No. CPU bound means that there is something that not allows on the way CPU-GPU for the GPU to be fully utilized.
DX11 was serial API that was CPU bound. It was how many Draw Calls were pushed to command scheduler of every GPU. Because there was not enough commands dispatched for GCN - therefore it could not be fully utilized. It is fairly easy way of explaining. And pretty rough
. DX12, because it is multithreaded, and multiengine API completely removes any bottleneck on GCN. Nvidia does not get that big improvement of performance, because it was not bottlenecked by API. At least - not in the way GCN was.
Why do I compare DX11/12 to PassMark? Because PassMark is like DX11. If it not be, 100% we would see any AMD GPU being faster than even old Kepler GTX 780 Ti.
Well said. Technical details are spot on. My understandings are roughly the same as yours - GCN was built to be a forward looking, parallel threading monster. It was seriously held back by DX11.
[doublepost=1460677004][/doublepost]
That's hundreds if not thousands of different systems with different CPU's, clock rates, RAM, etc. This is regardless if it's a AMD or nVidia GPU being tested. So are you saying all the AMD's testers scored lower than their nVidia counterparts because they must have been using slower processors when they did their tests?
If you read tech websites, you'll see that people budget their computers with similarly priced components that fit with each other. No one in the right mind would buy a AMD FX4300 CPU and pair it up with a GTX980Ti. It means that the cheaper GPUs in Passmark's database is heavily correlated with cheap CPUs (which are mostly AMDs at the moment), which for the GCN architecture, holds back its performance severely.
Also, I have not seen a reputable electronics enthusiast website that uses Passmark as routine testing. This non-Passmark list includes the giants - Anandtech, Tomshardware, HardOCP, Hexus, Techpowerup and more.
In addition, synthetic benchmarks are typically not useful as indicators of real world performance. One reason is that, no one buys their hardware to run synthetic benches all day, and a second one would be that, real world applications (games or otherwise) are all optimized for modern architectures. Who releases unoptimized, generic code applications ? It makes no sense. Even at the compiler-level, you can tick a few flags to output code that takes advantage of a few performance-boosting features.
[doublepost=1460677650][/doublepost]
Yes, boards are made by manufacturers. But the dies are made by Nvidia and AMD, which sells them to manufacturers. And thats where we have to look
. Of course boards can fail as always. But what is the percentage of fail of die of GPU, and what is the percentage of board fails? I suppose only IHV's know this.
Actually, this isn't hard to deduce if you want a qualitative answer. If you look at all the different brands of AMD gpus (ASUS, Gigabyte, MSI, Sapphire, XFX, PowerColor, etc) and get access to some reliability data (they are not always available, but do crop up from time to time in reliability studies), you'll see that the variability from manufacturer to manufacturer is in a certain range. Now, if you compare the size of that variability (statistically) to the variability of the reliability of GPUs as a function of the die maker - AMD/Nvidia, you'll see that the manufacturer of the board (OEM) probably makes a bigger impact to the reliability of the GPU rather than the die maker.
I'd wager that the OEM has far more to do with the reliability of the GPU than either AMD/Nvidia. I can't gather all the reliability studies, but here is one:
http://hexus.net/business/news/comp...eport-gigabyte-top-motherboards-msi-graphics/
Notice that there is >2x increase in failure rates between the bottom on that list than the top. I doubt AMD or Nvidia makes any modern GPUs which are 2x more likely to fail than the other.
