IBM Going with Samsung after dumping GlobalFoundries

By: anon (spam.delete.delete.delete@this.this.this.spam.com), December 21, 2018 6:30 am
Room: Moderated Discussions
Michael S (already5chosen.delete@this.yahoo.com) on December 21, 2018 3:46 am wrote:
> Maynard Handley (name99.delete@this.name99.org) on December 20, 2018 6:43 pm wrote:
> > itsmydamnation (no.delete@this.way.com) on December 20, 2018 2:02 pm wrote:
> > > Anon (Anon.delete@this.anon.com) on December 20, 2018 11:07 am wrote:
> > > > https://newsroom.ibm.com/2018-12-20-IBM-Expands-Strategic-Partnership-with-Samsung-to-Include-7nm-Chip-Manufacturing
> > >
> > > This was the thing i actually wondered about most when GF made the 7nm abandonment announcement.
> > > How much process development capability does IBM have left? Will IBM work with Samsung
> > > to develop things like edram for their 7nm process, can it already do it?
> > >
> > > Its going to be interesting to exactly what the 7nm power chips look like.
> >
> > There are many ways to skin a cat. IBM may perhaps have to give up on-chip eDRAM, but as a replacement they
> > may be able to use some sort of die-die stacking to put
> > a massive L3 cache as a separate die under the logic
> > die. (To goose the performance a little, you could also do mid-90s things like have on-logic-die tags for
> > the L3, with the L3 data on the separate die, but I don't know if that's worth the hassle.)
> >
>
> There are many ways to skin a power cat. Not so many way for skinning cat360.
>
> Traditionally S/360 designer don't trust SRAM for long-term storage.
> For years their SRAM caches were write-through (their preffered term is store-through)
> and most recently (z14) they got rid of SRAM caches completely.

I'm pretty sure the L1 is still SRAM.
I don't think it's a matter of not trusting SRAM, but rather not trusting caches in general.
Their whole RAS concept is based on the assumption that anything could fail and must have a backup.

> With L3 on separate die, write-through L2 will be very bad both for performance and for power.
>
> > The point of this exercise is that it would take advantage of directions that Samsung is probably
> > already good at or wants to be good at (die on die stacking) rather than requiring a very expensive
> > process stream that no-one outside IBM cares about or is willing to pay for.
> >
> > And once you've accepted this sort of model, multiple implementation options present themselves. The eDRAM
> > die could still be manufactured at GF. Or (if it makes sense) an MRAM die could be used in its place.
> >
> > One of the current roadblocks (perhaps THE big roadblock right now) to more aggressive die stacking
> > is not actually stacking the dies, or alignment, or thermal issues, or anything like that; it's that
> > design tools are still very far from being set up handle multi-layer/3D designs. But logic stacked
> > on a cache is about the simplest sort of SoC system one can envisage, and that is well-partitioned
> > enough that it seems to be a feasible first step, even with the current state of design tools.
>
>

< Previous Post in ThreadNext Post in Thread >
TopicPosted ByDate
IBM Going with Samsung after dumping GlobalFoundriesAnon2018/12/20 12:07 PM
  IBM Going with Samsung after dumping GlobalFoundriesitsmydamnation2018/12/20 03:02 PM
    IBM Going with Samsung after dumping GlobalFoundriesanon2018/12/20 04:50 PM
    IBM Going with Samsung after dumping GlobalFoundriesMaynard Handley2018/12/20 07:43 PM
      IBM Going with Samsung after dumping GlobalFoundriesMichael S2018/12/21 04:46 AM
        IBM Going with Samsung after dumping GlobalFoundriesanon2018/12/21 06:30 AM
          IBM Going with Samsung after dumping GlobalFoundriesMichael S2018/12/21 07:39 AM
            IBM Going with Samsung after dumping GlobalFoundriesanon2018/12/21 08:24 AM
            Definitely a mistakeDavid Kanter2018/12/21 09:09 AM
              Definitely a mistakeMichael S2018/12/22 08:45 AM
                Definitely a mistakedmcq2018/12/22 11:26 AM
                Definitely a mistakeRobert Williams2018/12/22 11:40 AM
                Definitely a mistakesomeone2018/12/22 01:46 PM
                L1 caches are SRAMDavid Kanter2018/12/22 02:27 PM
                Definitely a mistakeanon2018/12/22 02:30 PM
                Definitely a mistakecarop2018/12/22 08:09 PM
                L1 caches are SRAMjuanrga2018/12/23 05:23 AM
    IBM Going with Samsung after dumping GlobalFoundriesDavid Kanter2018/12/20 09:28 PM
      IBM Going with Samsung after dumping GlobalFoundriesKevin G2018/12/21 08:57 AM
        Samsung fabbed Alpha 21164A chips in 1996/7Mark Roulo2018/12/21 05:24 PM
          Samsung fabbed Alpha 21164A chips in 1996/7Kevin G2018/12/25 03:03 PM
      IBM Going with Samsung after dumping GlobalFoundriesAdrian2018/12/21 11:55 PM
Reply to this Topic
Name:
Email:
Topic:
Body: No Text
How do you spell purple?