply can issue. In this note we add this fourth term, calibrate its constant from the cuBLAS emulation path, and derive a closed-form operational-intensity threshold $\mathrm{OI}^{*} = c_q r P_{\mathrm{fp64}}/(8P_{\mathrm{int}})$ below which emulation cannot match native fp64 regardless of tensor-core throughput. On the NVIDIA B300 GPU the threshold is $\mathrm{OI}^{*}\approx 0.56$ FLOP/B. As a result, GEMV, SpMV, and low-batch GEMV, which are the memory-bound kernels the original paper claims to accelerate, are limited to 0.3-0.9x of native performance, and the 7-point stencil to 1.8x rather than the claimed 3.1x. Dense GEMM is unaffected as expected. We also show that precomputing and storing the residues moves the same cost into the bandwidth term, and we state the instruction count that an implementation would have to achieve to invalidate the bound.