I haven’t heard of anyone using the 5090 with flame? Is there anyone out there doing this in production or a home machine?
Is it worth upgrading a 5000ADA, and losing another Pcie slot? Is it stable or is it kind of like those titan cards back in the day? I’m guessing prices are going to drop even further in the next few months, so I’m investing this for my home machine.
instead of spending it on AWS credits you never wanted, just roast 44,000 marshmallows. It will be much more satisfying and take less time than learning AWS infrastructure.
It it helps - today I was working in Comfy and loaded a model that was quantized with fp8 4mf3fn, which is not supported on the A5000 generation of GPUs (Cuda 36). You need a Ada generation or Blackwell card for that, or use an older quantizing version of the model.
So it may not matter much in Flame, but that is a current issue I ran into.
DKU 20.2, released with Flame Family 2026.2 Update, adds support for the NVIDIA driver required for the Blackwell GPU generation. We qualified the RTX Pro 6000 and your mileage may vary using a non qualified card.
As far as Rocky Linux, please stick with RL 9.5 for Flame Family workflows.
I’ve been using Rocky Linux and Flame with an Nvidia 5090 for about 6 months. The drivers I installed from DKU or Nvidia’s website always gave me trouble, so I switched to installing them through ELRepo. Since then, I haven’t really had any issues.
I checked out prices for the RTX 5000ada, and they are super reasonable(refurbished). Ever the 42gb RTX 5000 Ada isn’t terrible, but the 72gb one is still around $8k.
So unless there’s an overwhelming reason to buy the blackell, I’m probably going to purchase the 5000ada.