Intel "Arrow Lake-S" to See a Rearrangement of P-cores and E-cores Along the Ringbus

[btarunr]

Intel's first three generations of client processors implementing hybrid CPU cores, namely "Alder Lake," "Raptor Lake," and "Meteor Lake," have them arranged along a ringbus, sharing an L3 cache. This usually sees the larger P-cores to one region of the die, and the E-core clusters to the other region. From the perspective of the bidirectional ringbus, the ring-stops would follow the order: one half of the P-cores, one half of the E-core clusters, iGPU, the other half of E-cores, the other half of the P-cores, and the Uncore, as shown in the "Raptor Lake" die-shot, below. Intel plans to rearrange the P-cores and E-core clusters in "Arrow Lake-S."

With "Arrow Lake," Intel plans to disperse the E-core clusters between the P-cores. This would see a P-core followed by an E-core cluster, followed by two P-cores, and then another E-core cluster, then a lone P-core, and a repeat of this pattern. Kepler_L2 illustrated what "Raptor Lake" would have looked like, had Intel applied this arrangement on it. Dispersing the E-core clusters among the P-cores has two possible advantages. For one, the average latency between a P-core ring-stop and an E-core cluster ring-stop would reduce; and secondly, there will also be certain thermal advantages, particularly when gaming, as it reduces the concentration of heat in a region of the die.



Every P-core would be no more than one ring-stop away from an E-core cluster, which should benefit migration of threads between the two core types. Thread Director prefers E-cores, and when a workload overwhelms an E-core, it is graduated to a P-core. This E-core to P-core migration should see reduced latencies under the new arrangement.

View at TechPowerUp Main Site | Source

 

[Daven]

Are all P cores and E cores on one tile in Arrow Lake or are E cores and P cores on separate tiles?

 

[btarunr]

Are all P cores and E cores on one tile in Arrow Lake or are E cores and P cores on separate tiles?

From what I gather, all cores are on the same die.

 

[dgianstefani]

From what I gather, all cores are on the same die.

Should be nice for latency compared to Zen, fingers crossed, considering it's single chiplet for CPU cores and also on Foveros.

 

[rv8000]

and secondly, there will also be certain thermal advantages, particularly when gaming, as it reduces the concentration of heat in a region of the die.

How does this provide thermal advantages? Localizing p-cores into a smaller area creates more of a hot spot/heat dense area. Less immediate surface area when transferring to the integrated heat spreader, on top of a denser process node.

 

[piloponth]

Intel using this ringbus thing since Nehalem. What a shame on their R&D not to bring any lowlatency, lowenergy & low silicon budget solution till this day.

 

[P4-630]

It doesn't seem intel rushed out this time around, they have clearly better thought about things.

 

[dgianstefani]

Intel using this ringbus thing since Nehalem. What a shame on their R&D not to bring any lowlatency, lowenergy & low silicon budget solution till this day.

Compared to what?

 

[P4-630]

fingers crossed

Right!

 

[JustBenching]

Was thinking about skipping arrow lake but it seems intel is going to pull another wow moment like alderlake.

 

[evernessince]

Should be nice for latency compared to Zen, fingers crossed, considering it's single chiplet for CPU cores and also on Foveros.

Intel's chiplets being physically closer together should also help latency. I'm not sure how the two cores on Intel's low power island will interact with the whole setup but I assume they will only be in use in non-demanding scenarios and thus won't impact any demanding app latency.

Looks pretty promising.

 

[Daven]

Was thingink about skipping arrow lake but it seems intel is going to pull another wow moment like alderlake.

Everything points to this being another architecture realignment since Alder Lake through Raptor Lake refresh hit another dead end with power budgets similar to how Intel hit a dead end with Comet Lake and had to scale back with Rocket Lake.

Lower clocks, no hyperthreading, etc should help with performance per watt but I’m not holding my breath on absolute performance being higher than a 13900k.

 

[Crazybc]

I am hoping that Arrow Lake desktop works out well .I so badly want to upgrade this year my system is 10 years old and I am hoping to go Intel again but if it turns out to be a dud I'll look at AMD. I like the idea of spreading out the cores for decreased latency and not centralizing all that heat . I will be watching the forums closely and I'm waiting for sept for way more information on the upcoming Z890 platform and Arrow lake desktop.

 

[JustBenching]

Everything points to this being another architecture realignment since Alder Lake through Raptor Lake refresh hit another dead end with power budgets similar to how Intel hit a dead end with Comet Lake and had to scale back with Rocket Lake.

Lower clocks, no hyperthreading, etc should help with performance per watt but I’m not holding my breath on absolute performance being higher than a 13900k.

Performance per watt was already a strong point for intel. That's the only thing they really don't need to improve.

 

[usiname]

Performance per watt was already a strong point for intel. That's the only thing they really don't need to improve.

 

[TristanX]

Intel presented ARL compute tile glimpse in their materials, and arrangement is unchanged. Ring bus is necessary, because cores must synchronize cache changes to ensure data integrity, so it is pare of cache system, not cores itself.

 

[Daven]

Performance per watt was already a strong point for intel. That's the only thing they really don't need to improve.

 

[Denver]

If the L3 cache is shared between all cores: 1.5MB cache/core... less than Lunar Lake.

 

[JustBenching]

View attachment 353568

View attachment 353569

Those are not ISO power lads. Let's not go over this again. If you care about efficiency you don't run 4096 power limits, 512 amp limits etc.

The point is on the vast majority of segments Intel has a much more efficient CPU at iso power in both ST and MT workloads. It's really not even competitive. R5 vs i5, R7 vs i7 etc.

Take your 7600x, it consumes 100w in CBR23 according to the same tests. Is there any Intel CPU that isn't actually faster than your 7600x while consuming the same watts? The answer is, maybe an i3 or something. Everything else handily beats the 7600x - in both performance and efficiency.

 

[Daven]

Those are not ISO power lads. Let's not go over this again. If you care about efficiency you don't run 4096 power limits, 512 amp limits etc.

The point is on the vast majority of segments Intel has a much more efficient CPU at iso power in both ST and MT workloads. It's really not even competitive. R5 vs i5, R7 vs i7 etc.

Look we can have a discussion about tech but we cannot have a discussion about your favorite company. Everyone here knows about the power draw and real world use of these products. It’s why close to 80% of TPU readers buy AMD and Nvidia for their builds.

We all see the same numbers and 99.9999% of us know Intel’s product are hugely inefficient. So don’t forgive Intel by ignoring their faults, wish for them to do better.

 

[TumbleGeorge]

another wow moment like alderlake.

I seem to have missed this. Was there really such a moment? I mean positive.

 

[dgianstefani]

I seem to have missed this. Was there really such a moment? I mean positive.

12900K

12900KS

13900K

So in answer to your question, yes.

Overall the 12-13th gen were massive improvements in IPC from the previous Skylake based systems, are still faster core for core against Zen, and didn't have the core count/process issues of Rocket Lake.

 

[JustBenching]

Look we can have a discussion about tech but we cannot have a discussion about your favorite company. Everyone here knows about the power draw and real world use of these products. It’s why close to 80% of TPU readers buy AMD and Nvidia for their builds.

We all see the same numbers and 99.9999% of us know Intel’s product are hugely inefficient. So don’t forgive Intel by ignoring their faults, wish for them to do better.

I don't, my favorite company is the one that makes the most efficient products. Currently it's Intel - it doesn't idle at 30-40 watts and it's incredibly fast in MT at low power. Let's say a 13700k / 14700k vs a 7800x 3d, both set at the same 90w, the i7 parts will literally fly past the 7800x 3d in MT workloads, being both faster and more efficient. Same goes for the i5 13600k vs 7600x and the 7700x.

If you strictly care about out of the box efficiency then TPU has tested that too, here is a simulated T and non k intel chip. Nothing is nowhere near in efficiency.

I seem to have missed this. Was there really such a moment? I mean positive.

Alderlake was a 40% jump in ST performance from cometlake within 15 months. It was also a huge uplift in MT performance, again during the same timespan. If you weren't impressed with that, I don't know man.

 

[Daven]

I don't, my favorite company is the one that makes the most efficient products. Currently it's Intel - it doesn't idle at 30-40 watts and it's incredibly fast in MT at low power. Let's say a 13700k / 14700k vs a 7800x 3d, both set at the same 90w, the i7 parts will literally fly past the 7800x 3d in MT workloads, being both faster and more efficient. Same goes for the i5 13600k vs 7600x and the 7700x.

If you strictly care about out of the box efficiency then TPU has tested that too, here is a simulated T and non k intel chip. Nothing is nowhere near in efficiency.

Alderlake was a 40% jump in ST performance from cometlake within 15 months. It was also a huge uplift in MT performance, again during the same timespan. If you weren't impressed with that, I don't know man.

This is from almost a year ago. It's probably close to 80% of AMD users on TPU now. It will probably hit 90% by next year. We DIYers like to buy AMD because of the efficiency and I don't see AMD letting go of this advantage anytime soon. How you see the exact opposite of what we see is mind boggling.

 

[TumbleGeorge]

Alderlake was a 40% jump in ST performance from cometlake within 15 months

Yes, but Intel did a bad job with the pre-Alder generations, I consider this to be back on the train tracks where it should have been the norm. That's why it's not a wow effect for me.