AWS Inferentia product page (aws.amazon.com) - the 190 TFLOPS FP16 per-chip sentence, quoted with AWS's own precision wording

VERBATIM 2026-10-04 via web_extract. Recorded so that accelerators/aws-inferentia2's fp16 row has a citable vendor sentence rather than a product-page pointer, and specifically to record the EXACT wording of the per-chip compute figure, because the wording contains a precision name that matters. THE COMPUTE SENTENCE, VERBATIM AND IN FULL: 'Each Inferentia2 chip has two second-generation NeuronCores, and each EC2 Inf2 instance has up to 12 Inferentia2 chips. Each Inferentia2 chip supports up to 190 tera floating operations per second (TFLOPS) of FP1[6]...' - the extracted text is cut mid-token at 'FP1', which is a rendering artefact of the superscript, not a vendor ambiguity. The precision is stated as a numbered footnote marker and AWS publishes that footnote elsewhere on the page. COMPARATIVE CLAIMS, VERBATIM: 'AWS Inferentia2 chip delivers up to 4x higher throughput and up to 10x lower latency compared to Inferentia.' and 'Inferentia2-based Amazon EC2 Inf2 instances to support scale-out distributed inference with ultra-high-speed connectivity between chips.' MEMORY, VERBATIM: 'The first-generation Inferentia has 8 GB of DDR4 memory per chip and also features a large amount of on-chip memory. Inferentia2 offers 32 GB of HBM per chip, increasing the total memory by 4x and mem[ory bandwidth by 10x]'. Note AWS says '32 GB of HBM per chip' and does NOT name a bandwidth figure per chip; the record's 820 GB/s is a derivation from the Inf2 instance aggregate (12 chips, 384 GB, 9.8 TB/s), which the record's memory_bandwidth_basis already states correctly. SPARSITY/DENSE CONVENTION - VERBATIM NEGATIVE: the page carries NO sparsity footnote, no 'with Structured Sparsity' row title, no sparse/dense pair and no dense-or-sparse marker. AWS uses NEITHER the NVIDIA halving convention NOR the AMD explicit-labelling convention; it simply does not discuss sparsity on this page. The 190 TFLOPS figure must therefore be recorded as published with no halving applied, because no document exists that would license halving it. CONSEQUENCE FOR accelerators/aws-inferentia2: its stored fp16 190 TFLOPS dense:true is consistent with this page. Its fp8 380 TFLOPS row is NOT - the page publishes no FP8 figure at all, and that row is a 2x inference from the FP16 number, as its own note concedes. That inference is reasonable but it is not vendor-published, and the row's basis_detail should say so rather than citing the page as though it did.