RTX PRO 6000 Blackwell

Added 2026-09-06 as part of a Blackwell/RTX 50-series freshness check: this site already tracked the previous-generation workstation flagship (RTX 6000 Ada Generation) but not its current-generation successor. Only two figures are directly published on NVIDIA's own product page (nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-6000/): fp32TFLOPS (125, "Single-Precision Performance") and a single AI figure, "4,000 TOPS," footnoted "Theoretical FP4 TOPS using sparsity." Every other compute figure here is DERIVED from those two using the same precision-ladder relationship NVIDIA's own Blackwell architecture whitepaper establishes for the RTX 5090 (same GB202 die family, see that entry's notes): each step down in precision doubles dense throughput versus the step above (tf32 = fp32; bf16 = fp16 = 2x tf32; fp8 = 2x fp16; fp4 = 2x fp8), and each precision's sparse figure is 2x its own dense figure. Working forward from fp32=125 gives fp4 dense=2000 -> fp4 sparse=4000 -- which lands exactly on NVIDIA's own independently-published 4,000 TOPS figure, a strong cross-check that the assumed ratios hold for this die too, not just the 5090. int8TOPS/int8TOPSSparse are set equal to fp8TFLOPSSparse/2x that (1000/2000), mirroring the RTX 5090 entry where int8TOPS matches its FP8-sparse rate exactly. transistorCountBillion (92.2) and processNode ("TSMC 4N") are carried over from the RTX 5090 entry, not independently re-published for this SKU -- both cards use the same physical GB202 die (this one effectively unlocked to 188 of the die's SMs, vs. 170 enabled on the 5090), and transistor count/process are fixed properties of the die itself, not of which SMs are enabled. computeUnitCount (188 SMs), shaderCoreCount (24,064 CUDA cores), and matrixCoreCount (752 5th-gen Tensor Cores) come from third-party reporting (Tom's Hardware/TechPowerUp/VideoCardz pricing coverage), not a spec sheet on NVIDIA's own product page (which lists no core counts at all) -- treated as reliable because they're internally consistent with each other (24,064 / 128 CUDA cores per SM = 188 SMs exactly; 752 / 188 = 4.0 Tensor Cores per SM, the same ratio the RTX 5090's whitepaper-sourced 680/170 = 4.0 figure shows). No msrpUSD: NVIDIA never announced an official MSRP for this card -- early 2025 retailer/preorder listings for the boxed Workstation Edition ranged roughly $7,700-8,600, and NVIDIA has since raised its own Marketplace listing price twice without an announcement (to $13,250 in June 2026, then $16,000 in August 2026, per contemporaneous Tom's Hardware/VideoCardz/TechPowerUp coverage) -- picking any single number here would misrepresent an unusually volatile, vendor-unannounced pricing history as a stable list price. A distinct "RTX PRO 6000 Blackwell Server Edition" (passive cooling, for OEM servers) and a "Max-Q" variant (300W, blower-style cooler) also exist; this entry covers only the workstation/active-cooling edition described above, matching this collection's existing convention of one entry per primary SKU rather than every cooling/form-factor variant.