NVIDIA RTX PRO 6000 Blackwell specs: Workstation Edition, Max-Q and Server Edition

The RTX PRO 6000 Blackwell family on one sheet

The NVIDIA RTX PRO 6000 Blackwell comes in three editions with the same GPU configuration: 24,064 CUDA cores, 96 GB of GDDR7 with ECC on a 512-bit bus and a PCIe 5.0 x16 interface. They differ in memory bandwidth (1,792 GB/s on the Workstation Edition and the Max-Q, 1,597 GB/s on the Server Edition), in board power (600 W, 300 W, up to 600 W) and in cooling. This page collects every figure NVIDIA publishes for the three cards, with the document each one comes from.

NVIDIA spreads the specifications over three product pages, three datasheets, a family page and a product brief for the Server Edition, and the documents do not always agree. Where two NVIDIA sources give different numbers for the same card, both are shown, with the one we plan with: the lower figure for compute, the higher figure for power. Choosing between the editions for a given machine is a separate question, answered in our comparison of the Server Edition, Workstation Edition and Max-Q.

NVIDIA datasheets 5349469 (Workstation Edition, June 2026), 5349650 (Max-Q, June 2026) and 4682150 (Server Edition, December 2025, the version linked from NVIDIA's product page); NVIDIA product pages and RTX PRO 6000 family page; compute capability from NVIDIA's CUDA GPU list, page dated 16 October 2026; all read 24 October 2026.

Edition by edition: compute, bandwidth and power

The table below holds the figures that separate the editions. NVIDIA marks the workstation AI TOPS as "Effective FP4 TOPS with sparsity" in both datasheets. The Server Edition page lists 4 PFLOPS of FP4, 2 PFLOPS of FP8, 1 PFLOP of FP16 and BF16 and 234 TFLOPS of TF32 with no footnote on sparsity, and its datasheet gives only "Peak FP4 AI: 4 PFLOPS".

NVIDIA product pages (Workstation Edition updated 2 October 2026, Max-Q 22 July 2026, Server Edition 10 March 2026) and datasheets 5349469, 5349650 and 4682150, read 24 October 2026. Peak rates are at GPU Boost Clock, as NVIDIA footnotes them.

The Workstation Edition's product page and its June 2026 datasheet differ by about one per cent, 125 against 126 TFLOPS of FP32 and 380 against 382 RT TFLOPS; the Max-Q figures match in both documents. We plan with the lower product-page figures, 125 and 380; our sizing rests on memory and bandwidth, which both documents give identically. In millimetres the cards measure about 137 × 305 mm (Workstation Edition) and 112 × 267 mm (Max-Q and Server Edition), by our conversion.

The Server Edition datasheet is the only one of the three that counts the cores beyond CUDA: 752 fifth-generation Tensor Cores and 188 fourth-generation RT Cores. Its December 2025 version 4682150 gives the same figures as the April 2025 sheet 3850600 and adds the liquid-cooled card under thermal solution. Its product brief gives a base clock of 1,852 MHz and a boost clock of 2,430 MHz. The two workstation datasheets name the Tensor and RT Core generations but no counts and no clocks.

Memory: 96 GB of GDDR7 and 95.6 GiB in nvidia-smi

Every edition carries 96 GB of GDDR7 with ECC on a 512-bit bus. Spread over 64 bytes per transfer, 1,792 GB/s is 28 Gbit/s per pin and 1,597 GB/s about 25 Gbit/s, by our arithmetic; the Server Edition's product brief lists a memory clock of 12,481 MHz and gives ECC as enabled. The driver reports 97,887 MiB, 95.6 GiB, which is the figure to size against.

Our users-per-card rule takes 90 per cent of what the driver reports and subtracts 3 GiB for activations and the runtime: 0.9 × 95.6 − 3 = 83.0 GiB per card for weights and KV cache, on any edition. Llama 3.3 70B in FP8, 67.7 GiB of weights, then leaves room for 12 conversations of 8K with an FP8 KV cache on one card. Our guide to how many users one RTX PRO 6000 serves derives the rule.

Bandwidth sets the ceiling for one stream of generated tokens, because each token reads the weights once. For Llama 3.3 70B in FP8, about 70.6 GB read per token, that ceiling is about 25 tokens per second on the Workstation Edition and the Max-Q and about 23 on the Server Edition, by our arithmetic; real runs stay below it. Our GPU memory bandwidth comparison sets GDDR7 against HBM3e and the other memory types on our cards.

Compute figures: what NVIDIA states and what it leaves out

NVIDIA prints the Tensor Core figure of the two workstation cards in AI TOPS, FP4 with 2:4 sparsity, which counts on weights pruned to that pattern; ordinary weights run at the dense rate, half of it. For the Server Edition NVIDIA states no basis. Its 4 PFLOPS of FP4 equals the Workstation Edition's sparse FP4 figure, rounded, and the 2 and 1 PFLOPS for FP8 and BF16 follow the same halving steps. We plan with the lower reading, that these are sparse figures and the dense rates half of them, until NVIDIA states the basis. Our comparison of GPU TFLOPS in FP4, FP8 and BF16 works through the dense rates and marks every value that rests on this assumption.

NVIDIA's software documents which precisions the Tensor Cores take. The support matrix of Transformer Engine 2.22 lists BF16, per-tensor FP8 and FP8 block scaling for compute capability 12.0, which covers all three editions, and NVFP4 as partly supported there for training. FP4 is named on the product pages of all three editions; our guide to NVFP4 models for the RTX PRO 6000 lists checkpoints for inference.

FP64 is not in any of the datasheets. For prompt processing, fine-tuning and batched inference the Tensor Core rate matters; for single-user generation the bandwidth line above matters more. Published measurements per edition are collected in our RTX PRO 6000 benchmarks.

Power consumption by edition

The Workstation Edition is a 600 W card and the Max-Q a 300 W card; both datasheets give "Total board power". For the Server Edition the documents differ. The product page says "Up to 600W (configurable)", the family page "400-600 W", and NVIDIA's product brief SP-12355-001_v02 of June 2025 defines two modes set by the strapping of the 16-pin cable: 600 W maximum by default, or 450 W maximum by default, each with a 300 W minimum. If the cable signals less than the card's default cap, the card does not boot. We plan feeds and power supplies with the brief's 600 W maximum, which the product page and family page also give as the upper end, and treat the family page's 400 W as a lower bound that the brief's 300 W minimum replaces.

NVIDIA datasheets 5349469 and 5349650, product brief SP-12355-001_v02 (27 June 2025), read 24 October 2026; four-card totals are our arithmetic for the cards alone, without processors and the rest of the system.

The figures are board power, the ceiling the card holds itself to. The processors, memory, fans and power-supply losses come on top, which our article on RTX PRO 6000 Server Edition power at 600, 450 and 300 W works out per server, and the commands to set a limit are in our guide to the RTX PRO 6000 power limit. We supply all three editions as cards or built into workstations and servers; send us the edition and the number of cards you plan, and our team replies with the power supply, cable and cooling the build needs.

Interfaces, video engines and software requirements

All three editions connect over PCIe 5.0 x16. The Server Edition's product brief also lists PCIe Gen5 x8 and Gen4 x16 as supported links, with lane and polarity reversal. The same brief gives NVLink as "Not supported", and neither workstation datasheet lists NVLink, so a model larger than one card is split across PCIe; our article on the RTX PRO 6000 without NVLink covers the layouts.

NVIDIA datasheets 5349469, 5349650 and 4682150 and product brief SP-12355-001_v02, read 24 October 2026.

The workstation datasheets list up to four 4096 × 2160 displays at 120 Hz or two 7680 × 4320 displays at 60 Hz, and DirectX 12, OpenGL 4.6 and Vulkan 1.4 among the graphics APIs. The Server Edition ships in Display Off mode, which its brief names as required for NVIDIA's vGPU software; NVIDIA's displaymodeselector tool switches the outputs on. Its BAR1 is 128 GiB in that default mode.

MIG and vGPU by edition

NVIDIA's MIG user guide, updated on 13 October 2026, lists the same profiles for the Workstation Edition, the Max-Q and the Server Edition: four 1g.24gb instances, two 2g.48gb or one 4g.96gb, each with a quarter, half or all of the streaming multiprocessors, plus +gfx variants for graphics workloads. MIG needs Linux, driver 575.51.03 or later, a current vBIOS and the card in compute display mode, which on a workstation card turns its display outputs off. Our RTX PRO 6000 MIG profiles show what fits in each size.

NVIDIA MIG user guide, supported profiles (13 October 2026); NVIDIA list of GPUs supported by vGPU software (2 October 2026); read 24 October 2026.

Virtual machines with a share of a card therefore need the Server Edition. Our guide to RTX PRO 6000 Server Edition vGPU profiles lists the profiles and licences.

Thermal and mechanical data

The Workstation Edition has a double flow-through cooler on a 5.4-inch, extended-height board; NVIDIA's family page describes it for single-GPU workstations. The Max-Q has an active cooler, a blower design, on a 4.4 × 10.5 inch board, and its product page states that one system scales "from one to four GPUs", 384 GB across four cards. The Server Edition has no fan. NVIDIA's product page also lists a liquid-cooled Server Edition, single-slot and full-height extra-length (FHXL); the figures on this page are for the air-cooled card unless marked.

For the Server Edition the product brief gives the conditions in full: a passive heat sink that needs at least 43 CFM through the card at a 25 °C inlet, 51 CFM at 30 °C, 65 at 35 °C, 84 at 40 °C and 125 CFM at 45 °C, an operating range of 10 to 45 °C, and a board weight of 1,194 g without bracket. It belongs in a server whose maker lists it for that slot and fan set, and our list of RTX PRO 6000 Server Edition compatible servers gives the models, up to eight cards per server.

Names that cause confusion

Searches for this card often use "RTX 6000 Ada 96GB" or "RTX 6000 Pro". NVIDIA's CUDA list has one RTX 6000 Ada, the previous generation with compute capability 8.9 and 48 GB, as our comparison of the RTX PRO 6000 Blackwell and the RTX 6000 Ada records; every 96 GB card in this class is an RTX PRO 6000 Blackwell. "RTX 6000 Pro" is the same card under a reordered name.

Cisco's data sheet for its UCS C845A M8, updated 24 August 2026, also lists an "RTX Pro 6000D" with 84 GB at 600 W. NVIDIA's product pages for the three editions do not describe it, and the figures on this page do not apply to it. In a quote, the edition name and the memory size on the line item tell the cards apart; our guide to buying the RTX PRO 6000 in the EU lists the checks before an order.

What we supply

You can buy these cards from us: our team sizes the configuration to your workload and budget and replies within one business day. We supply the RTX PRO 6000 Blackwell Workstation Edition, Max-Q and Server Edition across the EU, as cards (the two workstation editions also as OEM versions) or in AI servers built to order, assembled and tested before delivery, on one EU contract and invoice with manufacturer warranty on every component. Quantities and lead time are confirmed in writing with the quote. Ask us which edition your chassis takes, with the server or workstation model and the slots, and we check power, cable and airflow before the quote.

FAQ