I love these videos. They’re very insightful and informative. And it shows that there are talented and very smart people at Intel. I’m glad Intel is coming back (after the 10nm saga), and now with a laser focus on power efficiency. x86 lives on! My first x86 was a 286!! The last Intel chip with a Sky in its name became historic. sky lake. Skymont sounds beautiful and amazing. I wonder what’s in store with Darkmont ?
If Intel manages to achieve Snapdragon levels of power efficiency while maintaining comparative performance in Lunar Lake, it is going to seriously kneecap Qualcomm because why deal with all the compatibility BS? On the other hand Microsoft and laptop oems would love to have another option other than AMD or Intel because that competition also benefits them
@@RobertDunn310 indeed. Raptor cove levels of performance at much less power is a very attractive proposition for low power machines like thin and lights, mini PCs, handheld gaming consoles and so forth. And the cool thing with lunar lake is that it also has great graphics with battlemage which should make it a more potent solution than the snapdragon. And I agree with you wholeheartedly that one benefit Intel has going for it is native x86 support, no compatibility layer. If Intel produces (and consistently so) more powerful and efficient chips than Apple Silicon, I wonder if Apple might consider coming back ?!
thank you so much! We will keep these types of videos coming and if you folks would like for us to change something or have other type of content please feel free to drop it here in the comments
Thanks for the videos, it is a great learning resource. It is amazing how much die area the Skymont Core takes up, the Performance per Area is really good. At the same time we can see how memory is not scaling as well as logic. The System Side Cache is only 8MB but takes up a lot of die area. The die area taken up by memory from all the caches combined is huge.
Thread choices are a lot easier with LNL at least. Rather than go P-core main, E-core, P-core hyper, lpE-core, from the top down, it just goes P-core, lpE-core. Start on one, go to the other for either more threads (P-core down) or faster threads (E-core up).
Between Thread Director and W11 the scheduling is already pretty great. However it's going to be near perfect with Lunar Lake, as scheduling as Thread Director is getting enhanced and more importantly E-Cores are getting a huge performance boost, no longer will e-cores be weaker backup cores for multi-thread like they previously were, but now they are more like medium cores with Raptor Cove IPC but vastly more efficient and space efficient.
Many thanks to Alejandro again for his questions and to Stephen for chatting with us! 👏😁 It was really interesting the question about the new front-end and Stephen's answer about how they have augmented it, employing three-decoder clusters with very little transistor penalty. It had always been said that, in the x86 architecture, it was unfeasible to scale the front-end to increase the instructions per cycle, but Stephen and his team have shown that it is possible through ingenious and efficient solutions. So, another x86 myth busted! Bravo team! 👏😁 Skymont will be deterministic for multi-threaded tasks scaling several E-core cores given their small footprint and optimized consumption. Cloud server processors come to mind. Also in the consumer market, assuming more prominence E-core for everyday tasks as Stephen says, which will result in lower consumption. It also opens the door to simplify the design by eliminating HyperThreading. Looking forward to the next iterations! Cheers team! 💪😁
Have Intel considered to replace the ring with a full mesh since there are only 5 clients (4P+1E-Cluster)? In my super sensitive gaming scenario I'm bothered by noisy ring issue - 13900K any busy P-core will make other cores "noisy". Also now E-core gets P-core level IPC, then what's the next main difference between them?
P-cores still have a 14-18% IPC gain over Raptor Cove as well. The main differentiator now is in AVX support and clock speed. Skymont is likely a sub-4ghz architecture for density, while Lion Cove should go well above 5ghz and likely gets larger caches.
@@DigitalJedi Good points. One thing I found during my gaming is that Intel really masters the ring architecture. Raptor Lake ring has 12 clients, while Comet Lake has shown scalability issues with 10 or more clients on the ring. What Raptor Lake does is it runs 8 above the ring clock and 4 below and it's somewhat stable. I'd imagine in Comet Lake running 2 or 4 cores below the ring clock would also make it more stable but I don't have one to test. The ring clock limits both P- and E-core clocks because when the clock difference between core and ring is too large there is a huge performance penalty, and if it's too small then it's very unstable. Ideal clock is something like P=Ring+6, E=Ring-5, and that makes all the clocks fixed.
Hey Kevin don't try spend too much money on trying to look as if you are interested in making a good impression. We like the slovenly appearance to give us the impression that you are spending all of you time trying to improve Intel products.
That's the entire point of Lunar Lake. It's aggressively trimmed down on the P-cores and can shut off parts of the die to save power. They've shown perf/watt and system power comparisons to Meteor Lake in other presentations already. 9W to 32W TDP depending on the SKU, but the high end is mostly due to OEMs wanting to wring more performance out of it. It's a 15W class chip that will perform like the current 28W units.
I love these videos. They’re very insightful and informative. And it shows that there are talented and very smart people at Intel. I’m glad Intel is coming back (after the 10nm saga), and now with a laser focus on power efficiency. x86 lives on! My first x86 was a 286!!
The last Intel chip with a Sky in its name became historic. sky lake. Skymont sounds beautiful and amazing. I wonder what’s in store with Darkmont ?
If Intel manages to achieve Snapdragon levels of power efficiency while maintaining comparative performance in Lunar Lake, it is going to seriously kneecap Qualcomm because why deal with all the compatibility BS?
On the other hand Microsoft and laptop oems would love to have another option other than AMD or Intel because that competition also benefits them
@@RobertDunn310 indeed. Raptor cove levels of performance at much less power is a very attractive proposition for low power machines like thin and lights, mini PCs, handheld gaming consoles and so forth. And the cool thing with lunar lake is that it also has great graphics with battlemage which should make it a more potent solution than the snapdragon.
And I agree with you wholeheartedly that one benefit Intel has going for it is native x86 support, no compatibility layer.
If Intel produces (and consistently so) more powerful and efficient chips than Apple Silicon, I wonder if Apple might consider coming back ?!
im sure they know better regarding that task
thank you so much! We will keep these types of videos coming and if you folks would like for us to change something or have other type of content please feel free to drop it here in the comments
The new E-Core sounds very powerful, looking forward to testing it when it comes to desktop.
Thanks Alejandro and team
Correction: 13:36 says IPC of Skylake but means Skymont here
Thanks for the videos, it is a great learning resource.
It is amazing how much die area the Skymont Core takes up, the Performance per Area is really good. At the same time we can see how memory is not scaling as well as logic. The System Side Cache is only 8MB but takes up a lot of die area. The die area taken up by memory from all the caches combined is huge.
i thought that was a really well conducted interview. really enjoyed it. cheers
Brother is dancing around ArrowLake, its announced so you can say it 😁
Looking forward to see those Skymont E-cores in action in Arrow Lake CPUs.
Skymont Look's good, I already want to see it in action. 🤤🤤😎
> First hybrid...Alder Lake
Lakefield: 👀
Let's hope for good scheduler for operating system kernel.
I have to hope that the software side rises to meet that challenge.
Thread choices are a lot easier with LNL at least. Rather than go P-core main, E-core, P-core hyper, lpE-core, from the top down, it just goes P-core, lpE-core. Start on one, go to the other for either more threads (P-core down) or faster threads (E-core up).
@@DigitalJedioh yeahh good point
Linuxxx
Between Thread Director and W11 the scheduling is already pretty great. However it's going to be near perfect with Lunar Lake, as scheduling as Thread Director is getting enhanced and more importantly E-Cores are getting a huge performance boost, no longer will e-cores be weaker backup cores for multi-thread like they previously were, but now they are more like medium cores with Raptor Cove IPC but vastly more efficient and space efficient.
Many thanks to Alejandro again for his questions and to Stephen for chatting with us! 👏😁
It was really interesting the question about the new front-end and Stephen's answer about how they have augmented it, employing three-decoder clusters with very little transistor penalty. It had always been said that, in the x86 architecture, it was unfeasible to scale the front-end to increase the instructions per cycle, but Stephen and his team have shown that it is possible through ingenious and efficient solutions. So, another x86 myth busted! Bravo team! 👏😁
Skymont will be deterministic for multi-threaded tasks scaling several E-core cores given their small footprint and optimized consumption. Cloud server processors come to mind.
Also in the consumer market, assuming more prominence E-core for everyday tasks as Stephen says, which will result in lower consumption.
It also opens the door to simplify the design by eliminating HyperThreading.
Looking forward to the next iterations! Cheers team! 💪😁
Thanks Steve!
A little less conversation
Have Intel considered to replace the ring with a full mesh since there are only 5 clients (4P+1E-Cluster)? In my super sensitive gaming scenario I'm bothered by noisy ring issue - 13900K any busy P-core will make other cores "noisy".
Also now E-core gets P-core level IPC, then what's the next main difference between them?
P cores got even better
its true
P-cores still have a 14-18% IPC gain over Raptor Cove as well. The main differentiator now is in AVX support and clock speed. Skymont is likely a sub-4ghz architecture for density, while Lion Cove should go well above 5ghz and likely gets larger caches.
@@DigitalJedi Good points. One thing I found during my gaming is that Intel really masters the ring architecture. Raptor Lake ring has 12 clients, while Comet Lake has shown scalability issues with 10 or more clients on the ring. What Raptor Lake does is it runs 8 above the ring clock and 4 below and it's somewhat stable. I'd imagine in Comet Lake running 2 or 4 cores below the ring clock would also make it more stable but I don't have one to test. The ring clock limits both P- and E-core clocks because when the clock difference between core and ring is too large there is a huge performance penalty, and if it's too small then it's very unstable. Ideal clock is something like P=Ring+6, E=Ring-5, and that makes all the clocks fixed.
Qualcomm Lake
Hey Kevin don't try spend too much money on trying to look as if you are interested in making a good impression. We like the slovenly appearance to give us the impression that you are spending all of you time trying to improve Intel products.
That’s great, now make it so the battery can last more than three minutes
That's the entire point of Lunar Lake. It's aggressively trimmed down on the P-cores and can shut off parts of the die to save power. They've shown perf/watt and system power comparisons to Meteor Lake in other presentations already. 9W to 32W TDP depending on the SKU, but the high end is mostly due to OEMs wanting to wring more performance out of it. It's a 15W class chip that will perform like the current 28W units.