David Kanter (dkanter.delete@this.realworldtech.com) on October 28, 2020 8:40 pm wrote:
> Dummond D. Slow (mental.delete@this.protozoa.us) on October 28, 2020 8:27 pm wrote:
> > Mark Roulo (nothanks.delete@this.xxx.com) on October 28, 2020 8:15 pm wrote:
> > > Dummond D. Slow (mental.delete@this.protozoa.us) on October 28, 2020 5:32 pm wrote:
> > > > me (me.delete@this.me.com) on October 28, 2020 11:14 am wrote:
> > > > > > AMD tanked in the stock market today when under normal circumstances it would have rallied, having soundly
> > > > > > beaten guidance and predictions. I think this signals the uncertainty around the deal, the lack of clarity
> > > > > > on direction, potential and especially purpose, never mind the official management line.
> > > > > >
> > > > >
> > > > > I'm sure we will see FPGA chiplets
> > > > >
> > > >
> > > > You sure?
> > >
> > > Xilinx has been shipping FPGAs built around chiplets since at least 2011 with the Virtex-7.
> >
> > But the point is that AMD has not shipped CPUs with any non-CPU chiplets, ever. Despite using MCM
> > for Opterons since 2010. Intel is going to add HBM and whatnots to Xeons (Sapphire Rapids) it seems,
> > while AMD has nothing like that planned, even if they should be in a better position to do it.
> >
> > I'm just saying AMD very seldom realizes these "that would be great" and "absolutely makes sense" concepts.
> >
> > BTW, Intel had embedded Atom chip with on-package FPGA around 2009-2010 I think.
>
> You need something like EMIB or even higher bandwidth to attempt a decomposed GPU. The
> internal bandwidth is FAR too high for anything like what AMD is doing with Epyc.
>
> David

I didn't focus much on writing the comment, what I actually meant to say that AMD never integrated a GPU-based accelerator chiplet into its MCM server processor even if it might make sense, so I would not expect to integrate a FPGA-based accelerator into the processor's package (or on-die even) either.