Using the same exploit that recently allowed unlocking features of Nvidia mining GPUs, an unlocker for RTX 3090 has been created. Unfortunately the card doesn't have hidden VRAM to enable, but we can get some free floating point processing power.
The install instructions are a bit vague, but on my Fedora install I had to manually add the kernel command line arg (gpu_unlock.targets) and rebuild initramfs after running the build/install command.
I did some simple comparison using ComfyUI, and image generation became 33% faster when using fp8 and bf16 models. However, with an int8 convrot quantized model I only got around 8% speedup, which make sense since it would use the int8 tensor cores of the RTX 3090 which weren't locked to begin with.
I didn't write down my exact tokens/s in llama.cpp before installing the unlocker, but the numbers I got after were roughly the same as I remembered, which also makes sense since I use quantized models that afaik use integer tensor ops.
As with all hardware "hacking" you should be careful. This seemed relatively safe to me since it doesn't permanently alter anything on the GPU so it should be easy to revert. And it shouldn't mess with current limits or clock speeds, which could more easily damage the hardware.
The install procedure does involve changing some kernel boot parameters, so if you make a mistake it's possible that you soft-brick your OS.
Just to clear up some potential misunderstanding.. Ampere is the Nvidia GPU architecture, and your card is in the same category as RTX30xx. This can matter when it comes to hardware support for llama.cpp and other software.
Having tensor cores affects the speed your card does certain calculations, but AFAIK there are different types of tensor cores in GPUs and if you're using quantized models (which you most likely will be, unless you're training your own models) then it's INT8 performance that matters.
Your amount of VRAM determines how large (which usually correlates to how "intelligent" the model is) model you can fully load on the GPU. If the entire model doesn't fit in VRAM it can use CPU RAM swapping (partial offloading) which severely harms performance, though MoE models are less affected.
So both VRAM and tensor cores affect generation speed, but in completely different ways.