Moore Threads MTT S4000 product page - precisions and memory published, NO absolute compute figure published

VERBATIM 2026-10-04 via web_extract. Recorded because accelerators/moore-threads-s4000 has an EMPTY flops array and cites this page; this record establishes that the empty array is correct. MEMORY AND PRECISION, VERBATIM: 'MTT S4000 配备的 Tensor 核心算力、48GB 显存以及超高速卡间互连接口 MTLink，可以有效支持多种主流大语言模型训练' (48GB of memory, tensor-core compute, MTLink interconnect). And under '新一代 Tensor 核心': 'MTT S4000 支持 FP64、FP32、TF32、FP16、BF16、INT8 等主流精度算力' - the S4000 SUPPORTS FP64, FP32, TF32, FP16, BF16 and INT8 precision. Note carefully: this names the SUPPORTED PRECISIONS and attaches no number to any of them. INTERCONNECT, VERBATIM: 'MTLink 是基于 MTT S4000 自研的卡间互连技术... 卡间互连 I/O 带宽达到 240GB/s' - 240 GB/s interconnect I/O bandwidth. NEGATIVE FINDING - NO COMPUTE FIGURE: NO TFLOPS, NO TOPS, no numeric compute figure of any kind appears on this page. The 'Tensor 核心算力' phrase is unquantified. A full sweep of the extracted text found no absolute compute throughput number. The empty flops array on accelerators/moore-threads-s4000 is therefore CORRECT. SPARSITY/DENSE CONVENTION: no sparsity footnote, no 'with Structured Sparsity' title, no sparse/dense pair; the strings 'sparse'/'sparsity' do not occur. Moot given no figure is printed, and recorded so that a future reader does not assume a figure exists to be halved. CROSS-CHECK ON THE RECORD'S OTHER FIELDS: the page states 48GB of memory, which matches the stored vram_gb 48. It states NO memory type (the record's memory_type 'gddr6' is therefore NOT sourced from this page and should be marked as such or nulled), NO memory bandwidth (the stored memory_bandwidth_gbps 768 is not on this page), NO TDP (the stored tdp_w 450 is not on this page), and NO process node (stored process_nm is already null, correctly). Three of the record's stored numeric fields therefore rest on sources other than the cited vendor page.