These guy(s), do the same mistake (Octane devs and users do it too) in defending their speed in comparison to Vray/Corona, by using the argument, that it would be comparable when Vray/Corona, disabled their secondary GI kernels.
Well but...too bad, that's exactly what makes them fast. These GPU-raytracers are almost never faster in actual GI-heavy (like interior) scenes from what I've seen (applies to all of them, except for Redshift).
Redshift devs got it. You can use GPUs as fast as you can, unbiased pathtracing alone is not fast. Unless we talk car renders, HDRi-lit close-ups of decoration or some other simplicity.
I half agree, Redshift I have tried to get clean results for internal renders using the GPU GI caching interpolated GI and without high settings I can't get rid of the splotches, its like going back to Mental ray. So I tried there Path tracer and its slow as any GPU path tracer.
However Fstorm, its sampling methods are pretty damn quick :) and its by far the fastest GPU render I have used even for internals.
Now if Ondra put what he learnt with his Herpy derpy GI cache and made that for GPU, if at all possible, that would be very interesting, Its the cleanest GI cache thing I have ever used.
Things I like about Fstorm:
Glare like you mentioned is beautifully done, the quality and believability is great.
I like the Sun Sky system.
Its material is pretty solid, has lots of maps/shaders that are normally missing from a GPU renderer at the start development.
The other thing I like about GPU renderers, is the GPU seems to have a faster development.
Being able to upgrade a GPU or just adding another one into my system for more speed is awesome.
However any real work I do, I do in Corona, but enjoy testing any renderer I can find :) Corona is so reliable and easy to use.