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

What changed in one generation

The RTX PRO 6000 Blackwell has 96 GB of GDDR7 at 1,792 GB/s, twice the memory of the RTX 6000 Ada with 48 GB of GDDR6 at 960 GB/s. There is no RTX 6000 Ada with 96 GB, as NVIDIA lists that card with 48 GB only. The Blackwell card adds FP4 and MIG, which splits it into four 24 GB instances, two of 48 GB or one of 96 GB, and its Workstation Edition draws 600 W against the Ada’s 300 W; the Max-Q edition keeps the 96 GB at 300 W.

FP32 grew 37%, RT Core 80%, memory bandwidth 87%. Of the three, the bandwidth gain matters most for language models. (NVIDIA’s product page says 125 and 380 TFLOPS; its June 2026 datasheet gives 126 and 382, the architecture whitepaper 126.0 and 381.8.)

Memory: what fits in 48 GB and in 96 GB

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. StorageReview measured 31.84 tokens per second for Llama 3.1 70B and 163.1 for gpt-oss-120b in LM Studio on this card (October 2025, quantisation not stated).

Two RTX 6000 Ada cards do not add up to one 96 GB card. The datasheet says “No” against NVLink, so traffic between the cards goes over PCIe, and a model that does not fit one card has to be split across both, by layers or with tensor parallelism, at a speed penalty. A second card is more useful for keeping several specialised models resident at once, such as an LLM, a reranker and an embedding model.

FP4 and fifth-generation Tensor Cores

FP4 cuts the weights to a little over half their FP8 size (NVFP4 stores 4.5 bits per weight with its block scales). The headline 4,000 AI TOPS is an FP4 figure: the footnote reads “effective FP4 TOPS with sparsity”, so dense FP4 maths runs at half that rate. Lower precision also speeds up token generation, which is bound by memory speed, because for a dense model every token reads all the weights. For generation speed, what counts is how much a format reduces the size of the weights in memory.

600 W Workstation Edition or 300 W Max-Q

Puget Systems ran the 300 W Max-Q and the 600 W Workstation Edition on the same workstation in July 2025 and measured the Max-Q 5 to 14 per cent slower in content-creation applications; we found no published LLM test with the same method. Generation, which is bound by memory bandwidth, should lose least, and long prompts, which are bound by the Tensor Cores, should lose most. For chat the difference should be barely noticeable. RAG, agents and large documents would show it first, because their response time is set by prompt processing. For a 300 W budget, the Max-Q offers the same 96 GB of GDDR7 ECC memory in the RTX 6000 Ada’s dimensions.

We supply the Max-Q as well as the 600 W Workstation Edition. Tell us the workstation model and how many cards you want in it, and we will say which edition fits.

MIG and vGPU: Blackwell against Ada

MIG slices the card into isolated instances: four of 24 GB, two of 48, or one of 96. Our RTX PRO 6000 MIG profiles guide lists the profiles and their sizes. 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.

vGPU is supported only on the Server Edition, on the hypervisors NVIDIA lists for it; the Workstation and Max-Q cards get MIG but no vGPU. On those two, MIG needs a minimum vBIOS. Users on NVIDIA’s developer forum reported Workstation cards delivered with vBIOS 98.02.52.00.02, while MIG needs at least 98.02.55.00.00 on that card (98.02.6A.00.00 on the Max-Q), plus Linux, driver 575.51.03 or later and the display mode switched to compute, which turns the display outputs off. NVIDIA’s MIG guide says to contact the reseller or system provider for a vBIOS update, so ask for the vBIOS version before paying.

What the benchmarks show

Figures from StorageReview’s review of 7 October 2025 (Threadripper 7980X workstation; LM Studio quantisation not stated). Power figures are for the whole test bench, not the card.

Paired measurements of the two cards on the same day and the same build are rare, and almost all of them are about rendering rather than LLMs. Where pairs exist, StorageReview’s bench gives Blender 4.4 Monster 7,870 against 5,633 samples/min (+40%) and V-Ray 12,128 against 10,766 vpaths (+13%). In Puget Systems’ review of 11 June 2025, the Workstation Edition was about 50 per cent faster than the RTX 6000 Ada in both Blender and V-Ray RTX.

NVLink, ECC and driver requirements

We found no mention of NVLink in the RTX PRO 6000 datasheets and product pages, but NVIDIA’s product brief for the Server Edition lists it as not supported, as do Lenovo’s product guides for the Max-Q and the Server Edition, so everything between cards goes over PCIe Gen5 and two cards are not one 192 GB pool. NVIDIA’s GeForce whitepaper explains that GDDR7 carries ECC inside the memory chip and that it is always enabled on GeForce RTX cards with GDDR7 memory. NVIDIA’s RTX PRO whitepaper adds only that the GDDR7 implementation for professional products “includes support for Enhanced CRC for RAS”; we found no NVIDIA document on whether ECC can be switched off on the RTX PRO cards or whether it costs any capacity. Blackwell cards run only with NVIDIA’s open GPU kernel modules, so check the driver flavour your distribution installs. Plan time for matching firmware, driver and motherboard before the card goes into production.

When to upgrade from the RTX 6000 Ada

Upgrade to 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 uses all four gen-9 NVENC engines, or if your workloads read long prompts, provided the workstation has the PSU headroom, the chassis length and the airflow for it.

Stay on Ada if the model already fits in 48 GB and will not grow, if a unified fleet and identical drivers are worth more than the gain, or if the room’s power budget is fixed. Moving from 300 W to 600 W means a new PSU, sometimes a new chassis, and in multi-card builds a dedicated high-amp line. Neither generation has meaningful FP64 performance, so solvers that must run in double precision gain little. Single- and mixed-precision solvers gain from the 87 per cent higher bandwidth and from twice the memory for larger meshes; our GPU guide for CFD covers that choice.

What we supply

We supply the RTX PRO 6000 Blackwell in all three editions, EU-wide with manufacturer warranty, on one EU contract and invoice, as cards or in AI servers built to order: the Workstation Edition (600 W, double flow-through), the Max-Q (the same 96 GB in a 300 W budget and the RTX 6000 Ada’s dimensions) and the Server Edition (passive, vGPU support), plus the L40S for rack servers and the RTX 6000 Ada for fleets that stay on Ada. The Workstation Edition and the Max-Q are also available to order as OEM versions.

FAQ