Brett (ggtgp.delete@this.yahoo.com) on September 14, 2021 8:17 pm wrote:
> Mark Roulo (nothanks.delete@this.xxx.com) on September 14, 2021 3:51 pm wrote:
> > hobold (hobold.delete@this.vectorizer.org) on September 14, 2021 6:47 am wrote:
> > > Heikki Kultala (heikki.kultala.delete@this.gmail.com) on September 13, 2021 7:33 am wrote:
> > >
> > > > There are no little cores in Alder Lake. There are mediun cores. And these
> > > > medium cores gives twice the performance/area than the big cores.
> > > >
> > > > Now there is throughput performance worth 12 big cores, but area of only 10 big cores.
> > > >
> > > > The advantage would be even better if there was 16 instead of 8 of those medium cores.
> > >
> > > Why stop there? By that line of reasoning, one would want many different core sizes
> > > from big to little. Say, core N has performance proportional to 0.9^N, silicon
> > > area proportional to 0.8^N, and power consumption proportional to 0.75^N.
> > >
> > > This way, the fewer threads are running, the faster each
> > > thread gets to run. And the more runnable threads there
> > > are, the more efficient the cores are getting (both in terms of energy as well as silicon real estate).
> > >
> > > Punch line: total silicon area and power consumption of an infinite number of such
> > > cores is finite. :-) But performance is finite, too. And yes, the extreme version
> > > of this idea is a joke ... but what about three, four, five ... core sizes?
> > >
> > > One could always shrink and re-use previous generation cores, instead of having to re-design
> > > every core type from scratch. Or one could use parameterized synthesizable cores and let core
> > > proliferation be a semi-automatic thing. We do have AI now to handle the bulk of grunt work.
> >
> > Caches seem to have played out this way.
> >
> > The older microprocessors didn't have caches.
> >
> > The z15 had L1 (I and D), L2, L3 and L4 ...
> >
> > If one want to think of "memory hierarchy" rather than cache we have something like:
>
> You missed two levels:
> >

> >
1. Registers

> >
2. L1 cache

> >
3. L2 cache

> >
4. L3 cache

> >
5. L4 cache (on z15 anyway)

> >
6. DRAM

>
>
7. Compressed memory

>
8. Flash

>
> >
9. Disk for paging (HDD or SSD)

> >
10. Tape

> >

Lets add a few more:

1. Registers


2. L1 cache


3. L2 cache


4. L3 cache


5. L4 cache (on z15 anyway)


6. Local DRAM


7. Compressed/encrypted local DRAM


8. Local persistent flash memory (Optane)


9. Remote NUMA memory


10. PCIe device memory (GPU/accelerator)
11. 
    
12. NVMe flash storage


13. Spinning disk


14. Remote system memory


15. Remote system storage


16. Tape

And I'm sure others can be added.

Two things worth noting are the capacities of each tier as well has additional components (including software) in a system to add those tiers to the hierarchy. Things like compressed/encrypted memory can have accelerators to greatly reduce latencies but that adds die space. Remote systems in a cluster invoke numerous layers of hardware and software for access though technologies like RoCE have significantly reduced that.