RTX PRO 6000 Blackwell vs RTX 6000 Ada: what a generation changed

What changed in one generation

FP32 grew 37%, RT Core 80%, memory bandwidth 87%. The last number matters most. (NVIDIA’s product page says 125 and 380 TFLOPS; the datasheet PDF rounds to 126 and 382.)

Memory outweighs cores

In 48 GB, a 70-billion-parameter model fits only with aggressive quantisation and short context, with the KV cache trimmed. In 96 GB the same model sits in Q4 with a working context, and room to spare. Single-request llama-bench figures reported by owners of the RTX PRO 6000 (Q4_K_M unless stated; treat as indicative):

And the budget trap people keep falling into: two RTX 6000 Ada do not add up to 96 GB. The datasheet says a plain “No” against NVLink; traffic goes over PCIe, and a model that does not fit one card has to be layer-split with a speed penalty. Practitioners frame a second card’s value differently: keeping several specialised models resident: an LLM, a reranker, an embedding model.

FP4 and fifth-generation Tensor Cores

FP4 halves memory versus FP8, and that is where the headline 4,000 AI TOPS comes from: the footnote reads “effective FP4 TOPS with sparsity”, so a dense model scores lower. The gain from lower precision also arrives from an unexpected direction: token generation is bound by memory speed, since every token reads all the weights. The precision that wins is the one that makes the model physically smaller.

600 W: what capping the card costs

The doubled TDP is scarier than it should be. The same card was run at 600 W and 300 W limits on llama.cpp:

Half the power budget costs about 6% on generation and nearly half on long-prompt reading. For chat, barely noticeable; for RAG, agents and large documents, noticeable immediately: response time there is set by the reading phase. And for those who need 96 GB inside a 300 W budget there is the Max-Q: the same GDDR7 ECC memory in the RTX 6000 Ada’s dimensions.

What Ada never could do

MIG slices the card into isolated instances: four of 24 GB, two of 48, or one of 96. Ada has no hardware partitioning at all; its only path is vGPU with a ceiling of 32 VMs per card (16 in mixed sizing), and enabling vGPU switches off the RTX 6000 Ada’s physical display outputs, per its own datasheet footnote.

The Blackwell MIG story has its own catch: retail cards shipped with firmware 98.02.52.00.02 while MIG needs at least 98.02.55.00.00, and NVIDIA’s position was that the vBIOS update comes from the partner who sold the card. Ask for the firmware version before paying. Separately: vGPU is supported only on the Server Edition, on any hypervisor, Proxmox VE included; the Workstation and Max-Q cards get MIG but no vGPU.

What the benchmarks show

Power figures are for the whole test bench, not the card

And the honest weak spot of every comparison of these two cards: paired same-day, same-build measurements are rare, and almost all are about rendering, not LLMs. Where pairs exist: Blender 4.4 Monster gives 7,870 vs 5,633 samples/min (+40%), V-Ray 12,128 vs 10,766 vpaths (+13%) on the same bench. That is the truest answer to “how much faster”: it depends on the workload more than on the generation.

What the presentations do not mention

NVLink is absent from the RTX PRO 6000 documentation entirely; third-party sources say it is gone and everything goes over PCIe Gen5, so two cards will not give a 192 GB pool. Owners report that ECC on GDDR7 cannot be switched off: on earlier generations render shops disabled it, here nvidia-smi commands have no effect, with no official explanation of how many gigabytes ECC costs. Early drivers added work too: open-kernel-module-only support, a GSP-timeout thread that ended in two RMAs, and PCIe Gen4 instability on some boards cured by Spread Spectrum. The card is not bad; just budget time for firmware, driver and motherboard matching rather than installing it the Friday before a deadline.

When the premium is justified, and when Ada stays the smart choice

Take Blackwell if the model is 70B-class or 30B with long context; if you need MIG to divide the card between teams; if a video pipeline actually fills four gen-9 NVENC engines; if your workloads read long prompts. And if the workstation already has the PSU headroom, the longer chassis and the airflow.

Stay on Ada if the model already lives in 48 GB and will not grow; if a unified fleet and identical drivers are worth more than the gain; if the room’s power budget is fixed. Moving from 300 W to 600 W drags a new PSU, sometimes a new chassis, and in multi-card builds a dedicated high-amp line. Separately: neither generation really does FP64, so for numerical simulation the move changes nothing.

What we supply

Eurokommerz supplies the RTX PRO 6000 Blackwell Workstation Edition (600 W, double-flow-through; OEM version available), the Max-Q (the same 96 GB in a 300 W budget and Ada-sized dimensions), the Server Edition (passive, vGPU support), the RTX 6000 Ada 48 GB, and the L40S for round-the-clock rack duty, EU-wide with manufacturer warranty.

FAQ