Dear lord and this is the problem. NVlink is a interconnect, it isn't intended to make the CPU and GPU work as one, it's meant to improve scaling of multi gpu systems. Something you missed with all the articles from Nvidia comparing the interconnect to PCI-E 3, with all the articles with multi-gpu in the title, with every benchmark being about multi gpu.
The main premise of HSA is not to allow the cpu and gpu to work as one, it is ONE of the major premises. Having independent IP blocks available from different companies that will work together is frankly as large a premise and one of the most difficult to achieve.
One fundamental goal behind a cpu and gpu working as one is to have them on the same die to remove the communication latency. Currently ignoring unified memory, sending a smaller amount of data to a gpu across the pci-e bus to be processed before the cpu can continue has extreme latency and very low performance improvement if not a performance reduction.
99% of Nvidia's work is around having a single cpu in a system with up to 8 gpus and 99% of the work happening on the GPUs. When you do this bandwidth over the interconnect becomes saturated as you have 8 or even more GPUs using the same bus. This is entirely what NVLink is out to do, alleviate that bandwidth. It has zero goal, zero intention of enabling on die gpu/cpu to act as one unit quickly and efficiently.
You mentioned white papers, I did not. Funny I talk about Java, why, you think mentioning Java as well somehow links Nvlink and Java together to improve your argument, it doesn't. Java is a language, what you do with it doesn't have to be connected to what someone else does with it at all. You're inventing connections to support your argument where there are none.
You say it's a standard for markets where the incumbents have no interest in taking part... then say it will gain some traction in mobile. So mobile is a market, literally all the major players are involved.... but the incumbents have no interest, okay.
Thankfully there is zero overlap in desktop and mobile software. I mean people manipulate photos on mobile phones all the time but you'd never want to play with photo software or gpu acceleration on the desktop.
I'm not saying anything bad anywhere about Nvidia. YOU are making it AMD vs Nvidia.
YOU decided to poo poo HSA as unnecessary and pointless and attempted to imply that Nvidia have their own work to accomplish the same things. I am arguing against this utter nonsense you are coming out with. It's nothing to do with Nvidia. there is entirely nothing wrong with NVlink, it's not bad technology, it's not hugely interesting not least because it's only aimed at IBM compute machines which have and never will have any relevance to 99.99999% of users on this forum.
This is entirely about you attempting to deflect another AMD technology with absurd arguments on things you know very little about.
EDIT:- I would add that as a programmer who has studied computer science and still continues to learn more about programming, threading is pretty easily achieved, the incentive to do it is not there. It is expensive to have 30 extra people to continually code, bug check and update programs for every platform separately to achieve the required threading and gpu acceleration. The financial incentive, which is what man hours of coding are, a cost, is simply not that to push forward with wide spread gpu acceleration. When you remove all these barriers, allow less specialised coders easier ability to access gpu acceleration and make that code platform independent you dramatically reduce the cost of coding to take advantage of the available performance.
If you can gain 5% performance but it requires 1000 man hours of work on each of 5 platforms you might not bother, when the performance gain becomes 45%(because of things like unified memory) and it takes 200 hours of work total for all platforms because the code has been made higher level, easier and works on all HSA hardware you are changing the cost/benefit ratio so significantly that people will start to code for it. That is the main reason behind HSA, to make gpu acceleration/general compute of any kind usage cheaper to enable while also providing a larger performance benefit that gives the industry no reason not to move forwards.
Lastly, me and Pottsey agree on precious little, that we both agree should suggest just how painfully wrong you are on this.
