Perhaps the more tech savvy VG-ers will make more sense of this. It's like trying to read hieroglyphs for me:

View on YouTube

Tough subject matter to explain.
Watched it and felt like I could follow it.

I liked that he went into where the biggest gains from useing half floats would be.

More or less still no... It can be used to accelerate certain effects such as DoF, Motion Blur, HDR and etc but for most tasks that require and are higher percision, f32 is still required.

Really just sounds like a way to increase efficiency but only on certain things. So no, its not a 8.4tf beast, specially when you consider that current gen pc gpus and the upcoming scorpio can all do it.

Its a very well known technique, the scorpio will mostl likely also have it, instead of doubeling the Tflops (which is rediculous) they should have just said the graphical processor has benefits in some calculations and graphical settings/effects, which is true.

Obviously newer chips will handle things better and Tflops doesn't say everything, the newer the AMD technology, the better the performance overall.

Anyway, im waiting for the PS5 but its a very exciting month for people first getting their PS4 console and choosing the Pro model to get that little bit of extra performance. :)

If the Switch is 1.5 TFLOPS then the Pro is certainly 8.4 TFLOPS.

Eh... it's based on this belief in FP16 operations rather than FP32 for determining flops.

Quite honestly, very few people can actually explain it clearly or correctly. I doubt there's even real-world example of FP16 implementations out there just yet. For most people - it's really a bunch of magical thinking they rely upon to tell themselves PS4 is more powerful than Scorpio or that Switch is for sure as powerful as PS4.

at the end of the day... who cares how many flops...

All I know is, if this beast can make my favorite games look even better, it's got my 400 dollars.

In the video he said that processing things in FP16 isn't new. It is just being pushed on the GPU side. When I was talking about the Switch's 1.5 number I was drawing on the information from when CPUs did this. It just somethings don't need to use the full 32bit register so it only makes sense to spilt the register into two FP16 registers, so the things that would have usually left most the register empty or all zero's can now use that empty area for other calculations, so it speeds up certain processes.

And how much of a benefit developers and the games they make get out of this depends on the tools they are given from Nintendo/nVidia and Sony/AMD to utilize this.

its like this, if you can do a game all programed in 16bit FP the ps4 pro is capable of doing 8.4 trillion calculations per second, but if you have to use all 32 bit FP, it will do 4.2 trillion calculations. because there isn't just 16 bit games that value will never be achieved, but unlike before, where there was 32 bit FP registers, if they were all used or not was irrelevant, now that 32 bit space can be divided by 2 and do 2 different calculations at the same time, this means, that if it used, the GPU power is somewhere between the bottom 4.2 trillion calculations per second and the illogical 8.4 trillion calculations per second, where it lands, depend how much does the game use the 16 bit spaces for some stuff.

You won't use all calculations in FP16 like sergiodaly said. And although previous consoles could do it, the space and load on FP16 was equal to FP32, so what they have now is that in some situations they will be able to remove some of the calculations from FP32 and converte to FP16 so the 4.2TFlops will be used even more efficiantly than 1,8 TFlops of PS4 so the gains in GPU processing power is even more than a direct math calculation would show. But don't get the hopes too high before time.