Experimental

RTX PRO 6000 Blackwell, on the appliance

The appliance runs on Azure's RTX PRO 6000 Blackwell SE sizes. It works, we have proven it end to end on real silicon, and it is behind a manual enable step for reasons this page explains honestly. If you want the easy path, the recommended sizes need no enable step at all.

The sizes

Azure size names in this family follow the pattern Standard_<size>_RTXPRO6000BSE_v6. The card is 98,304 MiB; the smaller sizes are hardware slices of it, and what the guest sees is the slice.

SizeCardVRAMFitsWorth knowing
NC24lds_xlQuarter card24,576 MiB visibleModels up to 24 GBSlower than a full A10 at the same VRAM. Consider an A10 or A100 size first.
NC72ds_xlHalf card49,152 MiB visibleModels up to ~45 GBHalf the card, half the memory bandwidth.
NC144ds_xlFull card98,304 MiBModels up to ~90 GBThe size where this family makes sense.
NC288ds_xlTwo full cards196,608 MiBModels up to ~190 GBThe biggest Blackwell box today.

We do not publish tokens-per-second figures for this family yet. The runs exist, but they are internal sizing measurements and our rule is that a published number is a measured, dated, reproduce-it-yourself number.

What "experimental" means here

Not "it might not work". It means the deployment posture differs from every other size we support, and there are sharp edges we would rather list than have you discover.

Secure Boot and vTPM must be OFF, at deploy time

The only driver that binds this family is the GRID guest driver, and it is unsigned. Both settings are deploy-time choices that cannot be changed from inside the VM: a box deployed with Trusted Launch must be deallocated and updated before the enable step can do anything. Every other size we support keeps Secure Boot on; this family is the exception, which is a large part of why it is experimental.

One manual enable step

The driver cannot be redistributed inside a commercial image, so the appliance fetches it from Microsoft when you run the enable step from the menu. It is safe to re-run, reports status when run bare, and switches the appliance to the right engine build afterwards.

The partial-card sizes report partial cards

The quarter and half sizes are vGPU slices of a 98,304 MiB card, and the appliance sees only the slice: 24,576 MiB on a quarter card. That is expected, not a fault. Bandwidth is sliced with the memory, so a quarter card is slower than a full A10 with the same 24 GB. The appliance tells you when it detects a partial profile.

Licensing has a silent grace period

An unlicensed vGPU runs at full speed for around 20 minutes after boot, then the driver clamps compute several times slower, with no error and no warning. The enable step handles the licence; if a Blackwell box mysteriously slows down shortly after boot, re-run the enable step and check the licence state before blaming the model.

How to try it

  1. Deploy the appliance from the Azure Marketplace onto a Blackwell size with security type Standard (Secure Boot and vTPM off). This family is deliberately not in the one-click Deploy Now selector.
  2. Open the VM's serial console in the portal and run the Blackwell enable step from the llm menu.
  3. Serve a model that fits the slice you deployed. The menu declines models that do not fit.

Questions, or numbers from your own runs? We would genuinely like to hear them.