Thorsten wrote in Thu Aug 17, 2017 6:00 am:Very serious, absolutely serious, we do not know the true architecture of a GPU as it is covered by industrial secret.
Makes you wonder how on Earth they write those OpenSource drivers, knowing nothing about how the GPU internally works...
Unfortunately, the complexity of hardware products such as the CPU-GPU is so high that real architecture is not often known. To develop a drive, and not sign an NDA, an official architecture is made known, which often covers the one actually present in the device. This is also logical if I develop some architecture to make a system more performing, why do I have to tell my business competitors?
For example, if the GPU has a mixed architecture with a CPU (the Intel case), do you think there are no mechanisms similar to the predictive pipes that handle the code? The use of a compression method in transferring data to RAM through algorithms that are conceptually similar to MPEG can be a good strategy to increase the RAM access speed without too much reduction of the usable bandwidth from the CPU.
For example, watch this comparison:
http://www.game-debate.com/gpu/index.php?gid=3221&gid2=885&compare=hd-i7-6700k-vs-geforce-gt-740It does not seem surprising that 20-24 core can compete with 384 core and with a non-dedicated system RAM!
Certainly, the Intel GPU on the card goes slower: Texture Rate 5.6 GTexel / s (Intel) versus 31.8 GTexel / s (GT740) ... but the games at low resolution (800x600) goes at the same speed.
This seal means that the system is efficient!
With the my current Intel i7 compared to the previous GTX870M with 6 GB of RAM with FGFS I go slower, at the same resolution of about 5 times but GTX870M has 1344 core instead of 24 of Intel (ratio 1/56), theoretically it should go slower than 50 times not 5!
This was my surprise when NVIDIA broke and I had to use the i7 GPU. I thought I could no longer use FGFS, but instead, even at low resolution (1/4), I still get 15-20 fps ... as before!
Of course, to get similar performance, you need to work well on internal architecture, with methods and tricks that the manufacturer hides as they are the technological advantage over the competition.
Thorsten wrote in Thu Aug 17, 2017 6:00 am:Some of us have moved beyond that observations a whole ago, noted that it actually is Agriculture that needs to be set to full, used that to identify the precise bit of code that's running for these settings and even tentatively identified a line.
Too bad I have seen, tried and verified that this method does not use "agriculture-ALS.frag" and therefore only serves to skip the problem...
I'm interested in solving problems, understanding and not simply skip it!