In this short blog post, we are going to show benchmarking results of the latest RTX 2080ti. We also will try to answer the question if the RTX 2080ti is the best GPU for deep learning in 2018?
We use Ubuntu 18.04 with CUDA 10.0, Tensorflow 1.11.0-rc1 and cuDNN 7.3.
We only had a GTX 1080 Ti to compare. In all benchmarks we used the same hardware and software configurations, we just swapped the gpu cards. For the test, we will use FP32 single precision and for FP16 we used deep-learning-benchmark.
FP32 throughput
The throughput is the number of training samples processed per second. We will use some common models using synthetic data.
We found that the 2080ti is 27% in average faster than the 1080ti on FP32 and more expensive. We expect better performance improvement with FP16.
FP16 throughput
For some networks FP16 could be enough, some nvidia cards has faster computational unitfs for FP16 like the V100 processor.
In average the 2080ti was 38% better than the 1080ti in FP16 benchmark.
Hardware configuration
For this benchmark we are going to use the following hardware:
- CPU: Intel Core i7-8700K
- RAM: Corsair LPX 32GB DRAM 3000MHz C15
- Motherboard: Gigabyte Z370 AORUS
- GPU: ZOTAC GAMING GeForce RTX 2080 Ti AMP 11GB GDDR6 352-bit
- OS hard drive: Samsung 970 EVO 250GB - NVMe PCIe
- Data Harddrive: WD Blue 1TB SATA 6 Gb/s 7200 RPM
- Powersupply: Seasonic Focus Plus 750 Platinum
- Case: CORSAIR Carbide 100R
- CPU Cooler:Noctua L-Type Premium Quiet CPU Cooler