Intel "Bartlett Lake-S" Gaming CPU is Possible, More Hints Appear for a 12 P-Core SKU

[PrettyKitten800]

That Hybrid design gave them back the market. Until the 12th gen, Intel was losing the battle against AMD, because AMD was offering 16 core CPUs when Intel was having a difficulty to even bring out a 10 core CPU. And it was understandable considering AMD had a significant manufacturing advantage thanks to TSMC and Intel's manufacturing problems. With those Hybrid CPUs and those marketing cores, Intel managed to match AMD in core count and even surpass AMD in core count latter. Where AMD was offering 6 core CPUs, Intel was advertising 10 core CPUs. AMD was losing this battle, until X3D CPUs where introduced and Intel started facing it's own problems with degradation of 14K CPUs and performance regression with it's newer CPUs.

Those silly pointless efficient cores that I was calling from day one "marketing cores" was a genius move and I would have given to those who thought it, the biggest bonus Intel have ever given to an employee. And considering they dropped HyperThreading, I guess they will go the "efficient way" by adding more and more E cores in their future processors. Because almost no one reads "This CPU comes with X threads". Almost everyone reads "This CPU comes with X cores". And 99% of people can't even realize that their 10 core monster is slow, because only 2 of those cores are P cores.

Those little cores are a huge marketing advantage.

It’s really naive to label hybrid CPUs as a marketing gimmick. There are zero companies that do not utilize hybrid core designs, even AMD, the only company still holding onto hyperthreading.

 

[efikkan]

What makes people think that a CPU lineup intended for "edge" and "networking equipment" (in other words embedded/light servers) will be targeting the gaming market?
The reason why Raptor Lake have a slight edge in high-FPS gaming over Arrow Lake is thanks to its absurd boosting, it would be very strange for this kind of lineup to do this.

 

[igormp]

even AMD, the only company still holding onto hyperthreading.

Nvidia's latest Vera ARM CPU also has SMT.

 

[chrcoluk]

@AleksandarK you sure its not LGA1700? I tried to verify but its all still linked to 1700.

 

[matar]

Must be with HT to make it the perfect CPU otherwise not good.

 

[Lionheart]

Come on Intel make a comeback, and what others have said, give us more performance cores, Would love to see a 12 Performance Core only Intel Chip with HT returning.

 

[ejolson]

It’s really naive to label hybrid CPUs as a marketing gimmick. There are zero companies that do not utilize hybrid core designs, even AMD, the only company still holding onto hyperthreading.

To simply hardware and software engineering AMD uses the same Zen5 core with same instruction set from Ryzen to Epyc. The only thing that changes is the amount of cache. Moreover, except for the two X3D exceptions, all cores in a single processor package have the same amount of cache.

As far as I can tell software complications from having a heterogeneous assortment of cores turned out more difficult than expected. On the other hand, slowing the clocks and reducing voltage gives very similar energy savings for a performance core without having to divide engineering efforts between two types of cores.

An all P-core processor from Intel sounds very good to me.

 

[MaMoo]

There is nothing that just uses 2 or 4 threads. But why the heck are we even taking about 2P core cpus. Even the cheapest lowest of low end desktop i3 has 4.

I've been screaming in every thread about how crap the 12p core will be compared to 8+16 for a year now. It will be slower in everything.

Identical cores are better for symmetric multiprocessing. I use that a lot for prototyping scientific workloads and for stats. E-cores just do not work because the entire execution is waiting for the slower cores to finish before the next batch starts. E-cores will always be busy while the P-cores have completed their tasks. The cinebench type of benchmarks are all designed to be more asymmetrically scalable but there are plenty of workloads that interleave large and identical sized chunks of work that must retire together. I had a bunch of e-cores, I disabled them and got a much better throughput.

 

[Visible Noise]

So, an application that uses for example max 4 threads, will be faster on a 2P+4E CPU than on a 4P CPU. An application that uses max 8 threads will be faster on a 4P+8E CPU than on an 8P CPU. An application that uses max 12 threads will be faster on a 8P+8E CPU than on an 12P core CPU etc.

When Intel goes to 12P core CPUs people like you will be screaming about how much better a 12P core CPU is compared to an 8P core CPU because X game/application can utilize all those extra P cores. For now you will be simply saying "That's just absurdly wrong".
I am fine with that.

Open up task manager. How many threads does it say you have running?

Why would you want all those competing for P-core resources along with your foreground applications?

 

[JustBenching]

Identical cores are better for symmetric multiprocessing. I use that a lot for prototyping scientific workloads and for stats. E-cores just do not work because the entire execution is waiting for the slower cores to finish before the next batch starts. E-cores will always be busy while the P-cores have completed their tasks. The cinebench type of benchmarks are all designed to be more asymmetrically scalable but there are plenty of workloads that interleave large and identical sized chunks of work that must retire together. I had a bunch of e-cores, I disabled them and got a much better throughput.

Highly doubt it. Are you turning HT off for your symmetric multiprocessing? Cause those are much slower than ecores.

 

[john_]

Open up task manager. How many threads does it say you have running?

Why would you want all those competing for P-core resources along with your foreground applications?

If you have the option to buy 2 Ferrari or 1 Ferrari and 2 Smart for the same price, are you going with the Ferrari plus 2 Smart option because for example, Smart is more cost effective? Come on. It's funny how people who talk about efficient cores are the same people who celebrate when a 300W Intel Core CPU beats a 170W AMD CPU. Really funny, in the beginning. Pretty boring after awhile.

 

[JustBenching]

If you have the option to buy 2 Ferrari or 1 Ferrari and 2 Smart for the same price, are you going with the Ferrari plus 2 Smart option because for example, Smart is more cost effective? Come on. It's funny how people who talk about efficient cores are the same people who celebrate when a 300W Intel Core CPU beats a 170W AMD CPU. Really funny, in the beginning. Pretty boring after awhile.

Invalid comparison. A cpu with a hybrid design is as fast in st and faster in MT compared to a similarly sized (and therefore cost) cpu without ecores. That's the whole point of ecores, they are faster in terms of performance / die.

Take a practical example, my 12900k is 8+8. If you offered to take away the 8e to give me 2 extra pcores I'd laugh you out of the room.

The 170w amd cpu is tdp, it actually draws 220 to 250, the exact same wattage the 285k draws. I know, I currently have on an apex, peaks at 212w during cbr23. But that's completely irrelevant, why did you even bring that up?

 

[Zazigalka]

Finally, been waiting for a 10900 followup forever

 

[efikkan]

Must be with HT to make it the perfect CPU otherwise not good.

Arrow Lake has already proven that SMT is only going to become less and less useful.
SMT "extract" performance from two things; a) Idle cycles from a stalled core b) Competing for resources with another thread.
As architectures get more effective a) has become less of free performance to harvest, and this will only continue until its nearly pointless. There will be exceptions of course, but the exceptions are getting fewer and fewer. SMT will hold on for a while in the server space though, but eventually all the design and security implications are going to force it away.

Open up task manager. How many threads does it say you have running?

Why would you want all those competing for P-core resources along with your foreground applications?

Some applications (especially browsers) and background processes tend to spawn more threads the more "cores" a CPU have, so there is a strong argument for having fewer stronger cores than many weak ones. While most of these fairly idle, they do still require scheduling resources.

A cpu with a hybrid design is as fast in st and faster in MT compared to a similarly sized (and therefore cost) cpu without ecores. That's the whole point of ecores, they are faster in terms of performance / die.

Hybrid designs are mostly a gimmick outside mobile devices, but the big PC vendors largely sell new PCs based on specs, and when clock speeds and proper cores don't scale as before, they need another gimmick to sell upgrades. You can see this in their marketing; 24 "cores" and up to 5.7 GHz at 65W, misleading the customer to thinking this is a performance beast when it really isn't. E-cores are incredibly weak when you put load on all of them, as groups of them share L2-cache.

There are also the scheduling issues; even though Linux has proven to be far more efficient than Windows in this regard, still it's no cakewalk. If this is going to work nearly as you pretend it to be, we would either need hardware level scheduling or at the very least vastly different OS kernels than today.

Take a practical example, my 12900k is 8+8. If you offered to take away the 8e to give me 2 extra pcores I'd laugh you out of the room.

Then the joke is on you, as you clearly make your decisions on arbitrary specs you don't understand.
I'll happily prefer 10p over 8p+8e, preferably of the much more performant Arrow Lake generation, as it wouldn't be hit-and-miss with the scheduler and give much more consistent performance. But you can cling to your synthetic benchmarks all day long…

 

[JustBenching]

Hybrid designs are mostly a gimmick outside mobile devices, but the big PC vendors largely sell new PCs based on specs, and when clock speeds and proper cores don't scale as before, they need another gimmick to sell upgrades. You can see this in their marketing; 24 "cores" and up to 5.7 GHz at 65W, misleading the customer to thinking this is a performance beast when it really isn't. E-cores are incredibly weak when you put load on all of them, as groups of them share L2-cache.

If MT performance is a gimmick then sure, hybrid designs are a gimmick, since they are faster in MT than non hybrid designs at the same die space.

Then the joke is on you, as you clearly make your decisions on arbitrary specs you don't understand.
I'll happily prefer 10p over 8p+8e, preferably of the much more performant Arrow Lake generation, as it wouldn't be hit-and-miss with the scheduler and give much more consistent performance. But you can cling to your synthetic benchmarks all day long…

The joke is on me for preferring a faster CPU but not on you that youll happily prefer a slower one, and your preference is literally based on arbitary specs (a thing you accused me off, lol). I prefer what's faster, 8+8 is faster than 10+0. You on the other hand pick based on arbitrary specs. Enjoy I guess

 

[chrcoluk]

Windows needs a feature where you can hide core's from specific processes (or other way round only show to specific processes), that would solve hybrid scheduling headaches.
SMT is at least on Intel side inefficient, e-cores reign supreme over it. But bartlett lake interests me, because as a gamer and someone who is still dominated by per core performance I will prefer that CPU for the forseeable future. But no question for things like compiling e-cores are very nice.

 

[efikkan]

If MT performance is a gimmick then sure, hybrid designs are a gimmick, since they are faster in MT than non hybrid designs at the same die space.

Stop it with your straw man arguments, I never said multithreaded performance was a gimmick. Grow up.
The fact remains, in most real world scenarios, especially in desktop usage with applications (+typically a browser in the background) and various backgroud processes, one p-core outperforms one group of four e-cores easily. There are exceptions, but those are mostly either edge-cases or purely synthetic.

Windows needs a feature where you can hide core's from specific processes (or other way round only show to specific processes), that would solve hybrid scheduling headaches.

So kind of the opposite approach of setting affinity, interesting but arguably yet another "expert level" tweak to make a system usable. (I would like it of course, but also being able to configure memory and IO cache too…)
If a program is problematic but not super sensitive to responsiveness, you can encapsulate it in a VM and still run it "seamless" on your desktop BTW.
But if you're running into these kinds of issues, you might as well just build a massive workstation with a mighty CPU and loads of RAM…

SMT is at least on Intel side inefficient, e-cores reign supreme over it.

Only to the extent that the more saturated the CPU core is, the less of an advantage SMT will achieve (like I explained in #39), while the stronger the CPU front-end is, the less it will be idling its computational resources. This is why we see a relatively larger gain from SMT since the first Zen CPUs up until today vs. their respective counterparts, it is not that AMD's SMT implementation is somehow better.

SMT vs. E-cores, that's not apples to apples. It totally depends on the workload, and hard to make an overall fair statement. SMT do however have one advantage; the switching is done on hardware level.
Either way, I would much rather have those transistors spent on giving 5% higher IPC or something.

But bartlett lake interests me, because as a gamer and someone who is still dominated by per core performance I will prefer that CPU for the forseeable future.

But explain to me this;
Even if we assume Intel carves out a consumer SKU from the 12-core Bartlett Lake, and assuming there are no architectural benefits here, in best-case it will perform slightly lower than Raptor Lake in gaming (as it's the absurdly high boost that gives Raptor Lake an edge in gaming), and somewhat better in various productive workloads of course. But a 12-core model will almost certainly have lower base clock, and effective clocks at "mixed" threaded workloads, and presumably slightly less aggressive boosting than the "problematic" Raptor Lake. So what are you really gaining here? And how is this actually better than an 285K?

 

[Visible Noise]

Some applications (especially browsers) and background processes tend to spawn more threads the more "cores" a CPU have

Outside of encoders? I’ll need some evidence of that. I just looked at my work laptop, I have 375 processes and 5,234 threads running. You think those numbers will be significantly affected by the number of cores, to the point that fewer cores would be better?

That‘s a pretty extraordinary claim.

 

[MaMoo]

Highly doubt it. Are you turning HT off for your symmetric multiprocessing? Cause those are much slower than ecores.

The problem is not the threading mode, because I am referring to multiprocessing as opposed to multithreading. My jobs are divided into contiguous batches. Say I have 8 P and 8 E cores, then the slower cores will always finish last while the faster cores are idling after their tasks. The next contiguous batch for technical and memory reasons cannot be issued before the completion of the previous one. So the whole job moves at the speed of the slowest cores. HT actually helps because it helps to slow down the P cores to compensate. Still, benching using Linpack, I get a 11% boost in performance disabling e-cores on my 12700k, because the power budget moves to the P cores and their average clocks increase during boost. I don't OC for tasks needing perfect stability. This was observed by others, too. Only for multithreading and asymmetric or many many threads, would a hybrid arch perform better. We use same benchmarks and libraries that we developed internally on Xeons and Threadrippers, too.

 

[Sunny and 75]

Intel, now is the time to act! The 12 P-Core SKU, upholds our pact!

 

[Dr. Dro]

Will they reintroduce HT? since now it appears to target 1851 socket.

Previous rumours were pointing at socket 1700

Bartlett is supposed to run Raptor Cove architecture, it is not a new product so Hyper-Threading was never removed

What makes people think that a CPU lineup intended for "edge" and "networking equipment" (in other words embedded/light servers) will be targeting the gaming market?
The reason why Raptor Lake have a slight edge in high-FPS gaming over Arrow Lake is thanks to its absurd boosting, it would be very strange for this kind of lineup to do this.

Well, the LGA 1700 platform is established and this is a drop-in upgrade. 12 P-cores is meaty and tasty. My Z790 Apex Encore is ready, and by the time this ships, my wallet should be ready too (that dang 5090 hehe). Personally, I'd consider this a significant upgrade over the i9-13900KS at 8+16 if it maintains a similar clock curve target (5600MHz all core guaranteed), especially if removing the E-cores allows Intel to reenable AVX-512 support on the processor.

 

[efikkan]

Well, the LGA 1700 platform is established and this is a drop-in upgrade. 12 P-cores is meaty and tasty. My Z790 Apex Encore is ready, and by the time this ships, my wallet should be ready too (that dang 5090 hehe). Personally, I'd consider this a significant upgrade over the i9-13900KS at 8+16 if it maintains a similar clock curve target (5600MHz all core guaranteed)

Well, it's not going to go that high with 12 cores unless you feed it a significant amount of power. But if you're talking primarily about the OC market, then you might have a point. As I mentioned in #42, at stock this CPU will run at significantly lower clocks than Raptor Lake. I do wonder how easy it will be to cool when clocked at 5600 MHz all core…

especially if removing the E-cores allows Intel to reenable AVX-512 support on the processor.

Excellent point, Zen 5 has shown us how much of a difference this can make, and the uses of it is only going to increase.

 

[Dr. Dro]

Well, it's not going to go that high with 12 cores unless you feed it a significant amount of power. But if you're talking primarily about the OC market, then you might have a point. As I mentioned in #42, at stock this CPU will run at significantly lower clocks than Raptor Lake. I do wonder how easy it will be to cool when clocked at 5600 MHz all core…

Excellent point, Zen 5 has shown us how much of a difference this can make, and the uses of it is only going to increase.

Should be sameish? Since 4 E-cores use about the same die area of a single P-core, the footprint of both processors should be similar. It's not unheard of, the 14900KS does 5900MHz all core, boosting 1 of them to 6200 out of the box, so I suppose meeting the 13900KS/14900K targets shouldn't be too difficult for Intel.

AVX-512 would be the wild card, it's known to significantly increase power consumption on Intel designs, I suppose that would increase power consumption to a point it's untenable to have 5600 all core, but even 5200 would be solid if it held flat clocks throughout the entire workload IMHO

 

[efikkan]

Should be sameish? Since 4 E-cores use about the same die area of a single P-core, the footprint of both processors should be similar. It's not unheard of, the 14900KS does 5900MHz all core, boosting 1 of them to 6200 out of the box, so I suppose meeting the 13900KS/14900K targets shouldn't be too difficult for Intel.

The issue with cooling any of the top mainstream CPUs today during OC, especially Raptor Lake, is thermal density. Many excellent tower coolers will struggle*, as the heat pipes will stop working above a certain temperature, and water coolers will need the right contact plate and enough water flowing through it.
(* For comparison e.g. a single-towered Noctua NH-U14S caps out at ~680W on Xeon W, but may struggle to even hit 250W on a recent mainstream CPU.)

 

[Frizz]

I don't know about everyone else as a normal consumer, but since moving to AMD (not very long) I'm already tired of needing to use project lasso, checking if the cores are being utilized/parked properly and then needing specific settings such as ensuring the OS is on Balanced Power profile etc. etc.

Then on Intel Arrow Lake the overclocking experience was too weird, P-Core couldn't move an inch further from it's original P-Core clocks and overclocking the E-Core was what gave the biggest performance boosts....?

I am not saying any of the chips are bad as they are flagship after all however it would be nice not to worry about schedulers moving forward and just focus on raw performance of per say 10x Performant Cores either blue/red team that emphasize on general gaming/desktop performance, this I believe is the way forward and Arrow Lake hit the nail on the head with ditching HT and hopefully ditching E-Cores will yield better results than all hybrid solutions in gaming/general desktop performance.