Initial Linux Benchmarks Of The NVIDIA TITAN RTX Graphics Card For Compute & Gaming

Yesterday I unexpectedly found my hands on a NVIDIA TITAN RTX graphics card as the company's newest Titan graphics card built upon the Turing architecture and is now available via retail channels at $2499 USD. Here is an initial look at the NVIDIA TITAN RTX performance under Ubuntu Linux with a variety of compute workloads (including TensorFlow) as well as for entertainment are some Vulkan gaming benchmarks.

Simply put, the TITAN RTX is a beast for GPU compute workloads. While costing $2500 USD, the TITAN RTX offers 4608 CUDA cores compared to 4352 on the RTX 2080 Ti and 576 tensor cores compared to 544 on the RTX 2080 Ti while keeping to the same 1350MHz base clock speed but offering a 1770MHz boost clock (the 2080 Ti carries a boost clock of 1545MHz or 1635MHz on the Founder's Edition GPU). Not only is the TU102 operating at its full potential with the TITAN RTX, but there is also 24GB of GDDR6 video memory -- more than twice that of the RTX 2080 Ti or three times that of the RTX 2070 and RTX 2080 series, making it much more capable for memory intensive computations on the GPU.

Over the RTX 2080 Ti, the TITAN RTX also offers a 384-bit memory interface compared to 352-bit and 72 RT cores compared to 68. At the default frequencies the TITAN RTX is rated to provide an incredible 672 GB/s of video memory bandwidth.

All this GPU compute power of the TITAN RTX necessitates a 280 Watt TDP rating, only modestly higher than the 250 Watts of the RTX 2080 Ti. As such, dual 8-pin PCIe power connectors are required for sufficient power.

If running a display setup off the TITAN RTX, the card has three DisplayPort, one HDMI, and one USB Type-C (for VR VirtualLink) connections. There are more than 18.6 billion transistors making up this TU102 GPU.