Architecture

On the 14th of September, we published a comprehensive NVIDIA "Turing" architecture deep-dive article including coverage of its three new silicon implementations and the new RTX Technology. Be sure to catch that article for more technical details.


The "Turing" architecture caught many of us by surprise because it wasn't visible on GPU architecture roadmaps until a few quarters ago. NVIDIA took this roadmap detour over carving out client-segment variants of "Volta" as it realized it had achieved sufficient compute power to bring its ambitious RTX Technology to the client segment. NVIDIA RTX is an all-encompassing, real-time ray-tracing model for consumer graphics, which seeks to bring a semblance of real-time ray tracing to 3D games.


To enable RTX, NVIDIA has developed an all new hardware component that sits next to CUDA cores, called the RT core. An RT core is a fixed-function hardware that does what the spiritual ancestor of RTX, NVIDIA OptiX, did over CUDA cores. You input the mathematical representation of a ray and it will transverse the scene to calculate the point of intersection with any triangle in the scene. This is a computationally heavy task that would have otherwise bogged down the CUDA cores.

The other major introduction is the Tensor Core, which made its debut with the "Volta" architecture. These too are specialized components tasked with 3x3x3 matrix multiplication, which speed up AI deep-learning neural net building and training. Its relevance to gaming is limited at this time, but NVIDIA is introducing a few AI-accelerated image-quality enhancements that could leverage Tensor operations.


The component hierarchy of a "Turing" GPU isn't much different from its predecessors, but the new-generation Streaming Multiprocessor is significantly different. It packs 64 CUDA cores, 8 Tensor Cores, and a single RT core.

TU102 Graphics Processor

The TU102 is the largest silicon based on the "Turing" architecture and powers the GeForce RTX 2080 Ti. It's also the biggest GPU die NVIDIA ever built, with over 18.6 billion transistors sitting on a 775 mm² die that has been fabricated on the 12 nanometer process by TSMC. As we mentioned earlier, the essential component hierarchy on the "Turing" architecture hasn't changed. What has changed, however, is that the Streaming Multiprocessor (SM), the indivisible sub-unit of the GPU, now packs CUDA cores, RT cores, and Tensor cores, orchestrated by a new Warp Scheduler that supports concurrent INT and FP32 ops, which should improve the GPU's asynchronous compute performance.

At the topmost level, the GPU takes host connectivity from PCI-Express 3.0 x16, an NVLink interface, and connects to GDDR6 memory across a 384-bit wide memory bus. On the RTX 2080 Ti, the memory interface is narrowed to 352-bit and wired to 11 GB of memory. The GigaThread engine marshals load between six GPCs (graphics processing clusters). Each GPC has a dedicated raster engine and six TPCs (texture processing clusters). A TPC shares a PolyMorph engine between two SMs. Each SM packs 64 CUDA cores, 8 Tensor cores, and an RT core.

There are, hence, 768 CUDA cores, 96 Tensor cores, and 12 RT cores per GPC, and a grand total of 4,608 CUDA cores, 576 Tensor cores, and 72 RT cores across the TU102 silicon. The GeForce RTX 2080 Ti is carved out of the TU102 by disabling four SMs, resulting in 4,352 CUDA cores, 544 Tensor cores, and 68 RT cores. The GPU is endowed with 272 TMUs and 96 ROPs.


As we mentioned, the memory bus is narrowed down slightly to 352-bit, which holds 11 GB of GDDR6 memory, clocked at 14 Gbps, resulting in a memory bandwidth of 616 GB/s.

Features

Again, we highly recommend you read our article from the 14th of September for intricate technical details about the "Turing" architecture feature set, which we are going to briefly summarize here.


NVIDIA RTX is a brave new feature that has triggered a leap in GPU compute power, just like other killer real-time consumer graphics features, such as anti-aliasing, programmable shading, and tessellation. It provides a programming model for 3D scenes with ray-traced elements that improve realism. RTX introduces several turnkey effects that game developers can implement with specific sections of their 3D scenes, rather than ray-tracing everything on the screen (we're not quite there yet). A plethora of next-generation GameWorks effects could leverage RTX.


Perhaps more relevant architectural features to gamers come in the form of improvements to the GPU's shaders. In addition to concurrent INT and FP32 operations in the SM, "Turing" introduces Mesh Shading, Variable Rate Shading, Content-Adaptive Shading, Motion-Adaptive Shading, Texture-Space Shading, and Foveated Rendering.


Deep Learning Anti-Aliasing (DLSS) is an ingenious new post-processing AA method that leverages deep-neural networks built ad-hoc with the purpose of guessing how an image could look upscaled. DNNs are built on-chip, accelerated by Tensor cores. Ground-truth data on how objects in most common games should ideally look upscaled are fed via driver updates, or GeForce Experience. The DNN then uses this ground-truth data to reconstruct detail in 3D objects. 2x DLSS image quality is comparable to 64x "classic" super sampling.