David Kanter (dkanter.delete@this.realworldtech.com) on May 10, 2013 12:34 am wrote:
> Brett (ggtgp.delete@this.yahoo.com) on May 9, 2013 8:26 pm wrote:
> > Paul A. Clayton (paaronclayton.delete@this.gmail.com) on May 6, 2013 9:00 pm wrote:
> > > With respect to Silvermont's microarchitecture, I am somewhat disappointed by the dropping of SMT. This
> > > decision probably makes sense for tablets, phones, and some embedded uses, but it might be less good for
> > > server workloads (though SMT is less useful in a narrow [and somewhat shallow] OoO microarchitecture).
> > > (Since, as far as I know, Intel's SMT implementation does not support software setting of thread priority,
> > > embedded uses that could have real-time software benefits from multithreading presumably cannot fully
> > > exploit such multithreading benefits.) I admit, I am irrationally fond of multithreading.
>
> > Even in server land actual use of SMT is a niche market limited to certain databases and >a few other tasks.
>
> That's not true at all. If you look at server applications, the vast majority benefit from SMT. Pretty
> much every server chip with meaningful marketshare has some form of multithreading: Itanium, Power7,
> Fujitsu's SPARC, Sun's SPARC, Intel x86. The only ones missing it are AMD and IBM zseries.
>
> > Anyone want to start a betting pool on when Intel dumps SMT for mainstream CPU's?
> > The upcoming Tock, or the one after that? I guess you could also drop SMT on a Tick >update.
>
> Sure I'll take you up on that one. I wager Intel will continue
> to use SMT on Skylake, Skymont and whatever comes next.
>
> David

More so, if Intel ever again designs a mainline core with servers as the primary target (Nehalem was the last one like that and, probably, the only one in the whole Intel x86 cores history) then it will have no less than 4 threads.

However I expect that for foreseeable future laptops and high-end tablets remains the primary design target of mainline cores, and so, the number of SMT threads will remain at 2.
If Intel feels pressure from the competing "throughput" stuff on the server front, like somebody doing 50+ Cortex-A53 cores with competent cache/memory and right set of I/Os, or Oracle getting its acts together with pricing of their 1-2 way machines, or who knows what else, then I rather see Intel making throughput chip of their own by cutting KC's successor of 3/4th of its SIMD stuff (4-way SMT is already here), but not affecting mainline design goals. Of course, internal Intel politics can turn another way and they'd do something silly instead, like pushing lots and lots of Silvermont cores into single chip.