RTX Pro 6000 Blackwell: The better GeForce RTX 5090
While gamers have to make do with a much slimmed-down Blackwell GPU, there is a faster version for expensive workstations.
The RTX Pro 6000 Blackwell in the 600 watt version.
(Image: Nvidia)
The GeForce RTX 5090 gets a faster sister model with more memory: the RTX Pro 6000 Blackwell Workstation Edition uses the same GB202 graphics chip, but in a more powerful version with 24,064 instead of 21,760 active shader cores. On paper, this makes it up to 19 percent faster than the GeForce RTX 5090, with 96 GB of GDDR7 memory instead of 24 GB, including ECC error correction.
The RTX Pro 6000 Blackwell is intended for workstations with a focus on the development of AI algorithms. With the new generation, Nvidia is moving away from its tradition of only releasing energy-efficient versions with so-called "blower" coolers. These use a radial fan that blows fresh air through a very small heat sink.
(Image:Â Nvidia)
Such a variant is still available with the RTX Pro 6000 Blackwell Max-Q, but there is also a more powerful 600-watt version with what appears to be the same cooler as the GeForce RTX 5090 Founders Edition. Nvidia is pushing the GPU to its limits here, at the expense of efficiency – the electrical power consumption increases exponentially.
If you have the money, the RTX Pro 6000 Blackwell will be an even better gaming graphics card than the GeForce RTX 5090. Prices are not yet known, but PNY will explicitly launch the two new models in retail stores from May. Considering the predecessor RTX 6000 Ada Lovelace (starting from 7407,99 €), a five-digit price would not surprise us.
| RTX Pro 6000 Blackwell compared to the GeForce RTX 5090 | |||
| Graphics card | RTX PRO 6000 Blackwell | RTX PRO 6000 Blackwell Max-Q | GeForce RTX 5090 |
|
Graphics chip |
GB202 | GB202 | GB202 |
| Shader- / Raytracing- / Tensor Cores | 24.064 / 188 / 752 | 24.064 / 188 / 752 | 21.760 / 170 / 680 |
| Boost clock | ca. 2600 MHz | ca. 2300 MHz | 2407 MHz |
| FP32 performance | 125 TFlOPS | 110 TFlOPS | 105 TFLOPS |
| INT4 performance | 4000 TOPS | 3511 TOPS | 3352 TOPS |
| VRAM | 96 GByte GDDR7, 1792 GByte/s | 96 GByte GDDR7, 1792 GByte/s | 24 GByte GDDR7, 1792 GByte/s |
| Max. Power consumption | 600 Watt | 300 Watt | 575 Watt |
| Cooling | 2 Ă— Axial | 1 Ă— Radial | 2 Ă— Axial |
No NV-Link for coupling
Meanwhile, AI models with billions of parameters fit into 96 GB of memory, but not the flagship models from DeepSeek, Meta & Co. Unfortunately, neither the current GeForce nor the professional graphics cards now announced have NV-Link connections. This means that several GPUs can only communicate with each other via 16 PCI Express 5.0 lanes, which limits the transfer rate to 63 GByte/s. Without NV-Link interconnect, the memory for applications cannot be meaningfully combined.
Videos by heise
Nvidia thus separates the markets. If you want to connect several GPUs via NV-Link, you have to use the expensive accelerators. However, Blackwell accelerators of the B100 and B200 types are not yet available individually on the open market. Models from the previous Hopper generation, such as the H100 and H200 (starting from 33915 €), already cost dozens of thousands of euros.
The slower workstation graphics cards RTX Pro 5000 Blackwell (14,800 shaders, 48 GB RAM), RTX Pro 4500 Blackwell (10,496 shaders, 32 GB) and RTX Pro 4000 Blackwell (8960 shaders, 24 GB) are being released alongside the RTX Pro 6000 Blackwell.
Empfohlener redaktioneller Inhalt
Mit Ihrer Zustimmung wird hier ein externer Preisvergleich (heise Preisvergleich) geladen.
Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (heise Preisvergleich) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.
(mma)
