Best CPU for single-threaded performance by Q4 2014/Q1 2015

[DrMrLordX]

What do you think will be the CPU out there with the best overall single-threaded performance (overclocking included) by the end of 2014? Personally, I am curious as to how this Anniversary Edition Pentium will fare once fully overclocked.

To pre-empt thread derailment, let's just imagine that all the entrants will have very good air cooling (NH-d15 for example), and that multiple background threads will not be a significant issue for the purpose of this discussion.

 

[BrightCandle]

Most likely the Devil's cannon CPUs. By all accounts the Broadwell processors are low power varients and Haswell-E wont be competing for the highest clock speeds due to its core count although in theory it could its not normally how it plays out.

 

[jpiniero]

It's going to be the 4790K. This is Intel we are talking about, don't expect much from the Pentium.

 

[BSim500]

The unlocked Pentium looks a neat little thing to experiment with / for HTPC's / budget gamers. The higher end CPU's (i5/i7) do have more cache though, so overall, it'll probably be an unlocked Devil's Canyon. What would be amusing is if the Pentium was given the same "improved TIM" treatment and could be pushed to 5GHz. :biggrin:

 

[Rakehellion]

It's going to be the 4790K. This is Intel we are talking about, don't expect much from the Pentium.

This.

All the chips have the same basic technology, except i7's have a higher clock speed and more cache. Though the difference in single-threaded performance will be very small.

 

[DrMrLordX]

Hmm! Alright, I was sort of thinking the same thing myself. Technically there's no reason why you couldn't just disable two cores in the UEFI on a 4790k anyway, provided the option exists in the menus. Even that might be of no real benefit.

The unlocked Pentium looks a neat little thing to experiment with / for HTPC's / budget gamers. The higher end CPU's (i5/i7) do have more cache though, so overall, it'll probably be an unlocked Devil's Canyon. What would be amusing is if the Pentium was given the same "improved TIM" treatment and could be pushed to 5GHz. :biggrin:

No reason why the Pentium couldn't get a delid if it has the "bad" TIM. That's outside the scope of what most people want to do with their chips, granted . . .

 

[lamedude]

HP will see the writing on the wall, and to get back at Intel for never delivering on Itanium they will finish the Alpha 21464 along with a Linux port of FX!32. But they will forget to replace the RDRAM IMC so no will be able to afford it.

 

[DrMrLordX]

Linux port of FX32? Lulz. RDRAM is actually pretty cheap nowadays, but I'm sure if HP released that killer Alpha, demand would drive prices back up! Yup yup!

 

[Homeles]

It's going to be the 4790K. This is Intel we are talking about, don't expect much from the Pentium.

Broadwell-K might end up being better.

 

[jhu]

what do you think will be the cpu out there with the best overall single-threaded performance (overclocking included) by the end of 2014? personally, i am curious as to how this anniversary edition pentium will fare once fully overclocked.

To pre-empt thread derailment, let's just imagine that all the entrants will have very good air cooling (nh-d15 for example), and that multiple background threads will not be a significant issue for the purpose of this discussion.

ibm power 8

 

[guskline]

Devils Canyon.

 

[DrMrLordX]

ibm power 8

I was hoping someone would go off the x86 plantation in their reply.

Okay, so why POWER 8? I was under the impression that POWER 8 has the ability to handle an incredible number of threads per core, but that it wasn't (necessarily) faster than Intel's competitive offerings on a per-thread basis.

 

[SAAA]

I was hoping someone would go off the x86 plantation in their reply.

Okay, so why POWER 8? I was under the impression that POWER 8 has the ability to handle an incredible number of threads per core, but that it wasn't (necessarily) faster than Intel's competitive offerings on a per-thread basis.

Indeed Power8 scales very well with SMT2-4-8, to the point that it doubles or more the performance compared to single thread per core. But this means that 1) the core is too wide and underused when loaded with a single thread program 2) being clocked higher than Ivy Bridge and usually reaching only twice the performance (on very parallel server workloads) results in lower IPC, sure less than Haswell and of course you can overclock this to higher speeds.
So Devil Canyon is the right guess , now if that tiny Pentium can reach higher clocks I don't know, maybe under water...
Moral of the story: Intel has still the crown in IPC, the closest one are Apple with A7 (but that's much slower at only 1.4GHz) and IBM. Third and farther in the race is AMD, who hopefully will improve on that in 2016.

 

[DrMrLordX]

That's what I thought, not that there's anything wrong with POWER8 on that account. It's not built for single-threaded apps. It's built for highly-parallel workloads.

 

[SAAA]

That's what I thought, not that there's anything wrong with POWER8 on that account. It's not built for single-threaded apps. It's built for highly-parallel workloads.

Absolutely and it's a monster in its field, Haswell is intended for a broader range of uses instead, from servers to laptops/tablets: that sure brings some cost-savings to develop but has its disadvantages in specialized applications.

 

[shady28]

Spec for a 4 CPU (24 Core\48 thread) 22nm Xeon-8893v2 (3.4Ghz 37.5M cache) is 1010 base 1040 peak.

Spec for a 2 CPU (24 core) 3.5Ghz 32nm Power8 based IBM is 1280 base 1750 peak.

Core for core Power8 beats Intel's top line Xeon at roughly the same clock.

AFAIK Itanium has not been updated since Q4 2012. The closest thing I found in the spec DB from HP was a 32 core Itanium that scored 531 / 507.

Probably worth noting that most of the "run of the mill" "modern" 12-24 core Xeon or Opteron type servers are sub 500 scores though.


Reference :

www.spec.org

http://www.spec.org/cgi-bin/osgresults?conf=cpu2006

 

[DrMrLordX]

Right, but bear in mind that POWER8 can handle more threads per core than Xeon. If you just throw one fat thread at either system, all that parallel processing power ceases to matter.

That proposition seems absurd, and for server workloads, it is. But if you're doing something like . . . I don't know, running a Dead Souls LPMud using whatever bundled driver they're using now (FluffOS?) which is single-threaded, then, guess what, you've got a situation where the Xeon can come out on top. And where 23 out of 24 cores on either system would go almost completely unutilized. Talk about a waste of resources.

People don't live or die (or keep/lose their jobs) over the performance of an LPMud. That's not really the point of this forum thread. It was to figure out what CPU would have the best single-threaded performance within a particular time frame. That doesn't stop POWER8 from having some nice SPEC numbers. Hell, according to SPEC_Int rate, the 3.52 ghz POWER8 has better performance/watt than the Xeons I've seen listed on spec.org.

That being said, a Devil's Canyon chip, overclocked, will smoke both the 4P Xeon and 2P POWER8 machine in single-threaded performance.

 

[thunng8]

That being said, a Devil's Canyon chip, overclocked, will smoke both the 4P Xeon and 2P POWER8 machine in single-threaded performance.

IBM is announcing higher clocked POWER8 chips next week at 4.19Ghz. It was 3.52Ghz which already smokes any Intel chip in multithreaded server workloads (~2X performance per core).

In terms of single thread, the 4.19Ghz POWER8 is likely higher than Devil's Canyon.

 

[VirtualLarry]

Could AMD become one of the OpenPower alliance members, and then build a CPU with an x86/x64 frontend, and a Power8 backend?

 

[Idontcare]

ibm power 8

For single-threaded performance, "overclock included", per the OP?

I'm skeptical.

 

[crashtech]

For ST, nothing will beat an OCed G3258 per dollar. Within 10% or so for less than 1/4 the price. But as stated, the 4790K is gonna be the top of the heap until (maybe) Broadwell K.

 

[ShintaiDK]

4790K. And by the end of 2015 it will be the Skylake variant of 4790K.

 

[pw257008]

4790K. And by the end of 2015 it will be the Skylake variant of 4790K.

Skylake-K is next year, too?

 

[jpiniero]

4790K. And by the end of 2015 it will be the Skylake variant of 4790K.

You definitely won't see Skylake-K in 2015 now that Broadwell-K got delayed again into late 2Q. Unless they end up cancelling Broadwell-K, which could happen.

 

[DrMrLordX]

When I first asked this question, there was the hope that the G3258 might hit 4.8-5 ghz with good cooling. Now we know that doesn't happen so often. So with 4.6-4.7 ghz on Devil's Canyon being generally more common than G3258s hitting the same speeds (4.5 ghz seems more common with them, with some notable exceptions), I'd definitely have to agree that the 4790k wins that comparison in single-threaded performance, by a nose.

The suggestion that a POWER8 at higher clock speeds deserves consideration arouses skepticism, but it remains an interesting notion.

As for AMD cloning POWER8 with an x86 frontend for decode . . . that would be weird. But it would not surprise me if Keller is looking over the design for things he could use on Zen, provided the open license on POWER8 allows that sort of cherry-picking. The one thing AMD doesn't really want is to have yet another design with high power consumption and an intrinsic reliance on workload parallelism for performance. But hey, if they want to go SMT, POWER8's implementation is pretty darn good.

Now, I will say this: AMD is in the console market now. If they wanted to launch a high-clock, 1C/8T POWER8 for consoles, that would/could make a lot of sense, provided they don't want to carry on using their APUs. Not sure if that's really in their best interest though. I could see Nvidia going that way, perhaps.