NVIDIA RTX PRO 5000 72GB Blackwell Workstation Edition

NVIDIA product-page spec table prints 'Memory Bandwidth 1,344 GB/sec' on the shared row covering both the 48 GB and the 72 GB columns (source nv-rtx-pro-5000-blackwell-product-page, retrieved 2026-10-15). The linked datasheet PDF is image-compressed and was not text-extracted, so the bus width is NOT stored: Hardware Corner (Allan Witt, 2025-05-11) states a 384-bit interface, and 28 Gbps GDDR7 x 384 / 8 = 1,344 GB/s is arithmetically consistent, but that bus width is a secondary source and is left out of the field. Applied to the 72 GB column because the table prints one bandwidth cell spanning both SKUs, not because a separate 72 GB measurement was published.

Same NVIDIA product-page spec table as nvidia-rtx-pro-5000-blackwell, second memory column: '72 GB GDDR7 with ECC'. Bandwidth (1,344 GB/sec) and max power (300 W) are printed once on the shared row, so they are copied here and flagged as shared rather than independently measured for the 72 GB SKU. AI TOPS 2064 likewise has no per-SKU split. No public llama.cpp measurement for the 72 GB card was found in this pass — the Hardware Corner benchmark page covers the 48 GB card only. Do not apply the 48 GB tok/s table to this SKU: decode of a model that already fits in 48 GB should be bandwidth-bound and similar, but that is an inference, not a measurement.