OpenGL atomic counters vs atomics in a SSBO

Question

I came across this article that states there are no differences in performance between atomic counter buffers and an atomic variable in an SSBO:

https://software.intel.com/en-us/articles/opengl-performance-tips-atomic-counter-buffers-versus-shader-storage-buffer-objects

Is this actually true across nvidia and AMD GPU's now? I think I remember something about Radeon 5870 generation GPU's having specific faster support for the atomic counter subset? So I think it may have been an AMD specific thing at one point for performance?

From knowledge of nvidia CUDA I suspect it's never made a difference for them?

Does anyone know after which generation of GPU's from AMD/NVidia atomic counters are not worth it?

Answer 1

Mantle, AMD's low-level API, actually has specific support for atomic counters (they're part of queues, not memory). So there's every reason to believe that at least one piece of hardware doesn't just use SSBOs for them.