(*This post may contain affiliate links, which means I may receive a small commission if you choose to purchase through links I provide (at no extra cost to you). Thank you for supporting the work I put into this site!)
GPU is the heart of the graphics card and its working is defined by its architecture, which is called GPU architecture. Each graphics card family has its own GPU architecture and every new family or series of graphics cards comes with more advanced and powerful GPU architecture.
Every GPU manufacturer designs its own GPU architecture and GPU architectures of graphics cards from Nvidia and AMD are totally different in working, operation and naming. Examples of Nvidia GPU architectures are Fermi, Kepler, Pascal, Volta, Turing whereas from AMD we have GCN (1.0, 2.0, 3.0), Polaris (GCN 4.0) and Vega. Also, two GPU architectures from the same GPU manufacturer are different in working, efficiency and performance e.g. Fermi is different from Kepler, and Pascal is different from Kelper and so on. But here I am going to make a comparison of Nvidia’s three modern GPU architectures, which are Pascal, Volta, and Turing based on their specifications, performance, usage & applications in the graphics industry.
Pascal GPU Architecture
Pascal is one of the most popular GPU architectures from Nvidia and the successor of Maxwell GPU architecture. The GeForce 10 series and Quadro P series graphics cards are built on the Pascal GPU architecture. Pascal GPUs are built on the 16nm / 14nm fabrication process and uses CUDA Cores as their main pixel processing units and for Rasterization. GeForce GTX 1050, GTX 1050 Ti and GT 1030 are built on 14nm FinFET technology while others are built on the 16nmm technology process. Pascal GPU architecture brings significant improvements over the older architectures in terms of performance, power consumption (TDP) and heat generation.
Pascal GPU architecture supports GDDR5, GDDR5X and HBM2 memory. Only higher-end workstation graphics card, Nvidia Quadro GP100 comes with high bandwidth HBM2 memory. Nvidia Pascal GPU architecture supports DirectX 12, OpenGL 4.6, Vulkan, OpenCL, SLI (for high-end GeForce cards), NVLink (for top workstation GPUs only), NVENC, G-Sync, GPU Boost 3.0, DisplayPort 1.4, HDMI 2.0b, CUDA Compute Capability 6.0 / 6.1 and is VR Ready. Pascal based GPUs can also be found in Notebooks or Laptops.
Must Read: GDDR5 vs GDDR5X vs HBM2 vs GDDR6 Comparison
Volta GPU Architecture
Volta is the successor of Pascal GPU architecture and is built on the 12nm fabrication process. It supports high-speed and high-bandwidth HBM2 memory. Volta architecture is only designed to cater to workstation and datacenter needs. At present, there are no gaming graphics cards that use Volta architecture. This GPU architecture is powered by CUDA Cores and Tensor Cores. The Tensor Cores are used for Artificial Intelligence (AI), Deep Learning, Machine learning, Complex arithmetic calculations and provides an over 5X increase in performance compared to the Nvidia Pascal architecture.
The graphics cards that use Volta GPU architecture include Nvidia Titan V, Nvidia Titan V CEO Edition and Nvidia Quadro GV100. Volta architecture supports NVLink 2.0 technology which is much faster than the previous NVLink and allows much higher data transfer rates of 25 Gbit/s per data lane per direction. We may be able to see more graphics cards using Volta architecture but I think it would be only in the workstation graphics cards category.
Check out: Volta vs Pascal GPU Architecture Comparison
Turing GPU Architecture
Turing is the successor of Volta GPU architecture. It is one of the most advanced GPU architecture ever made. Turing GPUs are built on the 12nm FinFET manufacturing process and support GDDR6 memory which operates at very high speed can achieve much higher bandwidth compared to the previous GDDR5X and GDDR5 memories. This advanced GPU architecture comes with CUDA Cores, Tensor Cores, and RT Cores. Turing is the first GPU architecture to support Real Time Ray Tracing for creating lifelike lighting, shadows, reflections, refractions, and other advanced lighting effects. This Real Time Ray Tracing is handled by RT Cores and its performance is evaluated in a new metric called Giga Rays per second.
Tensor Cores in Turing GPUs are designed especially for Artificial Intelligence (AI) and Deep Learning for performing various types of complex calculations. It is also used for a technique called Deep Learning Super-Sampling (DLSS) or (Deep learning anti-aliasing) for smoothening out edges in games and for denoising, resolution scaling and video re-timing.
CUDA Cores are used for standard Pixel-based processing tasks or Rasterization, but in Turing, they feature a new streaming multiprocessor (SM) architecture that supports up to 16 trillion floating point operations in parallel with 16 trillion integer operations per second. This can now allow Developers to create complex simulations, such as particles or fluid dynamics for scientific visualization, virtual environments, and special effects. According to Nvidia, Turing GPUs provide up to 6X performance over the Pascal-based GPUs.
There are both workstation and gaming graphics cards based on the Turing GPU architecture. Turing Workstation Graphics Cards include Quadro RTX 8000, Quadro RTX 6000, Quadro RTX 5000, and Gaming Graphics Cards consist of GeForce RTX 20 series that include GeForce RTX 2080 Ti, RTX 2080, RTX 2070 and also the mighty Nvidia TITAN RTX. Turing supports multi-GPU setup for both Quadro RTX and GeForce RTX graphics cards using NVLink for doubling or tripling their computational performance. GeForce RTX graphics cards also support VirtualLink via USB Type-C connector for connecting VR Headsets on USB Type-C port for an amazing VR experience. On the other hand, Quadro RTX GPUs are capable of processing 8K videos in real-time which shows the power of this new architecture.
Pascal vs Volta vs Turing Comparison
A quick and brief comparison of Pascal, Volta and Turing GPU architectures from Nvidia.
|GPU Architecture ->||Pascal||Volta||Turing|
|Fabrication Process||14nm / 16nm||12nm||12nm|
|Memory support||DDR4, GDDR5, GDDR5X, HBM2||HBM2||GDDR6|
|VirtualLink (USB Type-C)||NA||NA||Yes|
|Multi-GPU support||Yes (in high-end cards), SLI and NVLink||NVLink 2||NVLink 2 / NVLink SLI|
|Graphics Cards||GeForce 10 series, Nvidia Titan X, Nvidia Titan Xp, Quadro P series workstation graphics cards, Quadro GP100||Nvidia Titan V, Quadro GV100||Quadro RTX 8000, Quadro RTX 6000, Quadro RTX 5000 / RTX series graphics cards|
|Applications||Gaming, Workstation||Artificial Intelligence (AI), Workstation, Datacenter||Artificial Intelligence (AI), Workstation, Gaming|
Well, it is no denying the fact that Turing is the most advanced GPU architecture and is leading ahead of Pascal and Volta in terms of technology. Turing is more of a multi-purpose GPU architecture that can perform all three things at the same time i.e. Pixel Processing, Artificial Intelligence (AI) and Real-Time Ray Tracing. Pascal is a very long-lived and successful GPU architecture but now it has been showing its age because of a lack of AI and Ray Tracing advanced functionalities. Volta is really a question mark for us because there are only a couple of graphics cards from it up till now and according to me it has been completely replaced by the newer and better Turing GPU architecture. The future of graphics definitely lies with the Turing architecture as the previous generations architectures fade out slowly. If you have anything to say then please leave a comment below.