David Kanter (dkanter.delete@this.realworldtech.com) on August 8, 2014 11:51 pm wrote:
> juanrga (nospam.delete@this.juanrga.com) on August 8, 2014 6:40 pm wrote:
> > Ungo (a.delete@this.b.c.d.e) on August 8, 2014 4:48 pm wrote:
> > > juanrga (nospam.delete@this.juanrga.com) on August 8, 2014 11:00 am wrote:
> > > > But some of those companies were showing benchmarks behind the doors. And Nvidia has already admitted
> > > > in public that those ARM64 CPUs can feed its faster GPGPUs the same than with traditional Xeons.
> > >
> > > "Admitted"? You're spinning this as if it's something NVidia only grudgingly accepts,
> > > and using that spin to imply credibility. But it's actually in NVidia's interests to
> > > promote this idea. They've spent a lot of money developing ARM SoCs and CPUs.
> >
> > They are promoting ARM CPUs from other companies, not their Denver project.
>
> Uh who? Half the companies doing ARM server chips have cancelled
> their projects (or delayed them): Calxeda, Qualcomm, Samsung.
>
> Nvidia's plan so far seems to be working with IBM.
>
> > I believe Nvidia when they claim that those ARM CPUs will provide performance similar to Xeons.
> > I believe because I know some details of the arch. You don't believe Nvidia. Case closed.
>
> Which architecture? I haven't heard of anything from any ARM-based vendor
> that would credibly challenge a Xeon in terms of performance.
>
> > > Your behavior in recent posts is interesting and revealing. When you find that (IIRC) one of three authors
> > > of a paper you don't like happens to work in the same lab
> > > as people funded by Intel, you shout about tenuous
> > > Intel connections being proof that the paper is ridiculous while providing little substantive criticism
> > > of the paper's contents. But when it's an idea you want to be true, suddenly this hyperskepticism about
> > > motives flies out the window and obvious corporate marketing is treated as gospel.
> >
> > I found the paper in an ordinary search of literature and read it, finding lots of weird stuff
> > and unproven statements. They didn't look arbitrary to me and then, only then, I decided to
> > search further info about the authors and found that the senior author is closely related
> > to Intel, which didn't surprise me because the paper did look as a marketing piece.
>
> What paper are you talking about?
>
> > I mentioned some issues with this crap paper in a previous post, but as another poster
> > mentioned that the paper was analyzed here before. No need to repeat what was said.
>
> For a 3-issue core, there is no way that using the ARM ISA will give more than a 5-10% advantage over
> Intel (all things being equal). However, things are not equal. Intel has larger design teams, Intel
> has better process technology, Intel has higher volumes and can bin more aggressively, etc.
>
> David

gcc code for ARMv8 for Linux seems to be on average almost exactly the same length as x86 code. The average length of x86 instructions tends to be just a bit over 3 bytes therefore x86 needs about 30% more instructions to do the same job. With 3 wide decode for both that is a disadvantage for x86. There's a lot of other things that confuse matters like address literals would be in a literals pool for ARMv8 whereas they can be inline in x86 code and the x86 decoders also go to the trouble of trying to put two instructions together sometimes and really performance is measured for code that is predecoded in the instruction cache for x86 but overall there is a bit of a hill for x86 to overcome.