Moore Threads MTT S5000 (PH100, Pinghu architecture) product page and architecture page - precisions named, NO absolute compute figure published

VERBATIM 2026-10-04 via web_extract of BOTH the S5000 product page and the companion Pinghu architecture page (linked as '了解 平湖 架构'). Recorded because accelerators/moore-threads-s5000 carries a single fp8 row of 989 TFLOPS whose own basis_detail concedes it is Moore Threads' footnoted description of an unnamed comparison TARGET rather than of the S5000. This source establishes that no absolute S5000 compute figure exists on the vendor site at all. S5000 PAGE, VERBATIM: 'MTT S5000 是一款面向生成式 AI 时代，专为大模型训练、推理及高性能计算而生的全功能 GPU 智算卡。其核心搭载了摩尔线程全新一代 PH100 芯片，凭借先进的 平湖 架构，提供从 FP8 到 FP64 的全精度算力支持。' (the card is built on the new PH100 chip and provides full-precision compute support from FP8 to FP64). Also: 'PH100 芯片已首批通过中国信息安全测评中心的安全可靠测评（2026年第2号）' (the PH100 has passed the China Information Security Evaluation Center reliability certification). 'FP8 原生算力' section, verbatim: 'MTT S5000 率先实现硬件级原生 FP8 计算精度支持... 针对大规模集群面临的带宽与算力瓶颈，MTT S5000 的 FP8 引擎全面支持 DeepSeek、Qwen 等前沿架构，训练性能提升超 30%' (hardware-native FP8; the FP8 engine gives over 30% training improvement). '原生 FP64 双精度算力' section, verbatim: '具备从 FP8 到 FP64 的全精度处理能力，确保科学模拟中极高的数据精度需求，为精密计算提供可靠保障'. PERFORMANCE CLAIMS THAT ARE RELATIVE, NOT ABSOLUTE - the complete set found on the page, all quoted: '在主流模型实测：在计算机视觉 (CV) 及大语言模型 (LLM) 核心训练场景中，综合性能达到国际旗舰产品① 的 74%以上' (over 74% of an international flagship product in CV and LLM training); 'Llama3-70B 训练：模型算力利用率 (MFU) 突破 60%'; 'DeepSeek-236B 训练：模型算力利用率 (MFU) 超过 40%'; '在超大规模 AI 模型训练中，我们实现了万卡级 GPU 集群训练线性度高达 95%'; '在 16k 长序列输入测试中，MTT S5000 单卡 Prefill 吞吐量是国际旗舰产品② 的 2.5 倍'; '在 2 节点 16 卡配置下，MTT S5000 实现了高达 61.83 samples/s 的训练吞吐量，模型算力利用率（MFU）达到 51%'; '单机性能达到国际旗舰产品① 的 64%-79%'; '在 DeepSeek-236B 万卡集群训练中，在前 3 万步训练过程中，Loss 曲线与国际旗舰产品① 集群的相对精度误差仅为 0.6%'. THE FOOTNOTE-MARKED COMPARISON TARGET, WHICH IS THE ORIGIN OF THE 989 TFLOPS ROW: the page uses the markers '国际旗舰产品①' and '国际旗舰产品②' - 'international flagship product 1' and '2' - and the footnote text defining those markers is rendered in the page's images and footnote block, not in the extracted text. The 989 TFLOPS figure is the density of that undefined comparator. No page on mthreads.com states an absolute TFLOPS figure for the S5000 or the PH100. SPARSITY/DENSE CONVENTION: NO sparsity footnote, NO 'with Structured Sparsity' title, NO sparse/dense pair, and the strings 'sparse'/'sparsity' do not occur. Since no absolute compute figure is printed, the convention is moot, but it is recorded that the vendor offers no basis on which any figure could be halved or doubled. PINGHU ARCHITECTURE PAGE, VERBATIM - the absolute figures that DO exist, all percentages: 'FP8 GEMM 算力利用率高达 90% 以上，比国际旗舰产品高出 15 个百分点'; 'Flash Attention 的效率高达 95% 以上，比国际旗舰产品高出 15-20 百分点'; '在矩阵乘累加 (MMA) 方面实现了 5 倍性能提升，并首次引入了原生 FP8 精度支持'; 'FP8 计算中间累加精度尾数 24bit，远高于国际旗舰产品的尾数 14bit'. CONCLUSION FOR THE REPO: accelerators/moore-threads-s5000's empty-ish state (one comparator-derived fp8 row, everything else null) is the correct reading of the vendor's own disclosures. The stored fp8 989 TFLOPS with vendor_claim false is correctly flagged as not the S5000's own figure. No correction is warranted; this source simply makes that judgement auditable.