Originally Posted by budgetgamer120
A teraflop is a teraflop no matter what archictecture. A teraflop means 1 thing and it's universal.
If a program was made to use 6 teraflop of processing power the gpu or cpu with less would fail.
I know on OCN all that matter are games. Games don't care how much teraflops a gpu has.
By saying a GPU has xx teraflops you are basically discerning how much peak power it has not necessarily what will be used. It just defines how many concurrent actions can occur on any given frame. You can process more light rays, shadows, physics, textures and have a higher headroom for either more pixels or less of a requirement to optimize due to the extra headroom.
Games do care how many teraflops a GPU can produce, especially with lighting effects which is one of the harder things to reproduce from a GPU perspective as there are so many variable floating points in a given light stream.
Think about a light going through a cloud with 60% opacity. The GPU now has to calculate how much light the light ray loses while going through the cloud, the reflections in the cloud of the light, distance and its correlation to light strength etc.
Teraflops is 1 trillion operations therefore teraflop on its own has no direct meaning outside of defining the peak load of concurrent operations. This is calculated by
floating point operations per clock cycle x clock rate giving you the max output.
The more important number is actually the floating point operations per clock cycle as this will directly give developers the knowledge in how much throughput they can achieve per full cycle.
Most workloads will not reach 100% and when it does, it may not doing it in an optimized manner. There are many other factors at play like memory bandwidth and latency from the stack priority that prevent you from maintaining peak processing.
These factors are why games may run better/smoother on a console over a PC given the same price range. Console developers increase memory bandwidth and custom PCB w/better component layouts can improve latency by moving components closer together.