NVIDIA’s RTX PRO 6000 Blackwell puts 96 GB of GDDR7 on one workstation card. That is three times the memory of the GeForce RTX 5090, at the same listed memory bandwidth, and enough to run models that otherwise need several consumer GPUs.

Specifications

NVIDIA lists three editions in the RTX PRO 6000 Blackwell series. All three have 96 GB of GDDR7 with error-correcting code (ECC) and a PCIe Gen 5 x16 interface.

Edition Power Build NVIDIA’s intended use
Workstation 600 W Dual-slot, 5.4 by 12 inches, double flow-through cooler Single-GPU workstations
Max-Q Workstation 300 W Dual-slot, 4.4 by 10.5 inches, active cooling Dense workstations of up to four GPUs
Server 400 to 600 W Passively cooled Multi-GPU servers

The Workstation Edition’s memory bandwidth is 1792 GB/sec. NVIDIA’s pages in our sources give bandwidth only for that edition, so we do not state it for Max-Q or Server. The product page also gives the Workstation Edition 4000 AI TOPS, and a footnote says this is theoretical FP4 with sparsity. Treat it as a ceiling, not a typical number.

What 96 GB makes possible

Memory is the first limit for local models, because the weights and the context cache must sit in GPU memory. 96 GB on one card makes room for:

  • Large mixture-of-experts models in 4-bit form. The gpt-oss-120b model card says that model runs on a single 80 GB GPU, which is below this card’s capacity.
  • Models that need more headroom for weights or long context windows than a smaller card allows.
  • More than one model loaded at a time, if their combined size fits.
  • Fine-tuning and generative AI work. NVIDIA’s page names fine-tuning LLMs, generative AI and running AI agents locally as uses.

Ollama’s hardware table lists the RTX PRO 6000 Blackwell among supported NVIDIA GPUs. The usual local tools work. The fit list on this page shows which models in our data fit.

Compared with the RTX 5090

NVIDIA’s pages give the GeForce RTX 5090 32 GB of GDDR7 on a 512-bit interface, 1792 GB/sec of bandwidth, 575 W total graphics power and 3352 AI TOPS. Against the RTX PRO 6000 Workstation Edition:

  • Memory bandwidth is the same, 1792 GB/sec on both.
  • Memory is 96 GB against 32 GB.
  • Maximum power is 600 W against 575 W.

If your models fit in 32 GB, the 5090 is enough. The PRO 6000 is for models that do not, or for loading more at once.

Compared with the DGX Spark

NVIDIA lists the DGX Spark at 128 GB of unified memory and 273 GB/s of bandwidth. The Spark holds more; the PRO 6000 reads memory about six and a half times faster. Capacity sets what loads and bandwidth sets how quickly tokens come out, so for models under 96 GB the PRO 6000 should generate faster, while the Spark can hold bigger models. The Spark is a complete small computer. The PRO 6000 is a card that needs a host PC.

FAQ

How much memory does the RTX PRO 6000 Blackwell have?

All editions in NVIDIA’s series table have 96 GB of GDDR7 with ECC.

What is the Max-Q version?

The 300 W edition, with the same 96 GB, built for workstations holding up to four GPUs. The Workstation Edition runs at up to 600 W.

Is it better than an RTX 5090 for local LLMs?

It has three times the memory at the same listed bandwidth. If a model fits in the 5090’s 32 GB, the extra capacity does not help that model.

Can it run models that the DGX Spark can?

Only if they fit in 96 GB. The Spark has 128 GB, and the PRO 6000 has much higher bandwidth for models that fit.