With the Titan RTX, internally called T-Rex, Nvidia releases a full version of the TU102 chip with Turing architecture. The graphics card is aimed more at developers than well-to-do gamers, but is less expensive than the previous Titan V with Volta GPU.
Nvidia has introduced the Titan RTX, the new top of the line RTX graphics card with Turing architecture. The manufacturer plans to launch the model in December 2018, and with 2,500 US dollars before tax (2,700 euros) the Titan RTX will cost less than the previous Titan V with 3,000 US dollars (3,200 euros). The technical basis of the new graphics card is a TU102 chip in its full version.
This is already used in the Quadro RTX 8000 for 10,000 US dollars, while the Geforce RTX 2080 Ti (Test) uses a trimmed version. All 4,608 shader units are available for the Titan RTX, including 576 tensor cores and 72 RT cores. The 384-bit interface is equipped with 24 GB of GDDR6 memory, as is the case with the Quadro.
According to the manufacturer, the theoretical FP16 computing power is rounded to 130 teraflops, which results in a GPU clock rate of 1,770 MHz boost – slightly more than the Quadro RTX 8000. The video memory achieves a data transfer rate of 672 GByte/s, so it runs at 14 GBit/s like the Quadro RTX 8000 or the Geforce RTX 2080 Ti. Nvidia sees the Titan RTX as a cheaper alternative to the Quadro RTX, as it has a very similar speed – although as usual the drivers are somewhat reduced in functionality.
Compared to a Geforce RTX 2080 Ti the performance increases by almost 15 percent, but with 24 GByte instead of 11 GByte the much larger memory is an important factor for machine learning or scientific calculations in general. The Titan V has only 12 GB instead of 16 GB, but thanks to HBM2 the bandwidth is hardly lower than that of the Titan RTX. With the Teraflops, the Titan V is slightly ahead, but the Titan RTX is practically faster because it uses faster ALUs per clock. The board power is 280 watts.