Testing of Palit GeForce RTX 2080 Super in machine learning tasks

Nvidia plans to update its graphics cards, transferring the entire product line to the new GDDR6 memory. In our review of the Geforce GTX 1660 Super, we already mentioned that in Machine Learning / Artificial Intelligence tasks, VRAM throughput is often crucial. And of course, if you choose a GPU not only for games, but also for scientific calculations, you are interested to see a video card with tensor cores and modern GDDR6 memory of 8 GB.


We are considering a new product from the company Palit. This manufacturer of video cards today is very popular with gamers, thanks to the good overclocking potential and quiet cooling system. By the way, the high TDP of the RTX2080 Super video card forces the use of thick heat-conducting tubes and two 4-pin power connectors.

Keeping the noise level low is a challenge, but here's the interesting thing: in all our tests, I failed to warm up the GPU properly. Apparently, many blocks of the video chip are not involved in computational calculations in contrast to the same GTX 1660, which easily reached its design capacity in Tensorflow tests.

Testing

The first part is synthetic tests, and we start by evaluating integer and floating-point operations.

Let’s continue with Geekbench 5 with tests using basic facial recognition algorithms.

Almost in the test Octane Bench 4.0, using a new rendering engine, the novelty shows very good results for a single GPU. Under OpenCL, the video card in question has only 30-40% lower speed than the Tesla V100.


Let’s move on to testing in real problems and measure the speed in the most popular Tensorflow / Keras framework.

Let’s start with the simple tests included in the keras package examples. We will compare with Tesla-mi provided in Google Colab. And while it’s undeniable Google shares GPU performance, it’s important for us to understand how much your on-premises GPU is comparable to what you’re given in the cloud.

In the real problem of training the model with the help of Markov chains, known as Textgenrnn with a source file of 2.59 MB and a Butch size equal to 256, the Board is 4 times faster than the GTX series and 2 times faster than the Tesla P100 variant, which by great luck can fall to you at night, during a low load on the Google Colab service. With this performance, Palit RTX2080 Super is a real and reasonable alternative to cloud computing in terms of 1 GPU.


We estimate the tasks of generating hashes by working in the mining client Minergate on Etherium. Here, the video card shows remarkable results, but at the current rate of the cryptocurrency works at a loss.

Conclusions

If the cloud is insecure for you and Tesla is expensive, then the Palit RTX 2080 Super is a great alternative for working in Tensorflow and other Python tasks involving machine learning and artificial intelligence. Very quiet, cool and thanks to FP16 support, more productive than previous generations of GPUS.

Before buying, make sure that you have a powerful enough power supply and a well-ventilated case. Most Telecom enclosures that are not optimized for GPU installation will not fit this graphics card, but universal 4U / Rack enclosures for workstations will fit without problems.