Question Intel 12th to 13th generation performance comparison

[GunsMadeAmericaFree]

I thought this was an interesting read - benchmark comparisons between Intel 12th generation & 13th generation:

Article with details

That's an average performance increase of 47% from one generation to the next. I wonder if AMD will have a similar increase?

 

[DrMrLordX]

In that situation, sacrificing a couple percent ST performance for ~30% more throughput performance is easily worth it, hence why SMT still exists.

Neither Intel nor AMD are sacrificing ST performance by implementing HT.

 

[Exist50]

Neither Intel nor AMD are sacrificing ST performance by implementing HT.

They do. If not directly, then certainly indirectly.

 

[DrMrLordX]

They do. If not directly, then certainly indirectly.

How?

 

[Exist50]

How?

Supporting SMT clearly takes non-zero engineering effort - effort that could be spent on ST performance (or anything else).

 

[Doug S]

Also, the benefit isn't just the extra ~5% or whatever transistors; it's also the engineer time. How many 10s of thousands or even hundreds of thousands of hours have been devoted to implementing, maintaining, and securing SMT? What could we get if those engineers were devoted to something else? Be that power, area, ST performance, features, whatever.

Whatever that amount of time is, it is a LOT higher today than it was a few years ago before the first SMT based exploits came to light. Given that more variants of the same theme were found since the initial ones, the story isn't over there and there is a lot of engineering time left.

There's also the risk that an exploit is found that's so bad the only defense is to disable SMT, rendering all SMT related expenses wasted for all affected shipping models and future models unless/until the issue can be effectively mitigated.

 

[TheELF]

Why stop now? Because their own creation (the P+E core hybrid x86 processor) is proving to be able to provide superior MT throughput with an e-core quad than the same die space allocated to a P core. Because performance of many applications is highly governed by the performance of just one or two threads, and providing superior single threaded performance for those few threads is important both for application performance and comparative marketing benchmarks. Because they want to make the best possible all-around processor for desktop usage?

Yes, that's why intel and amd have prefered cores that run faster than the others, if they would believe that they could make a core faster without HT/SMT and that it would make sense then they would add one or two much faster cores without HT, but gaining 1-2% without HT is not worth the effort for them having to design another different core because the loss in MT performance will reduce the desirability of these CPUs.
How much trouble did intel have because p-cores and e-cores are different, now they would add yet another different core to the mix, yeah that would go over really well with consumers...

How Intel Technologies Boost Your CPU's Performance - Intel

Innovations like Adaptive Boost Technology make 11th Gen CPUs faster. Learn how Turbo Boost, Thermal Velocity Boost, and ABT work together.

www.intel.la

 

[TheELF]

Also, the benefit isn't just the extra ~5% or whatever transistors; it's also the engineer time. How many 10s of thousands or even hundreds of thousands of hours have been devoted to implementing, maintaining, and securing SMT? What could we get if those engineers were devoted to something else? Be that power, area, ST performance, features, whatever.

We would get less than.
What, you think they don't analyse these things to see what is worth it and what not?! You think they survived for 50 years or what it is because they don't know how to make business?

 

[Exist50]

We would get less than.
What, you think they don't analyse these things to see what is worth it and what not?! You think they survived for 50 years or what it is because they don't know how to make business?

As I said, SMT makes sense in a world where you use one core for almost everything. That's no longer the world we live in.

I think this thread will age quite beautifully.

 

[TheELF]

As I said, SMT makes sense in a world where you use one core for almost everything. That's no longer the world we live in.

I think this thread will age quite beautifully.

Getting more performance out of your CPU will always be making sense,


this thread will age beautifully anyway the future goes since both opinions are represented...

 

[LightningZ71]

Yes, that's why intel and amd have prefered cores that run faster than the others, if they would believe that they could make a core faster without HT/SMT and that it would make sense then they would add one or two much faster cores without HT, but gaining 1-2% without HT is not worth the effort for them having to design another different core because the loss in MT performance will reduce the desirability of these CPUs.
How much trouble did intel have because p-cores and e-cores are different, now they would add yet another different core to the mix, yeah that would go over really well with consumers...

How Intel Technologies Boost Your CPU's Performance - Intel

Innovations like Adaptive Boost Technology make 11th Gen CPUs faster. Learn how Turbo Boost, Thermal Velocity Boost, and ABT work together.

www.intel.la

TVB is essentially a hack. Oh, it's officially supported in that it exists, but, it's essentially saying "what two cores can we push beyond our baseline target successfully on this particular die that is binning better than the rest?" This same die is used down the stack, yet TVB is only enabled on a small subset. Also, and this is a key point, TVB would STILL be available on a hypothetical ST optimized core.

It is my opinion that the 13900K could be a better all-around processor in well over 95% use cases with 6 P cores with 5% better IPC optimized for ST throughput and 24 E cores of the existing design. It would be better than the existing model in every ST benchmark. It would be better than the existing model in every MT benchmark that didn't specifically target 7-8 threads, and, depending on how that benchmark behaved, it still might best it. It would only require two core types. It would be immune to specifically targeted SMT attacks and would be no more vulnerable to shared L2 attacks than the existing product.

Why didn't they choose to do the above? Because they were ALREADY invested in the GoldenCove core, which is an evolution of the previous cove cores in Rocket lake, tiger lake and Ice Lake. If you've already got the product, there's no reason to walk away from it. What we're talking about is the next generation of core after the Coves are finished. Also, don't forget that marketing plays a role in all of this. Against a 16 core, 32 thread 7950x, the 6+24=30 thread conjectural 13900K would be down two threads, even if it outperforms the 7950x across the board, which it would in everything not AVX-512.

 

[DrMrLordX]

Supporting SMT clearly takes non-zero engineering effort - effort that could be spent on ST performance (or anything else).

Intel has supported SMT for years. If there's anything they know how to do right, it's that. It is highly-unlikely that (for example) Golden Cove or Raptor Cove could have been better products if Intel had chosen to remove SMT from the design.

 

[scineram]

Neither Intel nor AMD are sacrificing ST performance by implementing HT.

Maybe as long as you aren't using it.

 

[TheELF]

It is my opinion that the 13900K could be a better all-around processor in well over 95% use cases with 6 P cores with 5% better IPC optimized for ST throughput and 24 E cores of the existing design. It would be better than the existing model in every ST benchmark. It would be better than the existing model in every MT benchmark that didn't specifically target 7-8 threads, and, depending on how that benchmark behaved, it still might best it. It would only require two core types. It would be immune to specifically targeted SMT attacks and would be no more vulnerable to shared L2 attacks than the existing product.

But they do increase IPC by 3-5% every gen, and they do that even while spending all of these resources on HT...

 

[Kocicak]

Intel has supported SMT for years. If there's anything they know how to do right, it's that.

I believe that it was recently necessary to disable SMT in some Intel CPUs due to some security problem? I do not follow the news that closely, I am not sure about this.

 

[JustViewing]

Why we are having the debate that removing HT would increase performance of single thread? Did anyone from Intel or AMD mentioned like that? You can always disable it in BIOS if you chose to. For the rest of us, please let us use the extra 20-40% MT performance.

 

[Kocicak]

For the rest of us, please let us use the extra 20-40% MT performance.

Enjoy your fantasy-land extra performance. The rest of us has to live with the grim reality of weak and neglected second HT threads, which in some case and some load (13900K and Cinebench R23) bring just 15% improvement, and that only if you load the CPU 100% with 32 threads.

If you load it with 24 threads, your poor second HT threads are completely useless.

 

[JustViewing]

Enjoy your fantasy-land extra performance. The rest of us has to live with the grim reality of weak and neglected second HT threads, which in some case and some load (13900K and Cinebench R23) bring just 15% improvement, and that only if you load the CPU 100% with 32 threads.

If you load it with 24 threads, your poor second HT threads are completely useless.

Again what is the harm in having HT? It is not like removing HT somehow reduce the die size by any significant amount. Without HT, there is little motivation in designing wider core. Because, extracting high-IPC is difficult. If anything, future server processors may use 4 HT threads per core if they continue to increase the execution width.

 

[Markfw]

Enjoy your fantasy-land extra performance. The rest of us has to live with the grim reality of weak and neglected second HT threads, which in some case and some load (13900K and Cinebench R23) bring just 15% improvement, and that only if you load the CPU 100% with 32 threads.

If you load it with 24 threads, your poor second HT threads are completely useless.

You are the one in fantasy land. Right now on my windows box, I have 4750 threads according to task manager. Any CPU count up to 4750 would be doing something.

 

[BorisTheBlade82]

That doesnt work like this, ressources are fully shared when there s two threads :

2130 - 1st thread on a P core

Mt score being 2830 :

1415 - 1st thread on a P core.
1415 - 2nd thread on a P core.

Okay, my friend @Abwx, this topic has been bugging me all day. So I wanted to give it a proper test in order to prove my point.
What I did is this: On my only SMT machine (a 5650u Cezanne based HP Elitebook) I created two local copies of CB R23 and renamed one of the executables to "Cinebench II.exe". So I could run entirely separate Instances of it. I then started both in ST mode and assigned them to T0 and T1 of the same core.

The results speak for themselves and what can I say:
You were totally, utterly and undisputedly right, and I was totally, utterly and undisputedly wrong
Both separate instances achieved exactly the same score of 785 while only one instance in ST mode achieved 1264.

Needless to say how shocking this is to me, to realize, that I got something as fundamental as SMT wrong for more than 2 decades - after having spent countless hours of learning about technical aspects of CPUs. Somehow the assumption was burned into my brain, that the second thread only gets the resources that the prime one does not use - and that is also how Intel was explaining it in Layman's terms, when they introduced it. Also, most articles only mention the total benefit of SMT but do not go into detail if the throughput of these two threads is symmetrical or not.
But the positive take-away for me is that I learned (and proved) something fundamental today. Thanks for bringing me on that path

I applied the screenshots to prove your point.

 

[BorisTheBlade82]

Again what is the harm in having HT? It is not like removing HT somehow reduce the die size by any significant amount. Without HT, there is little motivation in designing wider core. Because, extracting high-IPC is difficult. If anything, future server processors may use 4 HT threads per core if they continue to increase the execution width.

Just as has been mentioned before: You spend around 10-15% of transistor budget for it. You are more prone to side-channel attacks like Meltdown/Spectre. Because of that many of your most important customers don't even use it. You spend engineering capacity for it - just to name a few.

 

[Starjack]

Why we are having the debate that removing HT would increase performance of single thread? Did anyone from Intel or AMD mentioned like that? You can always disable it in BIOS if you chose to. For the rest of us, please let us use the extra 20-40% MT performance.

If on the same processor, yes. But when it comes to the E-cores in latest gen processors, particularly in the Alder Lake at least in which is specified that the Gracemont cores perform similar to Skylake, would that be like single threaded E-cores equal Skylake HT cores or single threaded E-cores equal single threaded Skylake cores?

 

[Exist50]

Intel has supported SMT for years. If there's anything they know how to do right, it's that. It is highly-unlikely that (for example) Golden Cove or Raptor Cove could have been better products if Intel had chosen to remove SMT from the design.

Yes, they've supported SMT for years, but that support has surely not been for free, especially with the recent side channel vulnerabilities.

Why we are having the debate that removing HT would increase performance of single thread? Did anyone from Intel or AMD mentioned like that? You can always disable it in BIOS if you chose to. For the rest of us, please let us use the extra 20-40% MT performance.

The discussion is not about disabling SMT on existing architectures, but whether future architectures would not be better off without it. Or rather, better off devoting the resources to some other metric.

You are the one in fantasy land. Right now on my windows box, I have 4750 threads according to task manager. Any CPU count up to 4750 would be doing something.

The vast majority of those threads are likely to be idle at any given time, or otherwise demand so little compute they can be safely discounted.

 

[Exist50]

If on the same processor, yes. But when it comes to the E-cores in latest gen processors, particularly in the Alder Lake at least in which is specified that the Gracemont cores perform similar to Skylake, would that be like single threaded E-cores equal Skylake HT cores or single threaded E-cores equal single threaded Skylake cores?

Gracemont IPC is ~equal to Skylake IPC. So 1T Gracemont vs 1T Skylake at the same clocks will perform similarly. Skylake only has a frequency advantage going for it.

 

[Exist50]

Again what is the harm in having HT? It is not like removing HT somehow reduce the die size by any significant amount. Without HT, there is little motivation in designing wider core. Because, extracting high-IPC is difficult. If anything, future server processors may use 4 HT threads per core if they continue to increase the execution width.

I'm not sure why you conclude that there's no reason to go wide without SMT. The widest and highest IPC mainstream CPU arch you can buy right now is Apple's big cores, which are significantly wider and higher IPC than Intel or AMD's, yet lack SMT.

 

[VirtualLarry]

Are you anti-SMT guys all aware of the different types of parallelism present in instruction stream(s), and the means to exploit them for performance purposes?

There's ILP, TLP, etc., probably a few more that I forgot.

Point is, thread-level-parallism is a real thing. Ignoring it (not implementing SMT/HT) is just leaving performance on the table, so to speak.