Nvidia announces the RTX Pro 6000 Blackwell graphics cards

Professional graphics cards RTX Pro 6000 have been another of the great novelties presented by NVIDIA at the GTC 2025 conference held this week in California and in which we have seen solutions of all kinds, with the Blackwell Ultra superchip and the personal computers of AI, such as the most relevant announcements.

RTX Pro 6000 are based on the Graphic Architecture ‘Blackwell’ released for the consumer market in the dedicated graphics cards GeForce Series 5000, but with maximization of the graphic nuclei and with the necessary characteristics for use in the professional segment, servers, data centers and workstations where they are directed.

The series combines a AI computing and a visual avant -garde performance accelerating the entire spectrum of creative work and artificial intelligence charges. Nvidia specifically indicates the AI ​​of agents, physical AI, scientific computing, photorealist rendering, 3D graphics and real -time video processing.

RTX Pro 6000 versions

To cover all market segments, from computer desks to data centers, Nvidia has developed three different versions that we are going to detail you:

NVIDIA RTX PRO 6000 Blackwell Workstation Edition

The company says that it is the Most powerful desktop GPU ever createdcapable of addressing the most demanding creative workflows and with a processing power without rival on a graphics card destined to motorize the most powerful workstations.

With a plate design similar to that of the RTX 5090 (the most powerful dedicated in the consumer market) and a graphic core GB202 Manufactured by TSMC in 4 NM processes, the card includes 24,064 CUDA cudders distributed in the 192 SM physically present on the plate, in addition to 768 tensor nuclei, 192 RT nuclei, 768 TMU and 192 ROP. The card incorporates an impressive ECCDR7 memory capacity with 96 gbytes dedicated in a 512 -bit memory interface and with a bandwidth of 1792 GB/s. Its TGP consumption has been established in 600 W.

Nvidia says that the fifth generation of tensor nuclei offers up to 3 times the performance of the previous generation for a performance of AI of 4.000 TOPS. Add compatibility with FP4 accuracy and multiple dlss 4. cadres generation technology in terms of the fourth generation of RT nuclei offers up to twice as much performance as the previous one, which accelerates the rendering for the creation of multimedia content and entertainment, AECO design and creation of manufacturing prototypes.

In connectivityhighlights compatibility with PCI Express Gen 5 that offers double bandwidth that the PCIE Gen 4, improves the data transfer speed from the memory of the CPU and offers a faster performance for tasks with intensive data use, such as AI, data science and 3D modeling. For data output, the Displayport 2.1 interface opens, with four ports capable of feeding high resolution screens with up to 8K at 240 Hz and 16K with 60 Hz refreshment rate.

It supports all types of graphic APIs (Directx 12, Shader Model 6.6, Opegl 4.63, Vulkan 1.33) and computing (CUDA 11.6, OpenCl 3.0 or Directcompute). It will be available soon without departure Price.

NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition

This version is also intended for professional work stations and has essentially almost the same specifications as the previous model, but with a TGP reduced to 300 watts and a simpler motherboard design of 2 slots that use a side fan. Although it is not a rendering server, it is designed for machines that welcome the installation of several cards of this type.

Its degree of performance is slightly lower than the general model (AI performance: 3511 tops – RT nuclei performance: 333 Teraflops – Simple precision performance: 110 Teraflops), but also its consumption and price. It includes the same number of CUDA nuclei and the 96 gbytes of dedicated GDDR7 and error correction.

It also uses the same internal interface, PCIE 5.0, and the External Ports Displayport 2.1 for connection to monitors with resolution of up to 16k.

NVIDIA RTX PRO 6000 Blackwell Server Edition

The third variant is intended for servers and Share the general characteristics of the series As for its base in the ‘Blackwell’ architecture, the GB202 graphic nucleus, the 24,064 CUDA cores, 192 SM, 768 Tensor nuclei, 192 RT nuclei, 768 TMU and 192 ROP. Also identical the ECC GDDR7 memory with 96 dedicated Gbytes or the four Displayport 2.1 connectors.

Its objective for servers and data centers makes the difference in format, with a Optimized plate design for rack -mounted servers and large rendering farms. The dissipator is based on the air flow of the rack for refrigeration and its TGP consumption varies from 400 to 600 watts.

Highlight the capacity as a multi -sectors GPU (MIG), which expands the performance and value of these accelerators by allowing the creation of up to four totally isolated instances. Each instance of MIG has its own broadband memory, cache and computing nuclei, and guaranteed service quality (QOs), which expands the scope of accelerated computing resources to each user.