Editing Rendering (computer graphics) (section)

=== GPUs ===
{{main|Graphics processing unit}}
The [[Graphics card|3D graphics accelerators]] of the 1990s evolved into modern GPUs. GPUs are general-purpose processors, like [[Central processing unit|CPUs]], but they are designed for tasks that can be broken into many small, similar, mostly independent sub-tasks (such as rendering individual pixels) and performed in [[Parallel computing|parallel]]. This means that a GPU can speed up any rendering algorithm that can be split into subtasks in this way, in contrast to 1990s 3D accelerators which were only designed to speed up specific rasterization algorithms and simple shading and lighting effects (although [[Kludge#Computer science|tricks]] could be used to perform more general computations).{{r|n=AkenineMöller2018|loc=ch3}}{{r|Peercy2000}}

Due to their origins, GPUs typically still provide specialized hardware acceleration for some steps of a traditional 3D rasterization [[Graphics pipeline|pipeline]], including hidden surface removal using a [[Z-buffering|z-buffer]], and [[texture mapping]] with [[mipmap]]s, but these features are no longer always used.{{r|n=AkenineMöller2018|loc=ch3}} Recent GPUs have features to accelerate finding the intersections of rays with a [[bounding volume hierarchy]], to help speed up all variants of [[Ray tracing (graphics)|ray tracing]] and [[path tracing]],{{r|n=RayTracingGems_Forword_Stich}} as well as [[Neural network (machine learning)|neural network]] acceleration features sometimes useful for rendering.{{r|NvidiaDLSS}}

GPUs are usually integrated with [[GDDR SDRAM|high-bandwidth memory systems]] to support the read and write [[memory bandwidth|bandwidth]] requirements of high-resolution, real-time rendering, particularly when multiple passes are required to render a frame, however memory [[memory latency|latency]] may be higher than on a CPU, which can be a problem if the [[Analysis of parallel algorithms#Critical path|critical path]] in an algorithm involves many memory accesses. GPU design accepts high latency as inevitable (in part because a large number of [[Thread (computing)|threads]] are sharing the [[Bus (computing)#Memory bus|memory bus]]) and attempts to "hide" it by efficiently switching between threads, so a different thread can be performing computations while the first thread is waiting for a read or write to complete.{{r|n=AkenineMöller2018|loc=ch3}}{{r|Lam2021}}{{r|Gong2019}}

Rendering algorithms will run efficiently on a GPU only if they can be implemented using small groups of threads that perform mostly the same operations. As an example of code that meets this requirement: when rendering a small square of pixels in a simple [[Ray tracing (graphics)|ray-traced]] image, all threads will likely be intersecting rays with the same object and performing the same lighting computations. For performance and architectural reasons, GPUs run groups of around 16-64 threads called ''warps'' or ''wavefronts'' in [[Single instruction, multiple threads|lock-step]] (all threads in the group are executing the same instructions at the same time). If not all threads in the group need to run particular blocks of code (due to conditions) then some threads will be idle, or the results of their computations will be discarded, causing degraded performance.{{r|n=AkenineMöller2018|loc=ch3}}{{r|Gong2019}}