Table of Contents
Some of the technical data of the new NVIDIA Turing architecture for the new GeForce RTX graphics cards is leaked.
The new NVIDIA Turing architecture for gaming graphics cards that will support Ray Tracing technology, is a great unknown or was, since a part of the technical documentation regarding this architecture has been leaked. According to the data, this is a preview of the documentation that will be available from September 14. This preview contains the most interesting facts.
Simultaneous execution of integral and floating point instructions (INT32)
INT32 execution units have been added to allow the execution of floating point and non-floating point processes and instructions in parallel. According to NVIDIA, this improvement should provide an improvement of around 36% over floating point operations.
Parallel execution is achieved thanks to a unified architecture for L1 shared memory and texture caching. According to the company the INT32 / FP32 cores among other multiprocessing transmission systems allow 'a performance improvement of 50% with respect to each CUDA core)
Shading Advancements
- Mesh Shading: new shading model for vertex, tesselation, geometry shading (more objects per scene)
- Variable Rate Shading (VRS) / Variable Rate Shading - Developer control over shading rates (to limit shading where it does not provide visual benefit)
- Texture-Space Sharing / Shared Texture Space - Stores shading results in memory (no need to duplicate shared work for processes)
- Multi-View Rendering (MVR) - Extends the Pascal Single Pass Stereo to multiple views in a single pass.
Memory compression in the Turing architecture
Turing offers new memory compression techniques that prevent quality loss. These improvements are achieved by 'state of the art' algorithms over Pascal that offer a '50% increase in effective bandwidth in Turing compared to Pascal'
Video and display engine
New special video engine to support DisplayPort 1.4a that supports 8K @ 60Hz. The new Turing-based graphics supports two 8K @ 60Hz monitors via DisplayPort or USB-C, incorporating the improved NVENC codec that allows encoding H.265 in streaming at 8K @ 30FPS and also the new NVDEC decoder that supports HEV YUV444 10 / 12b HDR , H.264 8K and VP9 10/12 HDR.
NVLINK two-way
The TU102 silicon has two 8nd Gen NVLINK x2 ports while the TU104 offers only one x8 link. TU106 silicon does not have NVLINK support. The SLI connector and the possibility of creating configurations of more than two graphics cards disappear.
Regarding the characteristics TU102 practically duplicates the specifications of TU106. The TU104 silicon has a chip that has four TPCs per cluster unlike the TU102 and TU106 that have six TPCs per cluster.
|
NVIDIA TURING GPUs |
|||
| TU102 | TU104 | TU106 | |
| Manufacturing Node | 12nm FFN | 12nm FFN | 12nm FFN |
| Die Size | 754 mm2 | 545 mm2 | 445 mm2 |
| Transistors | 18.6 Billion | 13.6 Billion | 10.6 Billion |
| NVIDIA SKU w / full chip | Frame RTX 6000 | Frame RTX 5000 | GeForce RTX 2070 |
| CPGs | 6 | 6 | 3 |
| TPC's | 36 | 24 | 18 |
| SMS | 72 (12 per CPG) | 48 (8 per CPG) | 36 (12 per CPG) |
| Tensor Cores | 576 | 384 | 288 |
| RT Colors | 72 | 48 | 36 |
| FP32 Cores (CUDAs) | 4,608 | 3,072 | 2,304 |
| INT32 Cores | 4,608 | 3,072 | 2,304 |
| ROPs | 96 | 64 | 64 |
| TMU's | 288 | 192 | 144 |
| Memory Interface | 384-bit | 256-bit | 256-bit |
| L2 Cache | 6144 KB | 4096 KB | 4096 KB |