Upgrading to a Nvidia 5080/5090 from a 3080?

I know it's not a direct answer but maybe antecdotal. I upgraded from a 3060 to a 4070ti super and noticed about a 20 percent uptick in speed.

I feel like the newer cards use a lot of bells and whistles that AP app’s don’t care about. But sheer horse power isn’t that much greater. IMHO

I just went from a medium-high end 4 year old gaming laptop with a mobile 3070 to a brand new high spec desktop with a 5080.    Moved to a 61mp full frame sensor and things just felt so slow on the laptop.   Generally the XT tools run about 2-3x faster on a full frame image than on the laptop.   And wbpp is also over 3x faster.    

I imagine going from a 3080 desktop card to a 5080 wouldn’t be as big of a jump

Does anyone know of any external GPU enclosures that can house and power the 5090?  I would consider picking one up and testing since I would also make use of the 5090 for work, but am restricted to connecting over USB-C; for the work I do the data bandwidth constraint has negligible impact compared to the computational benefit.

Alex Nicholas:

Billy Buchanan:
Does anyone know of any external GPU enclosures that can house and power the 5090?  I would consider picking one up and testing since I would also make use of the 5090 for work, but am restricted to connecting over USB-C; for the work I do the data bandwidth constraint has negligible impact compared to the computational benefit.

You could make one - but none of the e-GPU enclosures I've seen are REMOTELY large enough for a 5090, and they certainly do not provide enough power for one.. 

Being restricted by USB-C (if its actually USB-C 3.1) means that regardless of the bandwidth you think you don't need, all the data to be computed and all the results post computation need to be sent backwards and forwards via a USB 3.1 bus... There is still a significant hit.. 

Honestly, I'd find a laptop with a 5080 discrete GPU if I were you rather than trying to find a way to e-GPU a 5090...

Given the batch sizes that are used for work the bandwidth isn’t really a limitation at all.  The GPU memory tends to be the bigger factor in terms of a constraint on the compute.  For image processing things could easily be way more challenging though.

My take is that it is faster in terms of percent but the actual time savings in minutes is not significant. Because of that unless you also are a gamer with that machine it might not be worth it. Of course depending on your power supply you might need to upgrade that as well….

I went from a 3060 to a 4080 and noticed only a small improvement in actual seconds to already fairly fast times for Blur X and such. This on a AMD  9950X  with 96  Gig RAM.

For a 571 based image BXT took maybe 25 sec on the 3060 and that went to about 15 sec on the 4080. If it were not for the gaming it would not have made much sense to upgrade, especially since I did need to go to a larger PSU.

You'd definitely have to run it a lot for it to add up to enough to matter.  Generally for a narrowband image for example I run blurx three times (pre channel combination to use the appropriate FWHM value to each), on starless data, I run noisex twice, once on starless linear data and once on starless stretched data, and starx four times, once on each narrowband master, and once to extract the stars from my RGB combined star image.  Out of the many, many hours I spend processing an image that amount of time is a tiny, tiny percentage.

I think once PI implements their upcoming CUDA optimizations throughout their whole suite, there will be a bigger bang for that large buck.   I only upgrade every 5 years or so, so I choose the higher end stuff so it lasts longer.   Over doubling the speed of the RC tools with a better Nvidia card was nice, but well over 3x the speed for WBPP was my main goal which I know has nothing to do with the card yet.    But if I can get another nice boost in WBPP "for free" with CUDA sometime soon, that'll be icing on the cake.

Plus my games all run great now 

Rick

Brian Diaz:
HI everyone

i have AMD RYZEN 9  5950X 16 core, 128  GB RAM, nvidia 3080 Ti
  
image with 373 MB full frame .  blurxterminator around 30s  , starxterminator around 25s


CS
Brian

Similar performance to mine.. BXT and SXT both being cuda accelerated, my 4080 does about the same BXT/SXT times.

I've got a Ryzen 9 9950X and 128GB RAM, and things like WBPP changed immensely when I upgraded to the Ryzen 9 from my old dual Xeon rig, but BXT/NXT/SXT didn't change almost at all jumping from my 1080 to a 4080, and no change at all when I upgraded the rest of my rig around that 4080.

Alex Nicholas:

Brian Diaz:
HI everyone

i have AMD RYZEN 9  5950X 16 core, 128  GB RAM, nvidia 3080 Ti
  
image with 373 MB full frame .  blurxterminator around 30s  , starxterminator around 25s


CS
Brian

Similar performance to mine.. BXT and SXT both being cuda accelerated, my 4080 does about the same BXT/SXT times.

I've got a Ryzen 9 9950X and 128GB RAM, and things like WBPP changed immensely when I upgraded to the Ryzen 9 from my old dual Xeon rig, but BXT/NXT/SXT didn't change almost at all jumping from my 1080 to a 4080, and no change at all when I upgraded the rest of my rig around that 4080.

With my new rig with the 5080 and a FF IMX455 image, BlurX takes about 18s and StarX about 15s.=flex flex-wrap sm:w-full=sm:w-3/4
Intel i9-14900KF , 96 GB [48 GB X2] DDR5-6000MHz, NVIDIA GeForce RTX 5080=flex flex-wrap sm:w-full=sm:w-3/42x4TB Samsung 990 Pro (one with the image data, the other the PI swapfiles)

For me bottleneck seem to be memory, I use full frame datasets with 6200MM, initialization take a thousand year and after that progress look not bad.