Kevin G (kevin.delete@this.cubitdesigns.com) on August 1, 2018 2:03 pm wrote:
>
> I had been tried on the PC a little bit with some older hardware (PDF):
>
> https://pdfs.semanticscholar.org/167c/651a235cf567ee8ca19b8d0e4d2f19e01b42.pdf


Thanks, stuck it onto the cloud drive to have a read of, next time I can.

The five million dollars amount of capital needed to invest in building an appropriate compiler, is far too small for Intel as a corporation to get distracted with. It would be a better capital size of project, for a start-up outside of Intel altogether, or even Angel funding for that matter. Intel might be able to assist that team in making the hardware sure. But it's a classic case of not building the right compiler technology first, before starting to bake out new silicon production.

Mark Ruolo did give us an insight into the engineering problem though. The focus needed on a restricted sub-set of C programming language that Mark Ruolo described overhead.

https://www.realworldtech.com/forum/?threadid=138897&curpostid=178881

And it asks the question too, given Mark's initial estimate of a compiler technology of five million dollars in development,... would it not make sense to invest a little more time and engineering and develop a compiler technology,... that would work WITH the dual socket system, as suggested originally by Adrian above?

https://www.realworldtech.com/forum/?threadid=138897&curpostid=178854

Where the compiler is designed to become intelligent enough from the beginning, to produce binary which will take full advantage of the kind of system that Adrian mentioned? And the kind of system that may become more common moving into the future - the strengths of having big cores combined with smaller cores - in one system. I.e. Where one, working with the sub-set of C language that Mark explained, one is able to produce execute-ables that don't suffer as much from a) vectorization performance not being good enough or b) memory access performance not being good enough either.

I.e. Smaller cores to deal with vectorization, larger ones for memory access constrained workload?

Standing way back from it all though.

Many technology analysts talk about the need for a 'third' force (i.e. Intel corporation), working in graphics acceleration hardware, and how good that would be for the market in GPU technology. However, I would argue that it's going to be a lot easier one of these days, for nVidia to start making CPU's to work with it's CUDA technology, and 'the whole system' working from that end. Instead of Intel who already has the CPU, trying to come up with a viable GPU, from it's end.

What do you think?

Maybe the kinds of systems that nVidia eventually comes up with, if it was something like I described would not be a general purpose computer system. Maybe something intended more for purposes of AI and machine learning inside the database cluster environment. But there's a potential threat there, to the dominance of Xeon in that environment, which I don't think that Intel sees coming either.

A lot of these 'also ran' companies like nVidia have opened my eyes at least in recent times, owing to their persistent efforts to explore new technology and new markets. Something that Intel hasn't been good at doing now, for a while. Yeah, Optane looks to be a game changer (i.e. from a point of view of an Intel product line, that can build 'all of the system' out of Intel parts that fit together neatly). However, I still have my doubts as to where this 'complete system' will come from - whether it will in fact be Intel or somewhere else altogether.

> > In fact, if you look at small systems being manufactured and shipped by Asrock nowadays, with nVidia
> > MXM graphics cards, is it a million miles away from what you describe? Albeit, it's a graphics card
> > and a cpu still, but it's starting to look an awful lot like a '2-socket machine' everyday.
>
>