Today, NVIDIA has announced its latest release using the Turing architecture with its Titan RTX. This follows the heels of the Quadro RTX that was announced in August. NVIDIA positions the card more for the prosumer than a consumer/gaming card but will still manage to be the fastest card that doesn’t use the Quadro naming (and a lot more expensive). The Titan RTX will cost $2499 and will be available later this month.
Titan RTX
If we take a look at the specifications, the Titan RTX looks like like a combination of RTX 2080 Ti, and RTX 6000 from the Quadro line. The new GPU is based on the TU102 GPU (same as 2080 Ti), but uses the full implementation of the silicon whereas the 2080 Ti is the cut down version.
| Titan RTX Specifications | |||
| Titan RTX | RTX 2080 Ti FE | Titan V | |
| CUDA Cores | 4608 | 4352 | 5120 |
| Tensor Cores | 576 | 544 | 640 |
| RT Cores | 36 | 36 | N/A |
| GPU Base Clock | 1350MHz | 1200MHz | |
| GPU Boost Clock | 1770MHz | 1635MHz | 1455MHz |
| VRAM | 24GB | 11GB | 12GB |
| Memory Interface | 384-bit | 352-bit | 3072-bit |
| Memory Clock (Data Rate) | 14Gbps GDDR6 | 1.7Gbps HBM2 | |
| Memory Bandwidth | 672GB/s | 616GB/s | 653GB/s |
| Singe Precision | 16.3 TFLOPS | 14.2 TFLOPS | 13.8 TFLOPS |
| Double Precision | 0.51 TFLOPS | 0.44 TFLOPS | 6.9 TFLOPS |
| Tensor Performance (FP16 w/FP32 accel.) | 130 TFLOPS | 57 TFLOPS | 110 TFLOPS |
| Transistor Count | 18.6B | 21.1B | |
| L2 Cache Size | 6MB | 5.5MB | 4.5MB |
| TDP (Watts) | 280W | 260W | 250W |
| Price | $2499 | $1199 | $2999 |
One of the major differences between the 2080 Ti and the Titan RTX (outside of its gold “make no mistake about it” color scheme) is the memory capacity, a whopping 24 GB, up from 11 GB on the 2080 Ti. The larger memory buffer will allow for much larger datasets for compute users and data scientists. For gamers, much of it would be left unused.
The tensor cores are the meat and potatoes of these cards which rides its success from the AI and neural networking areas with the Geforce based cards having them, but limitations were put in place. In GF cards, using the FP16 mode, the Turing architecture is able to accumulate at FP32 for higher precision but is limited to half-speed throughput. That limitation is gone for the Titan RTX which is capable of FP32 accumulation throughput on its tensor cores.
Titan RTX – TRex
The new Titan RTX, or T-Rex as NVIDIA adoringly calls it, will be available sometime this month.
Joe Shields (Earthdog)
