Nvidia Turing Architecture Details: TU102, TU104, TU106 Cores According To White Paper

Nvidia Turing architecture powers the upcoming Nvidia RTX 20 series graphics cards that are going to be coming out a few days from now. Here we are going to share some new information regarding the Nvidia Turing architecture that you might be interested in if you are looking forward to getting one of these upcoming graphics cards but are waiting on further information.

Key Features of Nvidia Turing Architecture

The following are the salient features of the Nvidia Turing architecture that will power the upcoming Nvidia RTX 20 series graphics cards.

INT32 Cores

One of the major features of the upcoming Nvidia Turing architecture is that the cards based on this architecture are able to execute floating point and non-floating point processes in parallel. According to Nvidia, this will allow the graphics cards to provide 36% better performance when it comes to floating point operations.

Nvidia also claims that this implementation allows for the graphics cards to perform 50% better per CUDA as compared to the previous graphics cards. That is very interesting indeed and is something that we will have to test ourselves.

New Shading

There are a couple of new shading features as well that you might be interested in. The Nvidia Turing architecture comes with Mesh shading which means better vertex, tesselation, geometry shading. Variable Rate Shading is also a feature and this means that the developer can control shading. If there are no visual benefits in a scene then the developer can limit shaders for the scene with this feature.

Another feature worth mentioning is Texture-Space Sharing. this allows for storing shader results in memory and no longer will it be duplicated for processes. The new architecture also offers Multi-View Rendering. This extends Nvidia Pascal’s Single Pass Stereo to multi-views in a single pass.

TU102, TU104, TU106 Core Specifications

According to the specifications of the core that we have right now, it seems that the RTX 2070 will be a mid-range game rather than a higher end graphics card. We know that the RTX The specifications of the TU102, TU104, TU106 are included in the table below:

TU106 TU104 TU102
Chip 10.6 billion transistors, 445mm² die,  12nm by TSMC 13.6 billion transistors, 545mm² die, 12nm by TSMC 18.6 billion transistors, 754mm² die, 12nm by TSMC
Specifications 3 raster engines, 36 shader clusters, 2304 shader units, 144 TMUs, 36 RT cores, 288 tensor cores, 64 ROPs, 4 MB level 2 cache, 256 bit GDDR6 memory interface 6 raster engines, 48 shader clusters, 3072 shader units, 192 TMUs, 48 RT cores, 384 tensor cores, 64 ROPs, 4 MB level 2 cache, 256 bit GDDR6 memory interface 6 raster engines, 72 shader clusters, 4608 shader units, 288 TMUs, 72 RT cores, 576 tensor cores, 96 ROPs, 6 MB Level2 cache, 384-bit GDDR6 memory interface
NVLink & SLI No NVLink NVLink 2.0 x8 port (50 GB / sec), support for up to 2x SLI two NVLink 2.0 x8 ports (100 GB / sec), support for a maximum of 2x SLI
Products GeForce RTX 2070 (fully equipped) GeForce RTX 2080  & Quadro RTX 5000 GeForce RTX 2080 Ti & Quadro RTX 6000/8000


Block Diagrams

The block diagrams of all the cores that are based on the Nvidia Turing architecture are included below:

Let us know what you think about the Nvidia Turing architecture and whether or not you are interested in getting an Nvidia RTX 20 series graphics card when they come out in the next few days.