AMD Radeon Pro VII archived product page - the GPU Specifications table, quoted in full, which supersedes a wrong negative finding about this same page

THE SAME DOCUMENT AS amd-radeon-pro-vii-product-page, RE-READ AND QUOTED because that source's earlier negative finding was wrong. This record exists so the spec table travels with the citation instead of being re-derived. THE TABLE, verbatim and contiguous: 'GPU Architecture Vega20 / Lithography TSMC 7nm FinFET / Stream Processors 3840 / Compute Units 60 / Peak Single Precision (FP32) Performance 13.1 TFLOPs / Peak Double Precision (FP64) Performance 6.5 TFLOPs / GPU Memory: Memory Size 16 GB, Memory Type (GPU) HBM2, Memory Interface 4096-bit, Memory Bandwidth Up to 1024 GB/s, Memory ECC Support Yes / Board Specifications: Form Factor PCIe Add-in Card, Bus Type PCIe 4.0 x16, TBP 250W, Cooling Active, Board Width Double Slot, Board Length 10.5" (267 mm), Board Height Full Height, External Power Connectors 1x PCIe 6-pin'. WHY THE EARLIER NEGATIVE FAILED. w4h-amd-radeon-pro-vii-product-page-dense reported zero matches for the literal tokens 'TFLOPS', 'TFLOP', 'TOPS', 'Peak ' and 'sparsit' across 18,346 characters. Re-extraction of the same URL yields 22,831 characters containing the quoted table above, with TFLOPs x2 and Peak x2. Two things went wrong and both are generalisable: the extract was shorter (18,346 vs 22,831 characters), so the earlier pull did not reach the specifications section at all; and a zero-match sweep for such common tokens is only informative if the same sweep is positive on a page known to contain them. A negative finding needs a control. THE SPARSITY ANSWER, which is the question the negative was opened to settle: 'sparsit' returns ZERO matches and 'Structured Sparsity' returns zero matches over the full extract. AMD's Radeon Pro VII page predates the Instinct 'with Structured Sparsity' labelling and carries NO sparsity footnote, so its single printed figure per precision is DENSE. Do not halve 13.1 or 6.5. This is a third convention alongside AMD's own 2x sparse labelling on Instinct and NVIDIA's footnote markers, and the only safe rule is per-page: quote the footnote, and where there is none, take the figure at face value. WHAT THE FP64 ROW ADDS. 'Peak Double Precision (FP64) Performance 6.5 TFLOPs' sits directly below the FP32 row. GCN5's FP64:FP32 ratio is 1:2, and 13.1/2 = 6.55, which is consistent with AMD's 6.5 at its stated precision - so the two figures corroborate each other and the card has no matrix cores of any kind, consistent with it being a gfx906 part. MEMORY. 'Memory Interface 4096-bit' and 'Memory Bandwidth Up to 1024 GB/s' are both first-party. Note the page states the bandwidth as a CEILING ('Up to'), not a typical or measured figure, and publishes no memory clock - so the 2.0 Gbps implied per-pin rate is a derivation from two printed numbers rather than a stated figure, and should be labelled as such wherever it is repeated.