I have a system with 8x RTX Pro Blackwell 2000 cards. The BARs are consuming all of the 32-bit PCI address space so that none is left for my other PCI devices.
The Display Mode Selector Tool v10_July25 does not support the 2000 Blackwell (only 5000/6000).
Is there a tool available to disable the physical displays on the card or reducing the PCI BAR size that support the 2000 Blackwell?
Thanks!
Hi,
Bumping this thread. Is there a way to change the size or disable BAR1 on the RTX Pro Blackwell 2000 card?
Thanks again!
Hi, the workstation GPUs already come with the smallest (supported) BAR1 setting by default, if your system doesn’t support the 8x GPUs, you will need to search for a different system, maybe check in our certified systems catalogue? [just checked for RTX PRO 2000, max #GPUs listed is like 4]. But honestly, 8x entry/midrange GPUs is not a very common config, OEMs would take the effort to run tests and certify.
Any 8x (bigger) GPU capable system should also be happy with 8x 2000s…
Why are some of the workstation GPUs listed as supported in our displaymodeselector tool? Because they can be changed to a BIGGER BAR1 address-range… (useful for very memory transfer intensive workloads…)…
Out of curiosity, could you describe the usecase, for 8x midrange GPUs, 32x outputs, in a single chassis? we always assume relatively demanding workloads, when thinking many-GPU systems, so wonder what load per GPU you expect here…
many thanks
-Frank
Hi,
I am using the 8x Blackwell 2000s for AI workloads, so I do not need the video output.
I also have 8x accelerators. The issue is that BAR1 uses 32-bit addressing. For the Blackwell 2000 card, BAR1 is 64M. The accelerator also has a small 32-bit BAR which is 8M. When combining 8x Blackwells and 8x accelerators the 32-bit address space is exhausted.
The displaymodeselector tool has 3 options: 2 different sizes and an option to disable BAR1 on the 5000/6000. I was hoping a similar option was available to disable BAR1 for the Blackwell 2000. This would resolve the 32-bit address space issue because I do not need BAR1 for my application.
Thanks,
Richard
HI Richard,
I’m sorry, but like I wrote the RTX PRO workstation GPUs already come SHIPPED with smallest BAR setting, and there only are options to INCREASE the BAR1:
Explicitly, the option with physical displays disabled has the BIGGEST BAR1 with GBs, not MBs, of address range needs!
“displaymodeselector --gpumode <mode_name> [-i num]
mode_name: physical_display_enabled_256MB_bar1
physical_display_enabled_8GB_bar1
physical_display_disabled
num: (Optional) GPU index”
the last option “physical_display_disabled” will always set to 64 or even 128GB of BAR1 - depending on SKU - the lack of a size descriptor in the option is NOT saying NO BAR1 regions address space needed…
I guess I see where a 1:1 between accelerator and dedicated GPU comes in easy, but maybe you can consider a 2:1 ratio, and use only 4x, maybe more powerful GPUs? Or consider like 5000 class GPUs, and use MIG to separate/isolate the GPU hardware a bit more?
Now that I think it more through. I also wonder what special system/platform you use, to actually have PCIe slots for 8x GPUs AND 8x accelerators? lesser GPUs might open the range of systems/platforms, you can use… - just a thought…
But for sure, there is NO option to go with a smaller BAR1 address range…
regards
-Frank
