From: anton@mips.complang.tuwien.ac.at
Michael S writes:
>The question is what is most likely meaning of the fact of patenting?
>IMHO, it means that they explored the idea and decided against going in
>this particular direction in the near and medium-term future.
>
>I think that when Intel actually plans to use particular idea then they
>keep the idea secret for as long as they can and either don't patent at
>all or apply for patent after release of the product.
>I can be wrong about it.
That would risk that somebody without patent exchange agreements with
Intel patents the invention first (whether independently developed or
due to a leak). Advantages of such a strategy: Companies with patent
exchange agreements learn even later about the invention, and the
patent expires at a later date.
I remember an article about alias prediction (IIRC for executing
stores before architecturally earlier loads), where the author read a
patent from Intel and did some measurements on a released Intel CPU,
and confirming that they actually implemented what the patent
described.
If you find that article, and compare the data when the patent was
submitted to the date of the release of the processor, you can check
your theory.
>Some of them 1 year ago gave representations
>about advantages of removal of SMT.
I did not read any accounts of that that appeared particularly
knowledgeable. What are the advantages, or where can I read about
these presentations?
>Removal of SMT and this super-core
>idea can be considered complimentary - both push into direction of
>cores with smaller # of EU pipes.
What do you mean by that? Narrower cores? In recent years cores seem
to have exploded in width. From 1995 up to and including 2018 Intel
produced 3-wide and 4-wide designs (with 4-wide coming IIRC with Sandy
Bridge in 2011), and since then even the Skymont E-core has grown to
8-wide, with 26 execution ports and 16-wide retirement. And other CPU
manufacturers have also increased the widths of their CPUs.
It seems that there has been a breakthrough in extracting ILP, making
wider cores pay off better, a breakthrough in designing wider register
renamers and making other structures wider, or both.
Pushing for narrower cores appears unplausible to me at this stage.
Concerning the removal of SMT, I can only guess, but that did not
appear unplausible to me with Intel's hybrid CPUs: They have P-cores
for fast single-thread performance, and lots of E-cores for
multi-thread performance. You allocate threads that need
single-thread performance to P-cores and threads that don't to
E-cores. If you have even more tasks, i.e., a heavily multi-threaded
load, do you want to slow down the threads that run on the P-cores by
switching them to SMT mode, also increasing the already-high power
consumption of the P-cores, lowering the clock of everything to stay
within the power limit, and thus possibly the performance? If not,
you don't need SMT.
Still, after touting the SMT horn for so long, I don't expect that
such considerations are the only ones. There must be a significant
advantage in design complexity or die area when leaving it away
(contradicting the earlier claim that SMT costs very little).
Concerning super cores, whatever it is, my guess is that the idea is
to try to extract even more performance from (as far as software is
concerned) single-threaded programs than achievable with the wide
cores of today.
>Anyway, couple of months ago Tan himself said that Intel is reversing
>the decision to remove SMT.
On the servers, they do not follow the hybrid strategy, for whatever
reason, so the thoughts above don't apply there. And maybe they found
that the cloud providers want SMT, in order to sell their customers
twice as many "CPUs".
- anton
--
'Anyone trying for "industrial quality" ISA should avoid undefined behavior.'
Mitch Alsup,
--- SoupGate-Win32 v1.05
* Origin: you cannot sedate... all the things you hate (1:229/2)
|
