Intrinsic on 21 October 2016
I didnt see this on here, so I'm breakimg it down to the bestb of my understanding from MC's recent tech synopsis. these are the key features and improvements that allows the PS4pro do what it does.
- Custom Hardware that allows an ID Buffer in the render pipeline. This allows for better AA and is key in sonys upscale tech (there are two methods suggested but more can be used). The ID Buffer basically marks or identifies the properties and position of every rendered pixel on a hardware level which makes it possible to track them with no performance cost.
- 1GB of DDR3 ram added for smaller apps. This allows apps like Netflix, PSVue, Amazon prime. YouTube...etc to be moved out of the GDDR5 pool and into slower ram when not in use. This in turn frees up 1GB of GDDR5 ram on the PS4pro half of which is used to accommodate the 4k render targets and the other half for a 4k UI front end.
- Known factors are increase in memory bandwidth, cpu clock and GPU core count and increased GPU clock.
- Inproved FP16 performance. A GCN first in the PS4pro APU and due in the Vega architecture. This allows FP16 operations to be run in actual 16bit environments as opposed to 32bit addresses. This basically doubles the performance of FP16 since you can fit twice as many instructions in the same address space as before.
- More efficient GPU design all round, can run more wavefronts, new primitive discard culling implementation (removal of triangles that are too small too see which allows more room to render the stuff you can actually see)
- Advanced multi resolution support for increased performance in VR.
And on 4k. The two primary ways of accomplishing this on PS4pro is using simpler geometry only rendering (minimal cost and is a post process upscale of a base 1080p image albeit with a 4k render target. Downside is that even tho you get a 4k final output image texture detail is still identical to 1080p image) and the second is checkerboard rendering (1920x2160 base which is basically half the pixel count of native 4k and using the id buffer data missing pixels are recalculated per frame to arrive at a final 4k frame buffer)
It's possible for both systems to be mixed and matched to improve results or for completely different systems to be used. It's also possible to natively render up to 1800p (think 1080p's 900p for 4k) and checkerboard from there.
The checkerboard is a more costly process compared to geometry rendering (it can take up to 3 weeks to implement as opposed to the other being implemented in as little as a day) but yeilds significantly better results. May require you to stay as little as 6-12inches from the screen to be able to spot the difference between it and native 4k.
This is a brief of all the stuff discussed so far. For more in-depth info read this Digital foundry article.
