--- (---.delete@this.redheron.com) on January 26, 2023 9:37 am wrote:
> The NYT article on SPR
> https://www.nytimes.com/2023/01/10/technology/intel-sapphire-rapids-microprocessor.html
> is as journalistic and NYTish as you'd expect, but does include a few actually valuable nuggets.
> For example we are told that SPR discussions began in 2015. This is interesting in so far
> as I've suggested before that Intel's primary problem (whether caused by x86 or other concerns)
> is slow turnaround, taking ~7yrs to complete what Apple and ARM are doing in ~4yrs.
>
> Likewise they seem to not have adequate control of complexity, given that the SPR delays
> are blamed more on design/validation flaws rather than process delay. Of course covid
> hurt (like it hurt everyone) so I'm willing to give some slack on that.
>
> One could also throw marketing in the mix insofar as (see for example Patrick's STH analyses) most customers
> will probably not be using SPR's accelerators, but they were part of that complexity and thus the delay.

I feel like a lot of the complexity had to do with shoving a ton of bandwidth across chiplet links. L3 cache accesses can go across tiles, so that you get unified >100 MB cache. Another source of complexity is adding optional HBM support. Contrast that with other chiplet designs where you only have memory bandwidth going across dies, and large caches before the chiplet links.

I don't think accelerators were main source of complexity. Intel already put some accelerators (QAT) in their Xeon D line. And they have a history of integrating random accelerators (iGPU, GNA, IPU) into their client chips. That pretty much involves tacking another block onto the on-chip interconnect and isn't a big deal.

> A wiser company might, perhaps, have developed the accelerator IP on a different track, ready to add to
> the system when it was validated, but not delaying the bread-and-butter product for untested tech?
>
> Overall I find little in the article to make me positive as to Intel's execution capabilities.
> There's a whole lot of "mea culpa, mea maxima culpa" but every flaw that they admit to was a
> flaw that was clear five years ago, so it's not obvious why THIS time the mea-culpa'ing is any
> more serious. ( Claims about streamlining the product road map! Anyone believe that??? )

Unfortunately the article is paywalled :/