Experiences using an RTX 5080 with Resnet50 / Efficientnet_b3 / Inception_v3?

1

I’m looking for some info to compare to my own observations of an RTX 5080 I bought a couple weeks ago.

The architectures noted in the title and batch size 16 or 32.

What does ‘nvidia-smi dmon’ show for power, SM% for you?

–

My 5080 card is drawing between 225W and 270W depending on the architecture being trained, and shader utilization is higher than 95%. It is ~1.7x the speed of a 3080/12G in terms of time to process an epoch’s worth of files.

In the past, using RTX 3080 or A6000 or RTX 2080ti, the card would ramp up to near max power draw. This 5080is rather demure by comparison.

Imagine it! An RTX card that does not want to eat the world! It’s great.

Training ResNet50, Efficientnet_b3, Inception_v3 on Imagenet, 224x224 images, batch size 16 or 32.

Thanks!

2

Hi Emerth,

Thanks for sharing your experience with the RTX 5080! It’s interesting to see how it compares to previous models. Your power draw numbers are pretty consistent with what I’ve observed, too. I’ve seen similar ranges while training with ResNet50 and EfficientNet, and like you mentioned, the shader utilization is impressively high.

I also noticed that the 5080 doesn’t ramp up power as aggressively as the previous cards, which is refreshing. It seems to balance performance and power efficiency quite well.

3

It makes me wonder what would be possible with a similar product that did not contain hardware for ROPs, ray tracing support, image scaling support. Just all cuda shaders.