Compare the deep learning performance of NVIDIA RTX 2080 Ti with GTX 1080 Ti · Titan V · Tesla V100

For deep learning , "GPU" which is good at parallel computation than CPU is often used. Gaming graphics boards are often used for deep learning processing, so Lambda , which sells deep learning server equipment, is using the latest graphic board " GeForce RTX 2080 Ti " & " RTX 2080 " In order to measure the deep learning performance, we conducted comparison tests with existing GPU products such as GTX 1080 Ti · Titan V · Tesla V100, and we are publishing results such as cost performance.

2080 Ti Tensor Flow GPU benchmarks - 2080 Ti vs V 100 vs 1080 Ti vs Titan V
https://lambdalabs.com/blog/best-gpu-tensorflow-2080-ti-vs-v100-vs-titan-v-vs-1080-ti-benchmark/

In September 2018 NVIDIA announced the RTX series which replaces the conventional "GTX" series as a new GPU introducing the Turing GPU architecture and releases the flagship model "RTX 2080 Ti" at the same time as "RTX 2080" Did.

In order to investigate how much the deep learning performance has improved for RTX 2080 Ti & RTX 2080 which has evolved greatly as a gaming performance, such as real time ray tracing support, Lambda has developed "RTX 2080 Ti", "RTX 2080", "GTX 1080 Ti "," Titan V (Volta) "and" Tesla V 100 ", we compared processing performance of Google's machine learning library" TensorFlow ".

All the tests are using "single GPU system" with one GPU, the motherboard is ASUS Prime B 360-Plus, the CPU is Intel Core i 7-8700 K, the CPU cooler is Noctua L-Type Premium , The machine " Lambda Quad Basic " equipped with Intel Xeon E5-1650 v4 only Tesla V100 is used under the condition of 32GB DDR4 memory, case NZXT H500, power supply EVGA SuperNOVA 750 G2 (750W).

The benchmark code used for the test is published on the following GitHub page.

GitHub - lambdal / lambda - tensorflow - benchmark
https://github.com/lambdal/lambda-tensorflow-benchmark

The results of FP32 throughput (the number of training samples processed per second) are as follows. The red line is the baseline of GTX 1080 Ti.

Comparing the average speed with GTX 1080 Ti, the RTX 2080 is nearly equivalent, but the RTX 2080 Ti is thin to Titan V.

The average speed in FP16 looks like this. RTX 2080 Ti exceeds Titan V.

While the North American price of RTX 2080 Ti is 999 dollars (about 110,000 yen) and RTX 2080 is 699 dollars (about 79,000 yen), the Tesla V100 is $ 9800 (about 1.1 million yen), which is a mere processing speed It is not useful to compare performance with the user's eyes. So, Lambda also graphs performance performance divided by price for cost performance.

The FP32 performance per dollar is RTX 2080 Ti top, followed by RTX 2080 and GTX 1080 Ti.

Even in FP16 performance the ranking has not changed.

Regarding the above results, Lambda stated that "Tesla V100 is like Bugatti Veyron, in exchange for overwhelming performance, few people will pay not only premium vehicle prices but maintenance costs and insurance premiums" It is. On the other hand, if you compare RTX 2080 Ti to Porsche 911, it expresses that you can get high performance at a (relatively) cheap price.