Broadwell architecture seems fundamentally clock limited

[Arachnotronic]

With the reviews/overclocking results in for the Broadwell-E chips, I am forced to conclude that there is something fundamental to the Broadwell CPU core architecture that limits its clock speeds, even relative to 22nm Haswell-E. I don't blame it on the 14nm process because the Skylake core hits higher frequencies as well as higher perf/clock than Broadwell.

Thoughts?

 

[exar333]

No surprises there. There were a lot of issues with Broadwell all around and many products skipped from Haswell to Skylake directly. Skylake is fantastic though...

 

[FIVR]

FIVR



Broadwell has it, skylake doesn't.

 

[PPB]

And who says Skylake desktop is made on the same process variant that Broadwell is?

 

[Zodiark1593]

Broadwell (and probably the process as well) was likely tuned for reduced power use at the expense of maximum clocks.

 

[The Stilt]

Aren't all Broadwells made on P1273 (SoC) while Skylake is the only one on P1272 (HP, CPU)? Broadwell-E 10/15/24 are pretty large dies, so it would make sense to use the denser process on them to optimize the cost.

 

[BigDaveX]

FIVR



Broadwell has it, skylake doesn't.

So does Haswell-E, and that apparently overclocks better than Broadwell-E does.

 

[therealnickdanger]

So does Haswell-E, and that apparently overclocks better than Broadwell-E does.

Perhaps since HSWE is a larger node, the impacts are lessened.

 

[Arachnotronic]

Aren't all Broadwells made on P1273 (SoC) while Skylake is the only one on P1272 (HP, CPU)? Broadwell-E 10/15/24 are pretty large dies, so it would make sense to use the denser process on them to optimize the cost.

Nope, Broadwell is made on P1272. The only products made on P1273 are the Atom SoCs. None of the Core products are built on the SoC processes at Intel.

 

[IndyColtsFan]

I'm waiting to hear overclocking results from the community on the 6800, 6850, and 6900 before I buy. If Haswell-E is significantly better at overclocking on average, it might make sense for me to look at 5820K or 5930K.

 

[Arachnotronic]

I'm waiting to hear overclocking results from the community on the 6800, 6850, and 6900 before I buy. If Haswell-E is significantly better at overclocking on average, it might make sense for me to look at 5820K or 5930K.

My chip was supposed to be here yesterday but Newegg seems to have botched my order. Tracking says that the chip hasn't even actually shipped yet even though I paid for one day shipping. Grr...

 

[IndyColtsFan]

My chip was supposed to be here yesterday but Newegg seems to have botched my order. Tracking says that the chip hasn't even actually shipped yet even though I paid for one day shipping. Grr...

Which model did you get?

 

[Arachnotronic]

which model did you get?

6800k.

 

[IndyColtsFan]

6800k.

Cool, I'm looking forward to hearing your results. I am kicking myself for not getting the $275 5820K or the $420 5930K over the holidays.

 

[Hi-Fi Man]

It might indeed have something to do with the architecture. My i7 5775C hits a wall at 4.2GHz and needs tons of voltage to go higher. Even the L3 cache doesn't OC as high.

 

[frozentundra123456]

Isnt it the same basic architecture as haswell? I blame 14 nm.

 

[dark zero]

FIVR and FinFet itself.... both are not combination...

Also... seems that near all FinFets are not built for extreme overclock.

 

[MrTeal]

FIVR and FinFet itself.... both are not combination...

Also... seems that near all FinFets are not built for extreme overclock.

Umm... the 5GHz+ 6700k on 14nm is also Finfet, as is my (with a larger overvolt) 4.8GHz on standard water HW-E chip. If Broadwell-E has trouble moving past low 4GHz, Finfets themselves aren't likely to blame.

 

[dark zero]

Umm... the 5GHz+ 6700k on 14nm is also Finfet, as is my (with a larger overvolt) 4.8GHz on standard water HW-E chip. If Broadwell-E has trouble moving past low 4GHz, Finfets themselves aren't likely to blame.

The proportion of 5+ Ghz FinFet chips are noticeable lower than older process like SB or even Haswell.
Also remember that Haswell-E has more cores than 6700K that is just 4.

 

[frozentundra123456]

Umm... the 5GHz+ 6700k on 14nm is also Finfet, as is my (with a larger overvolt) 4.8GHz on standard water HW-E chip. If Broadwell-E has trouble moving past low 4GHz, Finfets themselves aren't likely to blame.

5ghz Skylake is quite rare I think. According to Silicon lottery, I seem to recall only about 3% of 6700k will reach 5ghz.

 

[SAAA]

Isnt it the same basic architecture as haswell? I blame 14 nm.

Rather than just 14nm I blame large dies on 14nm. It's been a difficult node since the beginning and while there might be less problems now in 120mm^2 Skylake, so they can finally reach high clocks, a die double or more the size might still have limits. Maybe in a few months it will get better.

Expect much more headroom in Skylake-X, both from architecture and improved 14nm process.

 

[rpmtl]

Tracking says that the chip hasn't even actually shipped yet

Are you in the US or Canada? They seem to have messed up my shipping too and provided a Canada Post tracking number even though their shipping confirmation says it used Purolator. I can track it directly on the Canada Post website (Purolator set me straight about who's number it was), but the Newegg site still says "We do not yet have any record of the tracking number you've entered".

I think Newegg's shipping system is having issues.

 

[rpmtl]

It will be interesting to learn how well a single core can be OC'd rather than all cores. That might OC higher while generating less heat. The 'Turbo Boost Max 3.0' could prove to be a great new feature if it allows one to OC only a single core while running a certain game/app while using the rest of the cores at their native clock for something like video rendering.

 

[MrTeal]

The proportion of 5+ Ghz FinFet chips are noticeable lower than older process like SB or even Haswell.
Also remember that Haswell-E has more cores than 6700K that is just 4.

You contradict yourself. How can the proportion of 5GH+ Finfet chips be noticeably lower than Haswell, when Haswell is also Finfet?

 

[FIVR]

Ivy Bridge was the first architecture with a fully integrated voltage regulator (FIVR)

Sandy Bridge -> Ivy Bridge = Huge reduction in average attainable overclock (over 500mhz), oddly enough coinciding with the introduction of FIVR!

Ivy ->Haswell = Essentially the same, or slightly better overclock on same process (22nm) both with FIVR

Haswell -> Broadwell = Reduction in peak overclock, coinciding with the introduction of Intel's new (and highly problematic and disappointing) 14nm, both retain FIVR and both never reach the average peak overclock of 2011 Sandy Bridge on 32nm.

Broadwell -> Skylake Increase in peak overclock, coinciding with the REMOVAL of FIVR and on the same 14nm process as Broadwell. STILL DOES NOT REACH AVERAGE PEAK OVERCLOCK OF 2011 Sandy Bridge on 32nm


My conclusion from the above information, is that both FIVR and 14nm are bad for overclocking and is why Broadwell - E will likely be inferior to Haswell- E and especially Sandy-E in terms of peak attainable overclock.


I believe Intel 14nm has superior characteristics in many other respects to 22nm and 32nm, but peak overclock in Ghz is not one of them. It is an inferior process for high clock speeds, and FIVR is part of the problem with that. However, I believe the benefits of FIVR outweigh the downsides for most people. FIVR is what allowed for the huge increase in battery life from 2011-2013 on x86 portables.