Nvidia GeForce RTX 5090 Linux GPU Compute Benchmarks (opens in new tab)

Basically it seems to be worth about 1.333X as much as a 4090, as it has about ~1.4X performance ~1.333X the VRAM but uses about 1.4X the power. So 4090 is still a reasonable option.

One thing that concerns me is running these dies at 500w+; that seems like an awful amount of heat over a very small area and increasing the power (vs the 4090) must mean also decreasing it's lifetime.

Sadly no lora/finetune benchmarks. Phoronix sort of missed the whole idea, imo.

From the last page (https://www.phoronix.com/review/nvidia-geforce-rtx5090-linux...)

> When taking the geo mean across 60+ benchmarks of CUDA / OptiX / OpenCL / Vulkan Compute, the GeForce RTX 5090 was delivering 1.42x the performance of the GeForce RTX 4090. A very nice generational improvement considering many of the Windows reviews yesterday were talking up around ~1.3x the performance when not dealing with DLSS4 or similar tech. The GeForce RTX 5090 was delivering great GPU compute performance whether you are into CUDA, OpenCL, or Vulkan.

> On a performance-per-Watt basis though the GeForce RTX 5090 tended to deliver similar power efficiency to the RTX 4080/4090 graphics cards.

It's really unfortunate they seem to be unobtainium, despite all these benchmarks I've yet to see a place where they're actually sold outside absurdly priced ebay price-gougers. As I don't think things will change anytime soon, I've moved to coping with multiple lesser cards to try and get models to fit. Surprisingly I was able to get a mash of Intel and nvidia cards to work together for same inference job after two days of fiddling.

At $2,000 base and $5-7k scalper price, this is outside of the reach of most people forever.

To save the money I'll just sit back and chug away on my 3000 series until it explodes, TYVM