NVIDIA GeForce RTX 2080 Ti To GTX 980 Ti TensorFlow Benchmarks With ResNet-50, AlexNet, GoogLeNet, Inception, VGG-16
First up was ResNet-50 at FP16 where we can immediately see the performance impact of this new high-end Turing graphics card... 67% faster than the GTX 1080 Ti Pascal graphics card. The other graphics cards are interesting for reference purposes.
While there has been some concern about the reference cooler on the GeForce RTX 2080 Ti, while running this ResNet-50 benchmark the average GPU core temperature was just 47 degrees with a peak of 56 degrees -- lower than both the GTX 980 Ti and GTX 1080 Ti... Though if running multiple GeForce RTX graphics cards in a single chassis the temperature may become more concerning.
The peak AC system power consumption on this system paired with the RTX 2080 Ti was 336 Watts compared to 312 Watts with the GTX 1080 Ti. But interestingly for this ResNet-50 model the average power consumption was about 20 Watts lower on the RTX 2080 Ti than the previous-generation Pascal card.
On a performance-per-Watt basis with ResNet-50 with FP16 precision, the RTX 2080 Ti does extremely well.
With FP32 precision for ResNet-50, the advantage of the Turing card is much less but still a 35% improvement over the GTX 1080 Ti. Not all of the cards were able to successfully run this test configuration due to video memory pressure.
With FP32 precision the average power consumption was higher than the GTX 1080 Ti along with the peak power consumption.
But on a performance-per-Watt basis, the RTX 2080 Ti still came out with a 30% advantage over Pascal.