I’m currently running the Denoise MAX model and don’t think it’s any faster.
I can see that in the second step, communication is mainly happening over the PCI-E bus, and the whole system is stuttering because the GPU is busy with scheduling.
The GPU core is at 100% utilization, but power consumption peaks at 70 - 200 watts out of a possible 450.
Could it be that the new Neuroserver version is moving data from RAM to the GPU and back to RAM again?
Instead of using RAM and VRAM without overflowing the VRAM?
SYS: 9950X3D2 - 128GB ECC enabled - RTX 4090 ECC enabled.
