GLOBALFOUNDRIES Introduces New 12nm FinFET Technology for High-Performance Applications

[Yotsugi]

Show me one example where a Ryzen 6C/12T catches up with Intel CL 6C/12T.

Are we talking clock per clock?

 

[krumme]

The time difference was mentioned in relation to the quote "Intel has been ahead in process technology for quite some time"

1. The THG context was about AMD processors, not GPUs. But even then, it'd be 4 years instead of 5. Your point in disproving the original claim was....?
2. SMT has literally nothing to do with process technology. SMH.

Please, in your next try, on another topic, try to answer the original question without introducing irrelevant tangents.

Instead of beeing rude read the quote again slowly. You will find Samsung and Tsmc mentioned in the same sentence.
Yes. Your eyes caught GF but thats because your brain is obsessed with Intel vs amd as 95% in this forum. You are blind from it.
May i remind you that those companies isnt excactly dwarfs vs Intel. They have increasing been pouring B into process development the last handfull of years. Tools and knowledge is "shared". The arm eco system makes Intel small. We see the result now. Each day in the hands of tens of millions of people.

 

[ajc9988]

The only reason Ryzen catches up while running slower is because it is 8 cores vs 6.

Show me even one example where a Ryzen 6C/12T catches up with Intel CL 6C/12T.

Wrong. You have to control for core and thread count, then ipc and speed advantage. Only after those are controlled for can you then compare threading performance. Also, you can see the 1600X vs the 7800x that in heavily multithreaded tasks, it wasn't, percentage based, as far behind as single threaded, if the ipc and speed is controlled for. That decrease in percentage in threaded tasks over single thread tasks shows HT versus SMT. It is also why single threaded performance of coffee went up, but multithreaded was practically flat as the extra two cores only amounted to about 50% performance increase over Kaby. Learn what is being controlled and look at proper reviews which show these things.

 

[ajc9988]

Instead of beeing rude read the quote again slowly. You will find Samsung and Tsmc mentioned in the same sentence.
Yes. Your eyes caught GF but thats because your brain is obsessed with Intel vs amd as 95% in this forum. You are blind from it.
May i remind you that those companies isnt excactly dwarfs vs Intel. They have increasing been pouring B into process development the last handfull of years. Tools and knowledge is "shared". The arm eco system makes Intel small. We see the result now. Each day in the hands of tens of millions of people.

In fact, Intel used to be the third largest fab and is now the fourth. In fact, Samsung, this year, surpassed them in microprocessors. The reason those fabs are bigger is they do so many types of products, as you suggest. So, you are definitely correct there and people need to realize Intel isn't number one in fab, nor is it any longer the top semiconductor producer. Things have changed...

 

[krumme]

Yes.

Agree. Only fanboys can miss the words Samsung and Tsmc in a sentence because it contains the word GF that draws all the attention. That takes some weird glasses.

 

[PeterScott]

Wrong. You have to control for core and thread count, then ipc and speed advantage. Only after those are controlled for can you then compare threading performance. Also, you can see the 1600X vs the 7800x that in heavily multithreaded tasks, it wasn't, percentage based, as far behind as single threaded, if the ipc and speed is controlled for. That decrease in percentage in threaded tasks over single thread tasks shows HT versus SMT. It is also why single threaded performance of coffee went up, but multithreaded was practically flat as the extra two cores only amounted to about 50% performance increase over Kaby. Learn what is being controlled and look at proper reviews which show these things.

Adding 50% more cores is only going to ever add 50% more performance at maximum. You really don't know what you are talking about if you think it could be more.

 

[ajc9988]

Adding 50% more cores is only going to ever add 50% more performance at maximum. You really don't know what you are talking about if you think it could be more.

You misunderstand, ipc and speed of coffee single core increased, but the additional cores result in flat scaling. That means, due to inefficiencies of HT, the increased speed and IPC over Kaby was eaten up by the inefficiency of the thread scaling, so that 6 cores acts as just adding cores, not as cores+. Let's say IPC increased 10%. So single thread is 110%. But, if 6 cores is only 600%, not 660%, then the inefficiency is somewhere in the scaling. No one has 100% scaling currently. But this is why you calculate, after controlling IPC AND SPEED, the difference between the chips under single AND multi. If the multi gap closes from the single, but the speed on all cores was equal, and the same program under the same conditions was used, and the program was built for heavy multithreading, what does that leave. You then attribute the amount accordingly.

Evidently you don't understand what is being discussed.

Edit: also, there is more than one element that effects core scaling. You have SMT/HT, cache systems, latency, etc., that can effect thread scaling. So, this is a down and dirty review described above, whereas in real life you would need to control for more variables. Intel wins on latency and still loses on scaling, which suggests implementation of the mesh, the cache system isn't optimized for in software, and multithreaded system of HT. So please do more homework before responding.

 

[PeterScott]

You misunderstand, ipc and speed of coffee single core increased, but the additional cores result in flat scaling. That means, due to inefficiencies of HT, the increased speed and IPC over Kaby was eaten up by the inefficiency of the thread scaling, so that 6 cores acts as just adding cores, not as cores+. Let's say IPC increased 10%. So single thread is 110%. But, if 6 cores is only 600%, not 660%, then the inefficiency is somewhere in the scaling. No one has 100% scaling currently. But this is why you calculate, after controlling IPC AND SPEED, the difference between the chips under single AND multi. If the multi gap closes from the single, but the speed on all cores was equal, and the same program under the same conditions was used, and the program was built for heavy multithreading, what does that leave. You then attribute the amount accordingly.

Evidently you don't understand what is being discussed.

Edit: also, there is more than one element that effects core scaling. You have SMT/HT, cache systems, latency, etc., that can effect thread scaling. So, this is a down and dirty review described above, whereas in real life you would need to control for more variables. Intel wins on latency and still loses on scaling, which suggests implementation of the mesh, the cache system isn't optimized for in software, and multithreaded system of HT. So please do more homework before responding.

All the handwaving in the world doesn't disguise your basic mistake.

As I stated, adding 50% more cores, the best you can hope for in a perfect Mult-threaded situation, with perfect scaling is 50%. That is the maximum theoretical improvement. Your complaint about "only" 50% is absurd, because that is perfect scaling.

Address your basic mistake before digging a deeper hole.

 

[ajc9988]

All the handwaving in the world doesn't disguise your basic mistake.

As I stated, adding 50% more cores, the best you can hope for in a perfect Mult-threaded situation, with perfect scaling is 50%. That is the maximum theoretical improvement. Your complaint about "only" 50% is absurd, because that is perfect scaling.

Address your basic mistake before digging a deeper hole.

So, I'll say this one more time. This comes from Intel comparing coffeelake to Kaby. They showed about a 10% scaling over Kaby in ipc on a specific task. They showed only about 50% scaling over Kaby in multi-thread application. That means that the IPC and speed gains were destroyed due to multi thread scaling inefficiencies. It is specifically always a comparison to something and what you failed to grasp was the comparison point. If a single thread is faster, but the scaling makes it so that adding cores makes it perform the same as a quad core, but with two more cores, then you have loss somewhere in the core scaling. Period.

Then, you have the comparisons of a 1600X and a 7800X in cinebench 15. In single threaded, with both chips overclocked and it set to all core overclock, Intel beats AMD by X%. In multithreaded, Intel only beats AMD by Y%, with Y%<X%. So, there are factors that contribute to the multithreaded performance that are showing a better scaling with AMD than Intel because the percentage was lower. Cinebench uses small enough caching that this isn't really an issue, although latency can be to get it to the cache. So, it doesn't perfectly get to HT versus SMT, but it gets close, which is how you address thread scaling, in part.

So think more critically next time, because a flat scaling when single threaded performance increases actually means you are not at 100% scaling, otherwise you would get the perfect single thread performance increase to happen on every core, which a flat scaling shows you no longer have that. Do you understand yet? In other words, you are not getting 10% over the 50% that came with the extra cores due to scaling inefficiencies.

 

[PeterScott]

So, I'll say this one more time. This comes from Intel comparing coffeelake to Kaby. They showed about a 10% scaling over Kaby in ipc on a specific task. .

Keep digging that hole, by adding more errors to your argument.

There is no IPC improvement in CL, perhaps you just forgot the single thread clock speed boost improvement?

At locked clockspeed. KB vs CL perform the same. Check the top two performers:

https://www.gamersnexus.net/images/media/2017/CPUs/8700k/8700k-cinebench-1t-production.png

The top one is 2% faster, and is running at 2% higher clock speed.

So no change in IPC at all. IPC of CL = IPC of KL.

 

[ajc9988]

Keep digging that hole, by adding more errors to your argument.

There is no IPC improvement, perhaps you just forgot the single thread clock speed improvement?

At locked clockspeed. KB vs CL perform the same. Check the top two performers:

https://www.gamersnexus.net/images/media/2017/CPUs/8700k/8700k-cinebench-1t-production.png

The top one is 2% faster, and is running at 2% higher clock speed.

So no change in IPC at all. IPC of CL = IPC of KL.

Fine, I meant "performance" not "IPC" and was using it as an example from an Intel slide. My overall point, though, is correct. Instead of digging critically on the measure, you are attacking superficial numbers from an Intel slide. The point remains, when you compare six core to six core, with all cores fixed, between Intel and AMD chips (1600X and 7800X), you see the percentage of performance over AMD be a lower percentage than the single thread lead over AMD. That drop in performance is related to thread scaling, as it says AMD's chips don't drop in scaling as much as Intel's. How do you not understand this? You could replace any name and any generations in. Make up a company and plug those numbers in. It is to explain, generally, the concept. Also, it shows Intel rebranding/badging tech, which is also laughable, but not the point.

I'll find the comparative data to show this later (have work and trying to find comparable data from the same reviewer to best compare what I'm discussing.

 

[PeterScott]

Fine, I meant "performance" not "IPC" and was using it as an example from an Intel slide. My overall point, though, is correct. Instead of digging critically on the measure, you are attacking superficial numbers from an Intel slide.

You overall point is wrong. For the third time 50% scaling from 50% cores is perfection, not poor performance. You made that incorrect claim, and compounded it by defending your position by equating clockspeed increase with an IPC improvement.

As far as AMDs small extra benefit from HT/SMT vs Intel, you need to realize, that how much room there is for SMT/HT improvement, depends on how good your front end is at efficiently utilizing your execution units.

The worse your front end is, the more idle time your execution units have, and the more room there is for HT/SMT to benefit.

So bragging about the SMT/HT boost, is bragging about how bad your front end is at utilizing your execution units.

 

[SpaceBeer]

I mentioned SMT just as an example that if company A has been doing something for some time (e.g. 15 years), that doesn't neccesseraly means, company B needs the same amount of time to catch up. So if intel started with FinFET 5 years ago, and all the others have been doing that for 2 yeas, that doesn't mean they are exactly 3 years behind intel, and they'll reach intel's level in 2020.

 

[DrMrLordX]

So how 'bout that 12nm process?!?! Yeah! Gotta love it. Wait are we still discussing that?

 

[Excessi0n]

There really isn't anything left to discuss about the 12nm process unless GloFo or AMD talk about it more.

 

[itsmydamnation]

Given CFL hasn't seemed to eat into Ryzen sales yet, i dont think they will talk unless t. When CFL actually becomes available and if it majorly impacts Ryzen Enthusiasts sales then expect to start to hear stuff. Otherwise i dont expect to hear much up until it is launched, no need to Osborne if you dont have to.

 

[IRobot23]

Given CFL hasn't seemed to eat into Ryzen sales yet, i dont think they will talk unless t. When CFL actually becomes available and if it majorly impacts Ryzen Enthusiasts sales then expect to start to hear stuff. Otherwise i dont expect to hear much up until it is launched, no need to Osborne if you dont have to.

CF lake was complete paper launch (disaster), yet nobody says anything about it.

Yet again suddenly every review is pointing out that i7 8700/K is worht it? Seriously? 20% difference for double the price? Nonsense! I mean A YEAR AGO MOST OF THEM WERE COMMENDING i5 OVER i7 (KABY)... and suddenly i7 is MUST have? Sure.

Intel pushed CF lake... well why didn't they just do kaby lake 6 core on Z270. Well I won't be buying Intel for a long time.

 

[dark zero]

Meanwhile... Wondering if Mediatek starts to use the process to get cheap processors.