Initial Linux Benchmarks Of The NVIDIA TITAN RTX Graphics Card For Compute & Gaming
I've been hammering away benchmarks on the TITAN RTX for the past 24 hours so far and have some preliminary Linux performance data to pass along this holiday weekend... I should have some additional interesting benchmarks soon looking at other areas like the Blender rendering performance. Normally I re-test each graphics card in the comparisons as part of each new Linux GPU review on Phoronix but fortunately with having wrapped up some fresh tests earlier this month still using the NVIDIA 415 driver series and the still-stable Linux 4.19 kernel series, that data was used (as well as currently having an issue with the RTX 2070), so it was possible to deliver this TITAN RTX comparison in a timely manner.
Tests were done off the Intel Core i9 9900K system while running Ubuntu 18.04.1 LTS with the Linux 4.19.5 kernel. The GPUs tested for this GPU benchmark comparison included the GTX 980, GTX 980 Ti, GTX TITAN X, GTX 1070, GTX 1070 Ti, GTX 1080, GTX 1080 Ti, RTX 2070, RTX 2080, and RTX 2080 Ti FE. Note that the TITAN X is the Maxwell-based GM200 model and not the newer Pascal Titan X; the selection of graphics cards tested were based upon what I had available. For the OpenCL tests there are also the Radeon RX Vega 56 and RX Vega 64 cards for reference. I am also working on some Radeon ROCm 2.0 TensorFlow benchmarks against NVIDIA, but were not done in time for this initial TITAN RTX article but should be another fun comparison coming up shortly.
During this GPU benchmarking all of the tests were automated in a fully standardized and reproducible manner using the open-source Phoronix Test Suite benchmarking software. Additionally, the Phoronix Test Suite was polling and logging the GPU core temperatures as well as the overall AC system power consumption (fetched via a WattsUp Pro power meter) for being able to also deliver performance-per-Watt measurements in a reliable manner with PTS.
Following these preliminary GPU compute benchmarks are also some Linux gaming benchmarks, granted, the primary function of the TITAN RTX is not for gaming.
When kicking things off with TensorFlow using the NVIDIA GPU Cloud image, the TITAN RTX performance using FP16 came in at 16% faster than the already stunning RTX 2080 Ti Founder's Edition. Or compared to the GTX 1080 Ti, 91% faster. The TITAN RTX was 3.3x the speed of the Maxwell TITAN X for this ResNet-50 FP16 benchmark where the RTX cards' tensor cores come out to play.
Under this particular test, the overall AC system power consumption of the i9 9900K + TITAN RTX was 203 Watts on average with a peak of 350 Watts. That average power draw was actually 10 Watts lower than the RTX 2080 Ti FE while the peak power consumption was eight Watts higher.
On a performance-per-Watt basis, the TITAN RTX actually has a nice bump in efficiency over the other RTX cards.
At FP32 with ResNet-50, the gains for the TITAN RTX are far less dramatic, but still an improvement.