New GPU underperforming expectations

Help and Troubleshooting
1

I submitted a support ticket for this already, but thought I might check with the community.

I recently switched to a new laptop workstation purchased for it’s specs and ability to replace my previous system. From my understanding, it should have been able to render on the GPU maybe 30% faster than my previous system, but that hasn’t been the case so far in testing. It’s taking twice as long or longer.

Listed below are my specs:

Old workstation:
CPU: Intel® Xeon® Gold 5215 CPU @ 2.50GHz 2.50 GHz (X 2 processors)
GPU: Nvidia Quadro RTX 6000 (Turing) 24 GB VRAM
128 GB RAM
SSD hard drive

New workstation:
CPU: 13th Gen Intel Core i9-13950HX
GPU: Nvidia RTX 5000 Ada (laptop) 16 GB VRAM
128 GB RAM
SSD hard drive

Based on all of the comparisons and benchmarks I can find, the RTX 5000 should be faster than the older generation Quadro RTX 6000. But I am finding renders take approximately twice as long, if not longer.

Quadro RTX 6000: https://www.techpowerup.com/gpu-specs/quadro-rtx-6000.c3307

RTX 5000 Ada: https://www.techpowerup.com/gpu-specs/rtx-5000-mobile-ada-generation.c4097

Could you please provide me with suggestions for improving performance? Maybe it’s a driver issue? Any help would be appreciated.

2

I’ve also noticed, during rendering, the GPU is only using 8.3 GB of the 16 GB of VRAM available. Not sure if this has anything to do with it, but maybe…

The old GPU renders the scene only using 33% of it, in task manager. Also, the temperature goes up to 77 degrees C.

The new GPU uses between 90 and 100% and the temperature never goes above 53 degrees.

3

So if I’m correct you had a desktop and now you’ve a laptop.

Based on it’s CUDA cores I would indeed say that it should even be close to twice as fast. But that’s in theory especially with laptops and a laptop card is not comparable with a desktop card since they have a totally different power profile.

While your laptop has really awesome specs the i9-13950 and the RTX5000 are both power hungry animals. In a laptop that power converts to heat and a laptop has no way to cool that generated heat so the GPU will start throttling. Only the GPU can draw around 190W NVIDIA RTX 5000 Ada Generation Laptop GPU GPU - Benchmarks and Specs - NotebookCheck.net Tech

I’m sure you can download some tools so you can allow the GPU to get hotter before it actually starts to throttle. From the link above I see they mention you can configure the card from 80 to 175W so if you limits it’s TDP you get less performance (but also less heat).

If I look at V-Ray benchmarks the desktop RTX 6000 24GB sits around 940 points.
The newer mobile RTX 5000 scores in V-Ray just 780 points.

That’s quite a terrible difference since the performance is less while the CUDA cores are double. On the other hand, your desktop RTX 6000 could use almost two times as much power than the RTX5000. Nothing is for free in life.

You can overclock your card easily using some tool like MSI Afterburner but I think you can’t expect it to actually perform as well as your desktop RTX 6000. Heat will be the issue and it will throttle the card anyway, so an overclock can also make the card run slower on average since it runs too hot.

Task manager doesn’t show the CUDA core usage of your GPU so that’s misleading, if you would select that manually to CUDA you will see 100% or really close to hundred. You can change the graphs you see using the little down arrow.

My personal opinion is that laptops in this scene are a waste of money. A laptop can be great to work remote but I would never use it to render. For the same money you build a really simple PC with a 4090 and that will be 4.5 times as fast (around 4200 score).

The main problem with every laptop is heat and because of it all hardware has different profiles since a GPU can’t simply use 450W in a laptop let alone getting rid of all the heat. So a 4090 for a laptop has a much lower TDP and won’t perform like a desktop version.

Also why I’m not a fan of the current benchmark of KeyShot and I use the V-Ray benchmark results here. The KeyShot benchmark runs a scene till a certain amount of samples (I think) and records the time taken. Better would be to run a certain scene for one minute and see how much samples are done.

Sounds like the same but a laptop GPU will already start throttling within 15 seconds. With current benchmark the test is over before any fast GPU gets hot. It’s simply too short.

Sorry for the long story and bringing you bad news but I don’t think there’s anything you really can do about it. If you could return it I would. And don’t let them sell you a professional series card, they are not worth it for GPU rendering. It’s like paying twice for half the speed (well not half, but less).

2 Likes
4

Thanks for the response!

Some more data points:
The GPU in the laptop is never going above 53 degrees C. Thus, thermal throttling is unlikely.
I cannot find a CUDA dropdown in task manager. Not sure why, but I don’t see one.

Also, I rendered the same scene on another Lenovo laptop with a 4090 and it’s rendering the scene in 9 minutes. That’s half the time the desktop took and only 25% of the time the new laptop with the RTX 5000 Ada took.

And this RTX 4090 is supposed to be very similar if not better than the 4090. Something is definitely wrong. But we’re struggling to figure it out.

5

Just wanted to provide another update.

After some internal testing, we put Keyshot on another laptop in the building exactly like my new one with the RTX 5000 Ada GPU and it rendered the scene in 7 minutes. (2000 samples)

That’s more than twice as fast as the original desktop workstation with the Quadro RTX 6000 which took 18 minutes.

Long story short, there seems to be a problem with my new Lenovo laptop which took over 40 minutes and never go hotter than 53 degrees. The other laptop with the same RTX 5000 Ada immediately ramped up to over 70 degrees. Definitely a hardware issue.

It is pretty remarkable that these laptops with the new 4090 Ada based GPUs are so powerful! Happy to know that something is wrong and I was reading the specs right.

1 Like
6

Schermafbeelding%202023-10-18%20215852
Schermafbeelding 2023-10-18 215852.png794×327 13.1 KB

Well that’s weird, it used to be in this pulldown menu. I just installed fresh drivers but it really was there earlier. But if you want to know all sensor things of your GPU and if it’s actually working hard: GPU-Z Graphics Card GPU Information Utility (techpowerup.com)

That’s a really nice tool. I also use if for example to tame my cards a bit so they have same performance but use less watt.

I can imagine you think something is wrong, I would think exactly the same based on the number of CUDA cores which, in general, is quite a nice indication of the expected render performance.

Since NVidia is playing tricks with using the same RTX 5000 name for really different cards I hope I’m wrong. Looking at this makes more sense: Quadro RTX 6000 vs RTX 5000 Mobile Ada Generation : Which one is better? | TopCPU.net

Not sure how floating point calculations are in parallel with rendering but looking at the CUDA cores it’s more the performance you would expect. Instead of the stats from the V-Ray benchmark which might be the old RTX 5000.

I would start some renders while looking at the sensor tab from GPU-Z. You’ll see the GPU Load at 100% and it also shows all the temperature sensors if there are multiple. If the GPU Load ain’t 100% and you could check the clock speeds at the top of the window. They should be around the clocks mentioned in the specifications of the card. If not it could be something is holding back the GPU and put it into a different mode.

That’s basically what Windows does with is power/performance options. Especially with a laptop it’s balanced setting could be holding things back although KeyShot should make it run in highest performance mode via de drivers. But you could change it to ‘maximum performance’.

That most of the time also let the CPU run at high clocks all the time which you can manually adjust. No need for another little heater inside the laptop.

If the clocks are still low it might be a weird driver issue. Than I would just download the latest Creative Studio drivers from NVidia and select custom installation, install only the drivers and no GeForce Now things and put a checkmark at ‘clean install’. And do a quick restart afterwards to be sure it’s all fresh.

(if you do AI image generation as well the new Game Ready drivers might be more interesting since they accelerate image generation / torch).

Hope you get a bit further and I think/hope I was totally wrong about it’s actual performance because of the older RTX 5000 and benchmark results.

7

I was typing meanwhile :wink:

But great it showed I was wrong! Makes really much more sense as well looking at the cuda cores which was the thing that confused me as well.

Like mentioned, I would just try to do a clean install of drivers again. And if you really want it clean you can do a save boot and run DDU Display Driver Uninstaller (DDU) download version 18.0.6.8 (guru3d.com)

But I think currently NVidias option ‘clean installation’ currently works better than it did in the past.

Happy rendering!