AI Term of the Moment
AI/ML
Redirected from: Compute Unit
Definition: GPU
(Graphics Processing Unit) A programmable processor that renders 3D images and executes AI applications. NVIDIA is the inventor of the modern GPU and leader in the field (see NVIDIA).
The Engine Behind AI and Crypto
A GPU is a "math machine" that performs huge amounts of simple calculations simultaneously making it ideal for AI applications that require trillions of computations (see neural network and AI training passes). GPUs are also used to mine certain cryptocurrencies (see GPU mining).
The Streaming Multiprocessor (SM)
An NVIDA GPU manages more than a hundred streaming multiprocessors (SMs), the primary processing units. For example, the GPU in NVIDIA's RTX 4090 graphics card controls 128 SMs each with 128 CUDA cores for a total of 16,384 CUDA arithmetic units. For non-NVIDIA GPUs, Compute Units (CU), Workgroup Processors (WGP) and Xe Cores are equivalent to the streaming multiprocessor. See CUDA.
Ray Tracing Engine
GPUs for rendering 3D also include circuits for ray tracing, which simulates a light source falling on objects to create bright areas and shadows. Fast ray tracing enhances realism and has become mandatory for serious gaming (see ray tracing).
GPUs Are Often Integrated With the CPU
GPU and Everything Else
Graphics Evolution in a Windows PC
Blame It on the Gamers!
The Engine Behind AI and Crypto
A GPU is a "math machine" that performs huge amounts of simple calculations simultaneously making it ideal for AI applications that require trillions of computations (see neural network and AI training passes). GPUs are also used to mine certain cryptocurrencies (see GPU mining).
The Streaming Multiprocessor (SM)
An NVIDA GPU manages more than a hundred streaming multiprocessors (SMs), the primary processing units. For example, the GPU in NVIDIA's RTX 4090 graphics card controls 128 SMs each with 128 CUDA cores for a total of 16,384 CUDA arithmetic units. For non-NVIDIA GPUs, Compute Units (CU), Workgroup Processors (WGP) and Xe Cores are equivalent to the streaming multiprocessor. See CUDA.
Ray Tracing Engine
GPUs for rendering 3D also include circuits for ray tracing, which simulates a light source falling on objects to create bright areas and shadows. Fast ray tracing enhances realism and has become mandatory for serious gaming (see ray tracing).
Introduced in 2012, AMD's Trinity chip integrated a GPU with memory (DDR3) and two dual-core x86 CPUs, each with their own L2 cache (see x86, L2 cache). (Image courtesy of Advanced Micro Devices, Inc.)
In 2020, Qualcomm introduced the Snapdragon 888 system-on-chip (SoC). Along with a GPU, this single chip does just about everything in an Android smartphone. See Snapdragon and SoC.
On the left is the evolution of graphics circuits with 7 being the latest. On the right is performance with 1 being the fastest. Gamers plug in the fastest GPU card they can afford, and high-end graphics can cost more than the entire PC (see RTX).
Had gaming with 3D images not been such a force in the 1990s, NVIDIA would not have been compelled to make video cards with GPUs that would get faster every year. The enthusiasm for 3D games was singularly responsible for AI to flourish by the 2020s. See NVIDIA.