NVIDIA GeForce RTX 2080 Ti performance test

Let’s first take a look at the product appearance of 2080 Ti.

I won't go into detail about this part, just look at the pictures. What we got is a non-public version card from GIGABYTE, the specific product model is: GIGABYTE GeForce RTX 2080 Ti Windforce OC 11G.

Well, let's take a look at the actual photos of this product and compare it with the 1080 Ti. (To be clear, the 1080 Ti is the MSI 1080 Ti turbo fan used by Cortex Labs for deep learning. The specific product model is: MSI GeForce GTX 1080 Ti AERO 11G OC)

(2080 Ti outer packaging)

(1080 Ti and 2080 Ti real shot front comparison)

(Actual thickness comparison between 1080 Ti and 2080 Ti)

(Detailed parameters are from GIGABYTE official website)

After looking at the product, you need to prepare the test environment.

We used the SuperMicro SuperServer 4028GR-TR machine learning server available in the office as the test platform. The basic configuration of the server is as follows:

Motherboard: Super X10DRG-O+-CPU

CPU: intel® Xeon® processor E5-2600 v4†/ v3 family (up to 160W TDP) *

Dual Socket R3 (LGA 2011)

Memory: 2400 MHZ DDR4 SDRAM 72-bit

System: Linux Ubuntu 18.04

Then explain the mining algorithm tested:

Generally speaking, when testing the mining performance of GPU, the main choice is ETH's Ethash algorithm or ZEC's Equihash algorithm. However, mining machine manufacturers have developed and mass-produced ASIC mining machines for these two algorithms. However, due to the relatively high demand for video memory for the Ethash algorithm, the impact of Ethash ASIC mining machines on GPU mining is not very large, while the impact of Equihash ASIC mining machines on GPU mining is very large (A9, Z9 mini and other machines). Therefore, we choose ETH's Ethash algorithm for testing here.

Finally, let me introduce the selection of mining software:

We initially chose Claymore mining software, which is more familiar to miners, for testing. However, in actual testing, Claymore has not yet been perfectly adapted to 2080Ti. In the end, we chose EthMiner as the mining software.

All the preparations have been completed. We have officially installed both graphics cards on the server and installed the drivers.

The driver version of 1080 Ti is: NVIDIA-Linux-x86_64-390.87.run

The driver version of 2080 Ti is: NVIDIA-Linux-x86_64-410.57.run

(1080 Ti and 2080 Ti are already installed)

The exciting moment is coming soon!!!

After booting, we can see that both card drivers are normal.

(When checking GPU information in Linux system, the displayed information is incomplete)

(Another picture with complete information)

Running EthMiner, both cards can mine smoothly and display the local computing power. After stabilizing, we take a screenshot.

(GPU0 is 2080 Ti, GPU1 is 1080Ti)

From the figure we can see that the computing power performance of the two graphics cards under the Ethash algorithm is as follows when neither is overclocked:

1080 Ti Hashrate: 32.46 MH/s Power Consumption: 212W

2080 Ti Hashrate: 50.90 MH/s Power Consumption: 257W

Here we can also see that the improvement of 2080 Ti over 1080 Ti is quite obvious, with the computing power under the Ethash algorithm increased by 56.8%, while the power consumption only increased by 21.2%.

However, in the past, when using 1080 Ti for mining, there were actually many tools to optimize the computing power of 1080 Ti. However, there is no similar program for 2080 Ti at present. We choose a commonly used software to optimize the computing power of 1080 Ti and see how big the gap is.

(After using the 1080 Ti computing power optimization tool, the computing power performance of the two graphics cards)

After turning on the optimization tool, we can see that the computing power of GPU1 (that is, 1080 Ti) has increased from 32.46 MH/s to 45.19 MH/s, an increase of 12.73 MH/s. However, this optimization tool has no optimization effect on 2080 Ti. As a result, the computing power difference between 1080 Ti and 2080 Ti under the Ethash algorithm is only about 5.7 MH/s.

After testing this far, you may complain, is this the only performance improvement of 2080 Ti?

In fact, this is not the case. After all, the product cycle of 1080 Ti has entered its final stage. After so many years of development, countless people have repeatedly improved and optimized this product. However, not many people have received the 2080 Ti product yet, let alone optimized it. Although the final difference in our test is not very large, this is a contest between "the previous generation product optimized to the extreme" and "the new product is purely original", and it is only the performance under the Ethash algorithm. Therefore, this data test is only of reference significance under the specific algorithm at the current time.

Just after our operations team finished testing the mining performance, our AI team couldn't wait to start deploying machine learning tasks on the server. We implemented the following operations on the server at the fastest speed (the description comes from our AI R&D team): We used resnet50 to perform image classification tasks on the cifar10 dataset (10 categories, and resized the image to (224, 224)). In the latest cuda 10.0+pytorch 0.4.1+cuDNN 7.3 environment, we tested training + inference on the 2080ti and 1080ti graphics cards. The test results are as follows:

(1080 Ti running)

(The training time for 1080 Ti is about 280.66s, and the inference time is about 18.54s)

(2080 Ti running)

(The training time of 2080 Ti is about 235.30s, and the inference time is about 17.12s)

Through the above tests, we found that the 2080 Ti reduced the training time by 16.16% and the inference time by 7.7% compared to the 1080 Ti.

However, considering that the existing software support is not mature enough, the results this time are only a preliminary reference. We feel that this result is not satisfactory. Of course, we will have more in-depth tests in the future and share the test progress of our development team with friends in the whole community.