> 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.
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.
We're already seeing reports of the 12HPWR connector not being able to handle the power the 5090 demands [1]. It seems that for this generation, NVIDIA left little headroom. Might be best to run these undervolted to have some of that headroom back.
It's more that it's perfectly happy sending all 600w over a single wire if the resistance on the other wires is high enough. Since there's a single point of connection to the card and a single shunt, there's no way to do load balancing like when there used to be two or even three 8 pin cables.
> 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.