Technical Question: Why can Intel clock so much higher?

[ZenMonkey]

Comparing the 8700k and 2600x, both 6/12 chips, the Intel can get 700 to 1000MHz higher sustained clock speeds depending on silicon lottery without extreme cooling solutions, so I'm curious on the technical side of things why Intel's chips can sustain much higher clock speeds than their Ryzen counterparts and why AMD doesn't make changes to close the gap.

[Silentprototipe]

Im guessing its just how the architectures work. Coffee Lake and Zen+ are just made differently and take different clocks 

[Mayclore]

Oh, man, where to start.

 

There are a number of possibilities here. CPU architecture design and process node quality are two places to start; Intel's 14nm is a highly mature process at this point, while Zen was AMD's first crack at it and Zen+ is their first crack at 12nm. As for node quality, well, Intel owns the fabs that make its chips, like AMD used to. AMD now contracts out its fabrication to GlobalFoundries and Samsung, companies which may not have the same fabrication quality as Intel's fabs do. There's a reason Intel dumps so much dolla-dolla into those fabs.

 

It's also probably a function of the way the architectures work. Zen is a multi-core module design connected via the Infinity Fabric, while Coffee Lake is just a monolithic CPU die like we're all used to.

[Levent]

smaller core die size is also a big factor.

[Valkyrie Lenneth]

bcoz amd stands for anti mhz delivery

[Mira Yurizaki]

How fast a CPU can operate is solely based on its architectural design at the hardware level. The biggest factor is what the slowest stage in the execution pipeline is, and at that point you have two options to deal with it:

• Optimize the step
  • This is harder than it looks
• Split the step into more steps
  • This has to be done before any real hardware has been made, because it requires architectural changes.

Intel posted a blog related to this topic that is a good read: https://software.intel.com/en-us/blogs/2014/02/19/why-has-cpu-frequency-ceased-to-grow

[ScratchCat]

19 minutes ago, Levent said:

smaller core die size is also a big factor.

Core size would not matter - A53s only clock to around 2GHz while ARM A7X designs clock at 2.3+GHz.

Die size matters as the larger the die the more likely a frequency limiting fault will occur in the chip.

[ZenMonkey]

22 minutes ago, Mayclore said:

It's also probably a function of the way the architectures work. Zen is a multi-core module design connected via the Infinity Fabric, while Coffee Lake is just a monolithic CPU die like we're all used to.

 

This reminds me I read or watched something recently talking about Intel's Ring vs Infinity Fabric, and how the ring is much faster for now but loses out once CPUs pass the 10-12 core mark.

[Mayclore]

1 minute ago, ZenMonkey said:

 

This reminds me I read or watched something recently talking about Intel's Ring vs Infinity Fabric, and how the ring is much faster for now but loses out once CPUs pass the 10-12 core mark.

We'll see. It feels like we're finally headed toward a software ecosystem where everything wants as many cores as you can supply. 7nm Zen 2, if they can get a whole bunch of cores on Socket AM5 (or whatever they're gonna call it) is going to be a very interesting prospect indeed.

[Mira Yurizaki]

4 minutes ago, ZenMonkey said:

This reminds me I read or watched something recently talking about Intel's Ring vs Infinity Fabric, and how the ring is much faster for now but loses out once CPUs pass the 10-12 core mark.

Which is why Intel switched to using a mesh network in their higher core count processors. The thing with AMD is they're using two-level communication topology for inter-core communication whereas Intel seems to want to stick with one. There are pros and cons to each though.

 

However, I feel that Infinity Fabric will still run into scalability issues sooner rather than later. Especially with the multi-die approach that Threadripper and EPYC took (that's a lot of effing traces).

[ScratchCat]

11 minutes ago, Mayclore said:

We'll see. It feels like we're finally headed toward a software ecosystem where everything wants as many cores as you can supply. 7nm Zen 2, if they can get a whole bunch of cores on Socket AM5 (or whatever they're gonna call it) is going to be a very interesting prospect indeed.

The issue is that after a certain number of cores you tend towards a limit due to Amdahl's Law and that embarrassingly parallel tasks (which problem which can run on 32+ cores often is) simply are faster on GPUs. The main task where these high core count CPUs excel is rendering - the cores of the GPU are often not flexible enough to run the task (GPU based encoding uses hardware accelerators, Premiere uses some GPU accelerated effects though).

 

[agent_x007]

40 minutes ago, M.Yurizaki said:

How fast a CPU can operate is solely based on its architectural design at the hardware level. The biggest factor is what the slowest stage in the execution pipeline is, and at that point you have two options to deal with it:

• Optimize the step
  • This is harder than it looks
• Split the step into more steps
  • This has to be done before any real hardware has been made, because it requires architectural changes.

Intel posted a blog related to this topic that is a good read: https://software.intel.com/en-us/blogs/2014/02/19/why-has-cpu-frequency-ceased-to-grow

You know that Netburst (a.k.a. Pentium 4/Pentium D) was desinged to have the shortest stages possible ?
High frequncy number failed it at 5GHz and 65nm, while it was desinged to do 10GHz+ range.

Pesonally I think end frequency is ~70% transistor technology, ~20% architecture dependant and last ~10% is left for "fiddling around" (once you can do once previous two are established).

Reason :
Past 32nm for Intel, transistor switching time isn't most important - low power usage at "normal"/production frequncy was given priority (because of "Silicon power wall" hit first with Netburst at high clocks).
That's why Ivy Bridge can't reach 5Ghz speeds easily, while most unlocked Sandy Bridge class CPUs can do it without problems (despite having almost the same architectures).
On the other hand, Netburst could do Air cooloed 5GHz already on 65nm (but in it's case I think it was because of building it that way, and not really because it was designed for it [Prescott was designed for 5GHz+ clocks at minimum]).

Chip sizes matters for frequency in few areas : 
1) Multi-frequency regions (Core vs. "North Bridge"/IMC/L3 clock),
2) Heat output (size x frequency = big heat problems), 
3) Cost effectiveness (no point in selling the biggest core and highest frequency possible CPUs if you can't make money out of it).
^It can mitigated by Very High margins on some models (ie. high end Xeons and Extreme Edition CPUs)

[Mira Yurizaki]

1 minute ago, agent_x007 said:

Pesonally I think it's ~70% transistor technology, and ~20% architecture dependant.

 

Last ~10% is left for "fiddling around" you can do once previous two are established.

I don't think it's transistor technology so much as how you use them. Unless you're talking about the power wall. Then yeah, that'd be a problem. Otherwise, 5.0+ GHz processors has been a thing for 8 years now (the IBM z196).

[agent_x007]

30 minutes ago, M.Yurizaki said:

I don't think it's transistor technology so much as how you use them.

I guess you can hamper switching speed of fast transistors if you screw up your desing, but, I don't think it's something anyone would do (it's counter intuitive) ?
Implementation is meangless, if your basic building block is slow.
In other words : I don't see why/how a "slow-er transistor" (made for efficiency), could switch faster than a "fast" one (made for speed), if both were used for the same application.

Car analogy :
True, a car with a 1000 BHP engine can be slower than a car with 500 BHP engine.
But, why would you made it that way ?

[Mira Yurizaki]

56 minutes ago, agent_x007 said:

I guess you can hamper fast transistors if you screw up your desing, but I don't it's something would anyone do (it's counter intuitive) ?

The design will only present theoretical solutions. Until you actually implement it and analyze it, you can't say for certain a design actually meets a performance requirement or not.

 

Also optimizations can be counter intuitive if you don't understand the system you're working with very well. In software, the XOR swap method on the surface can be seen as faster or better than the temp swap method because you're using less memory. But in reality, the XOR swap method is only good if you are memory constrained, with the performance benefit being nothing to worse (especially if the CPU has implemented ILP).

 

I'm not going to pretend such things don't exist in hardware land.

Quote

In other words : I don't see why/how a "slow transistor" (made for efficiency), could switch faster than fast one (made for speed) if both were used in the same design.
Implementation is meangless if your basic building block is slow.

Well of course. I'm not arguing that the same design is magically faster on slower technology.

[ZenMonkey]

2 hours ago, M.Yurizaki said:

Intel posted a blog related to this topic that is a good read: https://software.intel.com/en-us/blogs/2014/02/19/why-has-cpu-frequency-ceased-to-grow

 

That was a really informative article.

[Biggerboot]

At it's most practical, Hz can only be taken at face value on the same architecture (an overclock cpu vs stock).  It's a single variable in a complex equation. 

I don't even think explaining some of the features of a specific cpu artictecture can accurately explain how intel's and amd's manufacturing technique yield different results.  Besides the fact that a lot of it is over my head, there's a lot of trade secrets involved in how they achieve what they achieve.

[DarkSmith2]

Haha, where to start.

because  Global Foundries 14nm =/= TSMC's 14nm.

AMD is cheating on us when talking about transistorsize. Like they did with Zen+, they are calling it 12nm even if it has the completely same size as before, 12nmLP=14nmLPP, they may have increased density a little but transistor size remained the same, there wasnt any shrink applied  

In reality its more about 22nm.. in size btw. Thats also the reason why its performing on Haswell levels. Haswell had 22nm aswell. Now imagine how much power a 4770k would need with 8cores to get above 4.3GHz  And how much heat it would produce. 

 

Its just all about the naming scheme.

 

They got pretty used to it. It started back in the day with the AthlonXP.. yea well AthlonXP3000 had 2.1GHz not 3GHz sorry mates. Or calling a Bulldozer "8Core" CPU. Just lieing straight in your face before you even buy it. That is what AMD does... strange marketing. And it works

[Biggerboot]

11 minutes ago, DarkSmith2 said:

---

I'll just put that all under the 'trade secrets' umbrella that I mentioned.

However AMD wants to market is fine by me, as long as it's making Intel work harder.  Benchmarks will always be your best friend in comparing chips.

[DarkSmith2]

1 minute ago, Biggerboot said:

I'll just put that all under the 'trade secrets' umbrella that I mentioned.

However AMD wants to market is fine by me, as long as it's making Intel work harder.  Benchmarks will always be your best friend in comparing chips.

yep, we all love thus Ryzens. Im thinking about buying one for myself to use it as a dedicated streaming rig. Cant get this SMT performance somewhere else for such little money.