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?

 

[dullard]

@GunsMadeAmericaFree Please put this in the Raptor Lake thread instead. Points to consider:
1) Calling this 13 Gen Refresh would confuse people. This is because there may be a future refresh of the 13th generation.
2) This is for multithreaded applications. More cores = better multithreading. So it is a pretty obvious result.
3) Because of #2, I think Intel is missing out on a potential multithreaded monster: 6 P cores, 24 E cores. Instead they went with 8 P Cores and 16 E cores. Both would use about the same amount of power and about the same amount of silicon area. But 6+24 would blow these away in multi-threaded applications. Alas.

 

[Wolverine2349]

@GunsMadeAmericaFree Please put this in the Raptor Lake thread instead. Points to consider:
1) Calling this 13 Gen Refresh would confuse people. This is because there may be a future refresh of the 13th generation.
2) This is for multithreaded applications. More cores = better multithreading. So it is a pretty obvious result.
3) Because of #2, I think Intel is missing out on a potential multithreaded monster: 6 P cores, 24 E cores. Instead they went with 8 P Cores and 16 E cores. Both would use about the same amount of power and about the same amount of silicon area. But 6+24 would blow these away in multi-threaded applications. Alas.

It would have been a disaster if Intel had only 6 P cores on 12th or 13th gen and more e-cores and I would have been one unhappy camper. They could have done a 6 + 24 config in addition to, but not instead of 8 + 16. Either both or just one with 8 P cores and less e-cores.

 

[Kocicak]

6+24 compared to 8+16 means just a small dip in performance of 7+ thread load intensity (decreasing with increasing load intensity) and significant improvement in highest multithreaded load capability.

Refreshed Raptor in 6+16 config and IPC improvement could overall mean notable improvement in performance compared to current 8+16 Raptor. Refreshed Raptor in 6+20 or 24 config would be a monster.

 

[Wolverine2349]

More P cores is better not more e-cores.

 

[Kaluan]

Give us power usage and temp numbers or this data has limited use.

We all know when Intel says 65W TDP they mean 200W++ and no restrictions.

RIP TDP & power states.

 

[Mopetar]

More P cores is better not more e-cores.

Better for what? The reason Intel went with 16 E-cores this time around is that it would take too large of a chip to get that performance with additional P-cores.

I'd say go further and make a 4 P-core, 32 E-core chip. That'd have enough fast high frequency cores to handle foreground applications and a monstrous amount of compute muscle.

 

[Wolverine2349]

Better for what? The reason Intel went with 16 E-cores this time around is that it would take too large of a chip to get that performance with additional P-cores.

I'd say go further and make a 4 P-core, 32 E-core chip. That'd have enough fast high frequency cores to handle foreground applications and a monstrous amount of compute muscle.

Better all around chip for gamers. 4 P cores and a bunch of e-cores ouch. Games have issues with 4 cores now a days and need ore and e-cores are trash for gaming

 

[Kocicak]

I'd say go further and make a 4 P-core, 32 E-core chip.

There could be an option to drop the hyperthreding (because the E cores are more powerful that the second threads anyway) and use all the freed space to make the P cores more powerful or to use that space for additional E cores.

There should also be an option to make the E cores a little bit more powerful (by fitting 3 larger ones instead of 4 in place of one P core).

4 P cores with HT and 32 E cores would transform to 4 P+ cores without HT and 24 E+ cores. This chip overall could have MT performance of a 18-20 P+ core chip, it would be a monster limited only by what power you would be willing to feed it.

 

[Wolverine2349]

More p-cores rather than e-waste cores is way way way better!!!

 

[ondma]

More p-cores rather than e-waste cores is way way way better!!!

Repeating something over an over does not make it correct. Intuitively, yes, it seems more P cores is better. But as others have pointed out, in heavily milti-threaded workloads, more e cores is in fact a better solution.

 

[Wolverine2349]

Repeating something over an over does not make it correct. Intuitively, yes, it seems more P cores is better. But as others have pointed out, in heavily milti-threaded workloads, more e cores is in fact a better solution.

More P cores is always going to be better for consumer as most apps scale to not infinite threads, but like 4-8. Only apps that scale to infinite threads are better with more e-cores. Though more P cores better for consumers as they are more well rounded cores and no hybrid gimmick crap and better for background tasks and multi tasking for apps that scale to finite number of threads.

 

[LightningZ71]

Having profiled many programs in performance troubleshooting, I can safely say that there are precious few desktop apps that most would consider to be non-pro that scale past 4 "performance-critical" threads that aren't specifically highly threaded rendering or encoding programs. Those highly threaded programs almost universally devour as many cores as you can throw at them. By "performance-critical", I mean that that the cores running those threads stay at or near 100% utilization and that the performance of those cores is the main limiting factor on program performance so long as there sufficient background thread resources available to handle all the needed support threads spawned.

In an ideal world, Intel would be able to produce a processor the size of Raptor lake with 6 performance cores that discard SMT for additional single thread performance at all costs along with 24 E-cores to pick up the MT throughput tasks. Yes, that losses 2 threads, but the gains pn the P cores would more than make up for it, and the increase in hardware threads would help overall throughput. There's not a game currently sold at retail that wouldn't gain performance in that case.

 

[TheELF]

Having profiled many programs in performance troubleshooting, I can safely say that there are precious few desktop apps that most would consider to be non-pro that scale past 4 "performance-critical" threads that aren't specifically highly threaded rendering or encoding programs. Those highly threaded programs almost universally devour as many cores as you can throw at them. By "performance-critical", I mean that that the cores running those threads stay at or near 100% utilization and that the performance of those cores is the main limiting factor on program performance so long as there sufficient background thread resources available to handle all the needed support threads spawned.

In an ideal world, Intel would be able to produce a processor the size of Raptor lake with 6 performance cores that discard SMT for additional single thread performance at all costs along with 24 E-cores to pick up the MT throughput tasks. Yes, that losses 2 threads, but the gains pn the P cores would more than make up for it, and the increase in hardware threads would help overall throughput. There's not a game currently sold at retail that wouldn't gain performance in that case.

Yeah but in contrast to your job intel makes desktop CPUs for the masses, people want to be able to run multiple things at once and if you have two or three things with 1-2 critical threads each you want a CPU with enough full cores to handle that.

For people that need an extremely high amount of MT intel makes server CPUs.

Also getting rid of SMT does not increase ST performance, unless it frees up power headroom to increase clocks, but that is a non-factor for modern CPUs, single cores get all the power they want.

 

[ondma]

More P cores is always going to be better for consumer as most apps scale to not infinite threads, but like 4-8. Only apps that scale to infinite threads are better with more e-cores. Though more P cores better for consumers as they are more well rounded cores and no hybrid gimmick crap and better for background tasks and multi tasking for apps that scale to finite number of threads.

Now you are arguing against yourself. If, as you just stated, most apps only scale to "4 - 8 threads" then you don't need more than 8 P cores right??? right??

We will see if the limit of 8 P cores for Intel ultimately comes back to bite them in gaming and perhaps gaming/streaming (unless streaming can be shifted to the e cores), but as of now, an 8 P core 13900k is as fast as, or a bit faster, than 7950x in gaming (with 16 "P" cores), so the lack of P cores does not appear to be a problem.

In any case, I am done with this conversation, as you clearly have a pre-determined idea which is not influenced by other viewpoints.

 

[LightningZ71]

Yeah but in contrast to your job intel makes desktop CPUs for the masses, people want to be able to run multiple things at once and if you have two or three things with 1-2 critical threads each you want a CPU with enough full cores to handle that.

For people that need an extremely high amount of MT intel makes server CPUs.

Also getting rid of SMT does not increase ST performance, unless it frees up power headroom to increase clocks, but that is a non-factor for modern CPUs, single cores get all the power they want.

SMT costs transistors. Period. There are many things that get duplicated for the two threads to work simultaneously. Removing that duplication can allow things like a larger OoO window, deeper buffers, larger caches, etc. All of those things can help single threaded throughput. However, we're very much deep into the area of diminishing returns, which is why both leading x86 chipmakers support SMT on their mainstream and server processors. The increase in MT throughput is, in their mind, worth the very minor hit to ST throughput it costs.

You absolutely can not judge what the possible ST throughput of a given core would be if the SMT transistor count cost was dedicated to ST throughput by turning off SMT of the existing core. All that tells you is if the thermal, power, and scheduling complexity cost of SMT is essentially free with respect to ST throughput.

 

[Carfax83]

Doesn't SMT help single threaded performance indirectly by masking memory latency as well? I read that somewhere a long time ago, that if a core is processing a thread and it stalls for whatever reason, the other thread can continue the process.

Is that true?

 

[Kocicak]

SMT costs transistors. Period.

HT brings 33% more performance in Raptor lake CPUs. There is A LOT of resources used for that and potentially big opportunity to improve ST performance when abandoning hyperthreading. There is no point for it in consumer computers any more, now that we have a lot of cores available and even more smaller cores.

 

[TheELF]

You absolutely can not judge what the possible ST throughput of a given core would be if the SMT transistor count cost was dedicated to ST throughput by turning off SMT of the existing core.

You absolutely 100% can do that for each and every piece of coding you write or just run.
Intel PCM will show you how many instructions per cycle a piece of software will use and there is no way for it to use any more than that unless you come up with a radical new way of coding that piece of code.
(And this is just an example other coding tools allow for a much deeper analysis)

Intel® Performance Counter Monitor - A Better Way to Measure CPU...

The Intel® Performance Counter Monitor provides sample C++ routines and utilities to estimate the internal resource utilization of the latest Intel® Xeon® and Core™ processors and gain a significant performance boost.

www.intel.com

All the CPU makers know how much IPC general code can use and make their CPU cores accordingly.
That's why servers have four way SMT because they use way more much "narrower" threads and desktops only use two way SMT because we use heavier threads in general.

Doesn't SMT help single threaded performance indirectly by masking memory latency as well? I read that somewhere a long time ago, that if a core is processing a thread and it stalls for whatever reason, the other thread can continue the process.

Is that true?

You are already talking about at least two threads, or the same thread cloned on two hardware threads, so you answered your own question.

HT brings 33% more performance in Raptor lake CPUs. There is A LOT of resources used for that and potentially big opportunity to improve ST performance when abandoning hyperthreading. There is no point for it in consumer computers any more, now that we have a lot of cores available and even more smaller cores.

HT brings 33% on average but also much more or much less depending on what you run, and that is because it's not using a lot of resources but because a lot of threads leave a lot of resources unused...
HT uses the UNUSED resources that the first thread leaves untouched, and it leaves them untouched because there is no way for that thread to use them, making the core have even more resources available for the first thread is not going to make that thread use more resources it will just leave even more resources untouched.

We know this stuff for 20 years already.

Intel's Pentium 4 3.06GHz: Hyper-Threading on Desktops

www.anandtech.com

 

[Kocicak]

...
HT brings 33% on average but also much more or much less depending on what you run, and that is because it's not using a lot of resources but because a lot of threads leave a lot of resources unused...
HT uses the UNUSED resources that the first thread leaves untouched, and it leaves them untouched because there is no way for that thread to use them, making the core have even more resources available for the first thread is not going to make that thread use more resources it will just leave even more resources untouched.

We know this stuff for 20 years already.

Intel's Pentium 4 3.06GHz: Hyper-Threading on Desktops

www.anandtech.com

It never hurts refreshing what you know (or should know).

If hyperthreading would be so easy to implement, it would be present on E cores as well. It is not, and the reason for that may not be just that these cores are streamlined and they have much less resources available.

It would really help to know how much physical stuff (and area used) is needed on P cores to make HT run. I still believe that deleting HT would enable increasing single thread performance somehow (the CPU may become simpler in some aspect making some performance increasing measure doable, etc).

One downside for the company is, that they would need to develop two versions of the P core, one for servers with HT and the second without HT for PCs.

Doing what I proposed above (making the P core as powerful as possible with aid of deleting HT and making the E cores a little bit larger) could be a challenging and expensive project, but since Intel is having troubles to capture interest of consumers lately, the resulting product may be very interesting and bring back some attention.

 

[igor_kavinski]

Doing what I proposed above (making the P core as powerful as possible with aid of deleting HT and making the E cores a little bit larger) could be a challenging and expensive project, but since Intel is having troubles to capture interest of consumers lately, the resulting product may be very interesting and bring back some attention.

You are exactly what Intel needs right now to dig a hole they can't get out of

PC enthusiasts salivate thinking about SMT4 and you want to take away SMT altogether. Your ST performance wish, though, may come true if this ever sees the light of day: https://www.blopeur.com/2021/10/30/Intel-VISC-Processor-Architecture-Patent.html

 

[Kocicak]

You are exactly what Intel needs right now to dig a hole they can't get out of

PC enthusiasts salivate thinking about SMT4 and you want to take away SMT altogether. Your ST performance wish, though, may come true if this ever sees the light of day: https://www.blopeur.com/2021/10/30/Intel-VISC-Processor-Architecture-Patent.html

They would just dig the hole larger to enable them running quicker to be able to jump out of it.

I believe that SMT is now with high core count CPUs available completely useless in home PCs for normal non professional use. When a CPU has 8 or 16 E cores - each one more capable than those second HT threads of P cores, who needs those second threads...

 

[TheELF]

They would just dig the hole larger to enable them running quicker to be able to jump out of it.

I believe that SMT is now with high core count CPUs available completely useless in home PCs for normal non professional use. When a CPU has 8 or 16 E cores - each one more capable than those second HT threads of P cores, who needs those second threads...

Yes intel and AMD should remove features and still keep the price at $5-600 and not sell any of the lower products at all anymore like AMD already does for the last few generations because...competition is good?! ...vote with your wallets?! Get less for the same price=good?!

I mean they do see how well this works for nvidia so why not.

 

[Kocicak]

Yes intel and AMD should remove features and still keep the price at $5-600 and not sell any of the lower products at all anymore like AMD already does for the last few generations because...competition is good?! ...vote with your wallets?! Get less for the same price=good?!

I mean they do see how well this works for nvidia so why not.

I am not sure I understand your post, because I was writing about ways to ADD performance to the CPUs. If removing HT could really have a positive effect, only a small fraction of performance (the extreme end of multithreaded loads) would suffer.

 

[TheELF]

I am not sure I understand your post, because I was writing about ways to ADD performance to the CPUs. If removing HT could really have a positive effect, only a small fraction of performance (the extreme end of multithreaded loads) would suffer.

Removing HT would reduce multithreaded performance of the main 6-8 cores by at least 30% and down to half as fast in some cases, the added speed to single thread would be nothing or at the very best next to nothing.
So how does that ADD performance?!