After a lot of fuss and leaks, Nvidia finally revealed their next-gen architecture that will provide the industry’s first Ray tracking abilities. To no surprise the new architecture is called the Turing architecture and Nvidia’s CEO Jensen Huang officially revealed the power of the RTX Ray tracing cores during their presentation at the SIGGRAPH graphics conference. They demoed the first Quadro GPU based on the Turing, the beast of a GPU has enough computing power that it can render the raytraced scenes efficiently in real time.
The new Turing architecture is a lot different than the last gen Pascal architecture, but it shares the tensor cores that Nvidia introduced with their Volta architecture. Unlike Pascal architecture, the Turing architecture is a revolutionary step towards ‘hybridness.’ Nvidia essentially added their Tensor Cores, new RT cores, and the traditional Shader & compute cores on a single huge die which appears to be GT 102 or GT 100. The sheer size of the die which is 754mm2 is making us think that it may have been the latter, but we cannot say as things are different for Nvidia this time and they can change their nomenclature anytime now. The massive die houses 18.6 Billion transistors, and the base clock speed of the cores is now 1.75GHz which is a huge leap over Pascal’s GPUs. The rumored GDDR6 memory has also found its place with a 384-bit memory bus design that is clocked at 14Gbps with the net bandwidth of 672GB/s.
The highlight of the new Turing architecture is the ray tracing abilities of the new ray tracing cores which are called RT cores. These RT cores enable real-time ray tracing of the objects and environments with physically making their precise shadows, accurate reflections, refractions and most importantly perfectly simulating global illumination.
The new Tensor cores help to accelerate the deep neural networks learning and inferences that are critical for AI-based applications and services. The improved NVIDIA NVLink combines two GPUs with a high-speed link to perfectly scale the memory capacity up to 96GB for the high-end Quadro RTX cards.
Lastly, the Turing architecture is made to cater to the needs of the ever-changing world of technology, most of our appliances have accepted the USB type-C and virtual link interface the new Turning GPUs will also be fully compatible with these interfaces to cater the needs of display and bandwidth demands of a VR set with a single USB C connector.
Quadro RTX series
The first Graphics cards to house the new Turing architectures are the new Quadro RTX series, in fact, the demo that Nvidia ran during their SIGGRAPH graphics conference was running on a single Quadro RTX Graphics card. They did not specify which model it was, but it can be safely said that it could be the high-end model. At first, there will be three different versions of the Quadro RTX series which will be available in October.
The high-end model is called Quadro RTX 8000 which will have massive 48GB of GDDR6 memory made by Samsung. The GPU itself is made with 12nm process.
However, the die size of this GPU is not revealed. The CUDA core count is 4,608, while the Tensor core count is 576 and lastly the Ray tracing spec is 10 GigaRays/s. Moreover, the GDDR6 module has a memory bus of 384-bit combined with 48GB size makes the total memory bandwidth of 672GB/s. According to initial information, the required TDP for this Graphics card to run will be 225W.
The Graphics card that can be thought of as the higher mid-end of the Quadro graphics the Quadro RTX 6000 may be the Graphics card that Nvidia used during their demo. The massive 754mm2 is made with 12nm process and based on the new Turing architecture. The Core count is similar to the RTX 8000 while the base clock speed is 1.75GHz.
The Single precision boost of this card is rated at 16TFLOPs while the Tensor performance is 500TOPs (INT4) which means that it can process 500trillion tensor operations in one second and it will use Integer points rather than floating points. The integer points are 4x more precise and thus require 4x throughput. That is why the comparison with the current non-consumer grade flagship Graphics card Nvidia Quadro GV 100 is somewhat deluded. However, the new Graphics card is significantly faster.
Lastly, the ‘cheap’ alternatives of these Graphics card the Quadro RTX 5000 will be packing 16GB of GDDR6 memory made by Samsung’s 10nm system module. The CUDA cores are reduced to 3,072 while the Tensor core count is 384 in this Graphics card, and the Ray Tracing spec is 6 GigaRays/s. The memory module is 256-bit making the total memory bandwidth to be 448GB/s. The total TDP of this card is speculated to be the only 160W.
All of these Graphics cards are bound to be released during October, stay tuned for more information on the new Turing architecture and importantly on the new gaming Graphics cards.