AMD vs Intel Single threading?

[Sargatanas2511]

But you can't really use games to benchmark a single component like that.

Nor could you, for example, use a system where there is a lot of (storage) io going on to benchmark the delta performance from a CPU overclock.

That might be "real world" but it does not in any way answer the question as to whether CPU performance scales linearly with clock frequency.

Any realistic way to test performance available shows less scaling as clocks increase so that's all that can be used as a reference, I have really been trying to find that old Intel video but it was something like 2005 and I just can't find it. Even things like Zen 1, Ryzen 1700 vs 1800 the clock speeds and performance results scale wasn't linear. This is harder to do now with Zen 2 as everything is more variable and relative though.

 

[Robert896r1]

Any realistic way to test performance available shows less scaling as clocks increase so that's all that can be used as a reference, I have really been trying to find that old Intel video but it was something like 2005 and I just can't find it. Even things like Zen 1, Ryzen 1700 vs 1800 the clock speeds and performance results scale wasn't linear. This is harder to do now with Zen 2 as everything is more variable and relative though.

no it doesn’t. You can easily see linear scaling in cinebench single and multi core as long as you don’t change other parameters except frequency. Why do you keep saying this? It’s easy to test and some of us have done so many times.

we don’t need 2005 videos.

 

[CAT-THE-FIFTH]

That's the first I'm hearing about it.

Why wouldn't it be linear, or near as dammit?

In any case, real workloads were shown where the advantage was ~xx%.

Not theoretical, but real, measured workloads.

e: finding the figures...

This is because you hit other limitations - one of these can be even a simple factor such as external memory bandwidth,let alone internal core bottlenecks.

 

[Panos]

Still this discussion? Impressed that is still moving in circles.

Some are arguing on a thread the OP hasn't contributed since 9th Dec!!!

 

[FoxEye]

This is because you hit other limitations - one of these can be even a simple factor such as external memory bandwidth,let alone internal core bottlenecks.

In that case surely the same applies to AMD's IPC advantage. If AMD has a ~15% IPC advantage over Intel, then using the same logic this doesn't translate into a 15% improvement (at the same clock speed).

IPC improvements and clock speed bumps are two ways of accomplishing the same thing: more CPU instructions per unit time.

People are strangely writing off Intel's clock speed advantage as "not translating to real world performance", but not doing the same for AMD's IPC advantage.

But all this theorycraft is pointless anyhow - the benchmarks show the real story. Intel is still ahead in workloads such as the workloads demonstrated by @MartinPrince. And not by the "3% or something" claimed by the poster above - by 10% or more wasn't it?

 

[CAT-THE-FIFTH]

In that case surely the same applies to AMD's IPC advantage. If AMD has a ~15% IPC advantage over Intel, then using the same logic this doesn't translate into a 15% improvement (at the same clock speed).

IPC improvements and clock speed bumps are two ways of accomplishing the same thing: more CPU instructions per unit time.

People are strangely writing off Intel's clock speed advantage as "not translating to real world performance", but not doing the same for AMD's IPC advantage.

But all this theorycraft is pointless anyhow - the benchmarks show the real story. Intel is still ahead in workloads such as the workloads demonstrated by @MartinPrince. And not by the "3% or something" claimed by the poster above - by 10% or more wasn't it?

Because it isn't linear - if you don't believe me,look at the Intel CPUs when under normal boost conditions and when overclocked,it does not follow that you get linear improvements in all games. You need to also understand,that games don't just respond to single threaded performance,they can respond to latency and better optimisations for Intel CPUs. I should know as I run one such game,ie,Fallout 4 which uses the Creation engine.

This game will never get a single update to Ryzen. AMD has to brute force that game - user testing shows despite Zen 2 having superior IPC,its still clock for clock similar to Skylake,so at best maybe around a mildly overclocked Core i7 6700K level.

This can be shown by a simple metric - look at the mesh and ring bus Intel CPUs. They have the same basic cores with the same IPC,and both can clock to similar amounts. Yet despite so much more memory bandwidth the X series can be inferior to the desktop ring bus CPUs,and even Ryzen. Fallout 4 scales with memory bandwidth in a way few games do.

You know what improved Zen performance for games like Fallout 4 in each generation?? It was latency improvements,and memory controller improvements. SKL-X was SLOWER than Ryzen in this game.

https://cdn.sweclockers.com/artikel/diagram/14689?key=51b6c375f557cb62e67cbdd5bfddb87f

** Do not hotlink images **

The Core i9 7900X boosts upto 4.3GHZ,which is more than the Core i5 7600K and it has more than 4 cores. The game uses upto 6 threads,with the first two prioritised. The SKL-X mesh bus is higher latency for games.

Fallout is one of the WORST games for Ryzen. People here didn't want to believe me.

In non-gaming benchmarks such as audio processing stuff,you saw how latency also affected performance in Ryzen,and as each generation made improvements it got better.

The same happens with AMD though if you just look at single threaded performance - have you noticed that a Ryzen 5 3600/3600X and a Ryzen 9 3950X are very close in gaming performance,despite the latter clocking much higher. In gaming a Ryzen 5 3600/3600X is most of a Ryzen 9 3950X despite having less than half the cores and lower clockspeeds.

The only games which do better are those which use more than 6 cores.

Also those benchmarks Martin Prince were only ran,because both of us use the same software and we were interested to see how DxO scaled in the workload he has. Like me he uses DxO for its Prime noise reduction,but he does a single image,and then sends it to PS or LR for further processing. However,I tend to run the software in batches as it will scale upto 16 threads or so,and then do any imports into additional software,so we do the same thing but in different way. So compared to the Core i5 8400 which I could also afford at the time(it was £50 to £60 more),the Ryzen 5 2600 I had was a better choice,which is was one of the main reasons I bought it. My original target was a Core i7 8700 non-K but the price went over £300,due to retailer greed,so they could keep it.

 

[CAT-THE-FIFTH]

Also Ryzen latency is more defined as inter-CCX latency and memory-CCX latency,so there are two areas to look at here. There are cache latencies too,but this was more of a problem for Zen1,which Zen+ and Zen2 kind of fixed.

Apparently the rumors I am hearing,is that Zen3 will move to an 8 core CCX,instead of dual 4 core complexes. That would actually help a lot in my view,in games,especially some of the older engines. I don't know how much memory-CCX latency can be reduced,but those large L3 caches we see in Zen2,AFAIK are meant to help in this regard.

But still you have the problems with optimisations for older games which might not happen,so in this case Zen has to brute force things.

 

[FoxEye]

I did say games weren't a good metric for determining the scaling of CPU performance with frequency.

As mentioned by a couple of us now, you'd need to run a synthetic CPU benchmark. Without using a lot of storage IO either.

Otherwise I can run a pure storage benchmark (with some slow storage) and show *no* difference between a 2GHz CPU and a 5GHz CPU

 

[CAT-THE-FIFTH]

I did say games weren't a good metric for determining the scaling of CPU performance with frequency.

As mentioned by a couple of us now, you'd need to run a synthetic CPU benchmark. Without using a lot of storage IO either.

Otherwise I can run a pure storage benchmark (with some slow storage) and show *no* difference between a 2GHz CPU and a 5GHz CPU

But AFAIK,you are buying a CPU for gaming,and unfortunately for AMD,a lot of older games,or games based on older engines won't be updated. Its why when you switch benchmarks to older games and indeed older APIs,Intel does relatively better. AMD has to brute force it,so its not just single core performance,but other design aspects people need to take into consideration.

Now I am also not saying AMD won't be "good enough" in those older gamers,but what is defined as "good enough" I suppose is dependent on the gamer,and now much the difference is in actual framerates and frame consistency.

Also its not only gaming,as I mentioned audio processing benchmarks,which are latency dependent,which again the first two generations of Zen were at a disadvantage to Intel,although not sure about Zen2 in that regard TBF,although I think it is better.

There are also other things for example - Adobe Lightroom,which does exceptionally well on Zen2 based CPUs:
https://i.imgur.com/IkN4Pwy.png

Batch processing in LR scales with threads(but not really after 16 threads apparently) and also single core performance,but look at where Zen2 is?? It defies the normal scaling you would expect with any other CPU,and people think it is down to the huge L3 caches. So in theory it has very high performance/core but you are not measuring single core performance only,its other aspects of the design.

Single threaded performance in any given application isn't just IPC and clockspeed,its also affected by latency,caches and other aspects of the uncore. Just as a mesh vs ring bus,can make cores with the same IPC and clockspeed perform radically different in certain workloads.

Its why as much as all these general benchmarks are fine and dandy,any CPU purchase has to be dependent on the workloads YOU are running. This is why people cannot just dismiss out of hand an Intel CPU being a better choice,for specific workloads you need to run and vice versa,as well as other non-CPU specific aspects which might influence a purchase.

If the Intel CPUs hadn't gone ******* crazy in price I might have even had a Core i7 8700 non-K instead of what I purchased right now.

 

[Panos]

Also Ryzen latency is more defined as inter-CCX latency and memory-CCX latency,so there are two areas to look at here. There are cache latencies too,but this was more of a problem for Zen1,which Zen+ and Zen2 kind of fixed.

Apparently the rumors I am hearing,is that Zen3 will move to an 8 core CCX,instead of dual 4 core complexes. That would actually help a lot in my view,in games,especially some of the older engines. I don't know how much memory-CCX latency can be reduced,but those large L3 caches we see in Zen2,AFAIK are meant to help in this regard.

But still you have the problems with optimisations for older games which might not happen,so in this case Zen has to brute force things.

The removal of CCX and the new Zen3 design is basically 8-core CCD with unified 32MB+ L3 accessible to all
Current Zen2 each CCD is 2 CCX of 4 cores sharing 16MB L3 (the 4 cores of the CCX).
Also we will get a faster new design I/O which would also drop latency also.

 

[CAT-THE-FIFTH]

The removal of CCX and the new Zen3 design is basically 8-core CCD with unified 32MB+ L3 accessible to all
Current Zen2 each CCD is 2 CCX of 4 cores sharing 16MB L3 (the 4 cores of the CCX).
Also we will get a faster new design I/O which would also drop latency also.

That is what I am thinking also,so I am hoping a Ryzen 5 4600 will work in my B450 motherboard.

 

[Phil2008]

Im guessing our cpu's are too fast for superpi now?

 

[MartinPrince]

Software power monitoring......


Yeah its really not that much of a stretch given that OC 5.2Ghz 9900KS chips run at around 250 watts in Cinebench.

Stretching and distorting reality (till it breaks) is pretty much all you seem to be doing in this thread:

• The thread is about single threaded performance bet yet you 'stretch' to focus on IPC as if that is the only factor.
• I refer to DXO Photolab yet you somehow 'stretch' to talk about Lightroom because you can find some charts which show Lightroom on top (even though they contain multithreaded components)
• I refer to a 9700k but you 'stretch' to refer to the 9900k because that uses more power so you can talk about power consumption which isn't the topic of the thread.
• I post up a this real world chart of the results from you, I and several other members here but you somehow 'stretch' to call them 'anecdotal' even though it is empirical data.

It's hard to grasp the gross hypocrisy when you say;

...especially if they claim as Intel do these days in their clown world that modern real world applications' performance is not what you're supposed to be using to measure performance.

but yet when modern real world application performance is used in the form of DXO Photolab (with yours 37% behind) you say you are being trolled and run back crying to your 'charts' because;

Charts are hard data, they trump anecdotal claims.

though, as has been established you don't really know your 'anecdotal' from your 'empirical'.

@Grim5
My 9700k @ 5.3Ghz needs 1.295v so the voltage is not as low as it appears in the HWInfo64 screenshot as I use DVID offset so you're just seeing the voltage fluctuation. It might be reasonable to surmise that I'm as you say a " lucky chip lotto winner" but that is not the reality.
I have built 8 systems for friends/colleague using the R0 9700K, all were L918 or newer which means they are all from Malaysia plant & later than week 18 2019. Six clocked to 5.2Ghz @ ~1.3v and two 5.1 Ghz @ ~1.3v. Non of them were using the same Z390 Aorus Master as me and some were on Z370 so I can't be sure that non of the 5.2Ghz could have done 5.3Ghz in a better motherboard.
Even Anandtech reached 5.3Ghz on a Z370
https://www.anandtech.com/show/13400/intel-9th-gen-core-i9-9900k-i7-9700k-i5-9600k-review/22

@CAT-THE-FIFTH
Puget have their own Photoshop and Lightroom benchmarks out. From viewing the per core performance meter, Lightroom has much more fully multithreaded components due to the batch tests. Photoshop has much less and is more an indication of apps/games that can't max out all threads. I think I'll start a new thread for people to post up both results so that we can have a good spread of results. I'll post up the Photoshop scores here as they're more applicable to this thread topic.

It's important to note that enabling OpenCL depending on the GPU can make quite a difference so I have shown both for the 3900x and 9700k.

If anyone else wants to run the Puget Photoshop benchmark it can be found here.
https://www.pugetsystems.com/labs/articles/PugetBench-for-Photoshop-1132/

you can get a 7 day free trial of Photoshop here (make sure to cancel before 7 days):
https://www.adobe.com/uk/products/p...5T&mv=other#mini-plans-web-cta-photoshop-card

3900X No OpenCL

9700K No OpenCL

3900X with OpenCL

9700K OpenCL

 

[CAT-THE-FIFTH]

@MartinPrince I think you are missing the point of why I talked about Lightroom. The Lightroom batch tests have nothing to do with core count as its beating Intel and AMD CPUs with much bigger core counts. The scaling shows this with Ryzen 3000 tests,as they defy normal scaling - its down to the huge L3 caches probably.



If you were using Lightroom exports as an indication of single core performance,Zen2 would have the highest single threaded performance of any CPU tested. But everyone knows this is not single threaded performance or multi-threaded performance being measured.

It also extends to other Lightroom tests too:
https://www.pugetsystems.com/pic_disp.php?id=59510



I cropped out some of the other Intel and AMD CPUs,but as you can see many of these operations are not multi-threaded(core counts don't affect the score in a significant way),and Intel CPUs don't always win or even if they win its rarely in any meaningful way despite this.

Not all tests are due to simple metrics such as single core performance,but things such as latency,uncore performance(L3 cache being one of them),etc.

In fact what is called IPC,is really just what you define as the physical core,outside things such as the L3 caches,or even support for certain extensions such as AVX technically are not defined as IPC(they are called the uncore). This is why forum estimations of single threaded performance might not be testing IPC,or even single threaded performance.

Also regarding your OpenCL benchmarks,you missed one thing there - they were done with a Nvidia graphics card,and in some games it was shown Nvidia drivers did have some issues running on Zen compared to Intel equivalents,when similar performance AMD cards were used. I would like to see an AMD graphics card run side by side to that - both Vega and Navi. Its never been really investigated outside gaming.

In fact I would argue most people running either Intel or AMD systems will be running slower RAM,so I would like to see some scaling tests in Adobe software with different speed RAM. Even memory bandwidth can actually affect performance,which is again not a single threaded measure,but something else(and yes Intel can run faster RAM so there is still an advantage there for such bandwidth dependent workloads as a platform so cannot be discounted).

 

[CAT-THE-FIFTH]

Puget also has tested most of the new AMD CPUs with Photoshop CC last month:
https://www.pugetsystems.com/labs/a...formance-AMD-Threadripper-3990X-64-Core-1655/



Looking at the test scores in seconds,there are few tests I can see Intel significantly is faster in,some AMD is faster in,but a whole bunch where they are margin of error between a whole lot of CPUs.

Also things get worse when you look at price. The Ryzen 7 3700X is now £260 from multiple retailers with a solid stock cooler which will give you most of the performance out of the box. The cheapest place I can see a Core i7 9700K is £339,plus the cost of the aftermarket cooler if you want to overclock it or even just run it at stock(I don't overclock BTW and haven't done so for a decade). The Ryzen 7 3700X,consumes as much power as my Ryzen 5 2600,so a £70 to £90 motherboard will do fine,and for someone using a mini-ITX rig,you probably suspect what is a better choice for me!

The saving is enough to buy more RAM,or more SSD storage for example. The problem is until the next Intel range is released Intel pricing isn't really competitive.

I do think the removal of the SMT tax will make the new Intel CPUs quite competitive in more multi-threaded benchmarks,unlike some here. The Ryzen 5 series will be now fighting against an equivalent thread count Intel Core i5.

 

[humbug]

A 12 thread 10600K will be much more competitive vs a Ryzen 3600 if the price is right, the 3600 is currently £155, if Intel charge their usual £210 its a lot more money.

Intel's 10K series is also going to be competing with Zen 2 soon after release.

 

[FoxEye]

A 12 thread 10600K will be much more competitive vs a Ryzen 3600 if the price is right, the 3600 is currently £155, if Intel charge their usual £210 its a lot more money.

Intel's 10K series is also going to be competing with Zen 2 soon after release.

And hopefully Zen 3 early next year at the latest

 

[MartinPrince]

@MartinPrince I think you are missing the point of why I talked about Lightroom. The Lightroom batch tests have nothing to do with core count as its beating Intel and AMD CPUs with much bigger core counts. The scaling shows this with Ryzen 3000 tests,as they defy normal scaling - its down to the huge L3 caches probably.

If you were using Lightroom exports as an indication of single core performance,Zen2 would have the highest single threaded performance of any CPU tested. But everyone knows this is not single threaded performance or multi-threaded performance being measured.

The chart you demonstrated above is a batch test. I'm not sure you've worded that correctly as ANY batch test is always FULLY MULTI-THREADED. Consequently, they have a lot to do with core count. Extra cache will aid performance though the main reason for the better performance of the 3900x is the full usage all core clock speed + IPC. The Threadripper 2950x is ~3.5ghz compared to the 3900x which is ~4.2Ghz plus it has higher IPC (~18% I believe). This alone could account for the 3900x's 12/24 better performance than the 2950x 16/32 not including the larger cache:
2950X = 16x3500 = 56,000
3900x = 12x4200 x1.18(IPC) = 59,472

Either way that graph is definitely a fully multi-threaded graph and I can pretty much guarantee you that every CPU on that list was maxing out all of its cores when doing that batch test. As such it is not really the topic of this thread, neither is power consumption or value for money etc. We're just concerned with single threaded performance or as I prefer, non-fully multi threaded performance.

With this being an overclockers forum, a few folks do run overclocked CPU's. The big issue for me is that most of the charts and reviews do not compare standard overclocked systems (High end air/AIO).

As I've demonstrated, there can be quite a disparity in the results when comparing stock systems with overclocked ones. This is more pronounced in software that uses less than the maximum threads available.

The fact that Humbug, an experienced member with over 30,000 posts, can be so off base believing his Ryzen 3600 with PBO would be faster than an overclocked 9700K in single threaded (or non fully multithreaded) software makes me wonder if there is a general lapse of understanding of;

1. How software utilises cores/threads
2. How overclocking on both platforms works.

If you look at DXO Photolab on Humbug's post you will see it uses about 8 threads.

My 3900X exhibits a similar behaviour:

Though look at the core clock speeds for the 3900X taken during just the 18secs while DXO Photolab was exporting during the test.
The columns are Current - Minimum - Maximum - Average

You can see that 8 cores are using the average of ~4.25Ghz

This is the 3900X core clock during the ~9min Single Thread run on Cinebench r20

You can see that 1 core did achieve 4.6Ghz though the average of that core (along with 1 other) was 4.5Ghz. The other cores are way below it.

We all know that the advertised clock speed of Ryzen 3900X is 4.6Ghz but what a few people don't understand is that under most conditions where software is using several threads then it does not reach that speed, plus you normally can not do an all core overclock exceeding 4.6Ghz. Most people will be able to do all core overclock 4.3-4.4Ghz. Though as Grim5 stated earlier in the thread, doing an all core overclock on Ryzen 3000 will lessen Single-Threaded performance.

Compare the DXO Photolab test of the 9700k at stock speeds.

and now overclocked.

The Cinebench R20 Single-Threaded difference between Humbug's 3600X and my 9700K is ~12%. The Geekbench single threaded benchmark difference is 20% (I made a mistake earlier running Geekbench in 64bit, the 32bit difference is 20%) but the real world performance in software that can leverage ~8 cores/threads then the difference is ~35%.

Also if you look at the Photoshop Puget Overall score for the top rated systems score is ~940 for the 3900X, 3960X and 9900K, my 9700K scores 1136, which is ~20% faster. This is simply the current position when it comes to overclocked systems on both platforms and has been for the last 6 months. This may well change in the next 6 months but that is another topic.

 

[Joxeon]

Though as Grim5 stated earlier in the thread, doing an all core overclock on Ryzen 3000 will lessen Single-Threaded performance.

Not always, maybe on the higher end chips but on the lower end 3600/3700x it can be advantageous as if those can manage a 4.2-4.4 all core which can match the chips max single core boosts.

 

[MartinPrince]

Not always, maybe on the higher end chips but on the lower end 3600/3700x it can be advantageous as if those can manage a 4.2-4.4 all core which can match the chips max single core boosts.

Actually that is very true, as the stock boost speed on the 3600X is 4.4Ghz then doing an all core clock of 4.4Ghz will only be beneficial. My 3900x will do all core of ~4.325Ghz and this gives me lower Single-Threaded performance.