Congrats to Intel for getting their 18A node working well enough that they managed to manufacture one single sample chip (which may or may not be functional).

ymdhisCongrats to Intel for getting their 18A node working well enough that they managed to manufacture one single sample chip (which may or may not be functional).

Why would a user go for product/SoC with incomplete ISA across its 3types of cores
“4P+8E+4LP” who needs this hustle?

AMD provided a performance demo of Zen a little over 6 months before release.
www.techpowerup.com/225129/amd-demos-breakthrough-performance-of-the-zen-cpu-core

And of it playing a game at all 8 months before launch.
www.digitaltrends.com/computing/amd-zen-cpu-doom/

Similarly OEMs showed working Meteor Lake and later Lunar Lake computers months before those chips launched. Assuming a December launch date, Intel should share Panther Lake running by June.

It looks to me like it mixes the best of Lunar Lake in with 8 more cores on the ring bus. Also where Lunar Lake has 8 Xe2 cores and Arrow Lake uses Xe1 cores, this uses 12 Xe3 cores. It could be an extremely competitive mobile chip.

T0@stthe top-most PTL-H SKU could appear with a 4P+8E+4LP+12Xe3 configuration.

I can hear the Windows Scheduler choking already. 4 whole P-cores? Meanwhile, AMD is shipping up to 16 of those in mobile, and even Qualcomm is already giving you 12. If that's the high-end, then lesser SKUs are gonna be rough. It might ship out on 18A, but it doesn't look like it's going to be all that impressive compared to the competition. Looks like one of the main objectives is to abandon the Tile approach and get back to monolithic.

Darmok N JaladI can hear the Windows Scheduler choking already. 4 whole P-cores? Meanwhile, AMD is shipping up to 16 of those in mobile, and even Qualcomm is already giving you 12. If that's the high-end, then lesser SKUs are gonna be rough. It might ship out on 18A, but it doesn't look like it's going to be all that impressive compared to the competition. Looks like one of the main objectives is to abandon the Tile approach and get back to monolithic.

AMD has 16 cores in Strix Halo, but that's more of a high-TDP mobile processor. (It also has special concerns for the Windows scheduler: preferred cores and two CCXs. Windows will try to use the preferred cores and it will try to keep related threads on just one CCX.) Most AMD laptops will have something closer to Strix Point, which has 4 Zen 5 cores on one CCX and 8 Zen 5c cores on another CCX. 4+8 cores in Strix Point or 4+8 cores in Panther Lake; what's the difference? (Panther Lake's 4+8 cores are easier for the scheduler because they share an L3 cache and there are no virtual cores to schedule. But Panther Lake does have 4 LPE cores outside the L3 cache which is a challenge Strix Point doesn't present.)

SquaredAMD has 16 cores in Strix Halo, but that's more of a high-TDP mobile processor. (It also has special concerns for the Windows scheduler: preferred cores and two CCXs. Windows will try to use the preferred cores and it will try to keep related threads on just one CCX.) Most AMD laptops will have something closer to Strix Point, which has 4 Zen 5 cores on one CCX and 8 Zen 5c cores on another CCX. 4+8 cores in Strix Point or 4+8 cores in Panther Lake; what's the difference? (Panther Lake's 4+8 cores are easier for the scheduler because they share an L3 cache and there are no virtual cores to schedule. But Panther Lake does have 4 LPE cores outside the L3 cache which is a challenge Strix Point doesn't present.)

I think that, unless the E cores in Panther Lake are far closer in performance to the E cores, the AMD CPU still has an advantage. When a thread ends up on the wrong core, the performance penalty of having it land on an Zen C core will be considerably less than if it hits a Panther E core. Also, I see it on my work PC regularly, where Windows 11 ignores the E cores entirely in threaded workloads, and then it isn't long before the P cores throttle down under the workload. With Zen C, at least it's a full CPU core with reduced cache, and not a different design with potentially reduced features.

Darmok N JaladAlso, I see it on my work PC regularly, where Windows 11 ignores the E cores entirely in threaded workloads, and then it isn't long before the P cores throttle down under the workload.

Just as a side note, that also happens quite often with Apple's AS (bar the throttling part). My M3 Max whenever stressed more often than not does not make use of the E-cores, only the P ones.

igormpJust as a side note, that also happens quite often with Apple's AS (bar the throttling part). My M3 Max whenever stressed more often than not does not make use of the E-cores, only the P ones.

Yeah, I’ve never been totally convinced of both core types. I think it’s especially bad when the core configuration favors lots of E cores over P cores. Fortunately Apple adds more P cores to their higher-tier products, where Intel ramps up the E cores.

Darmok N JaladI think that, unless the E cores in Panther Lake are far closer in performance to the E cores, the AMD CPU still has an advantage. When a thread ends up on the wrong core, the performance penalty of having it land on an Zen C core will be considerably less than if it hits a Panther E core. Also, I see it on my work PC regularly, where Windows 11 ignores the E cores entirely in threaded workloads, and then it isn't long before the P cores throttle down under the workload. With Zen C, at least it's a full CPU core with reduced cache, and not a different design with potentially reduced features.

The gap between P and E cores shrank going from Alder/Raptor Lake to Meteor Lake and again from Meteor Lake to Arrow Lake. Intel claims Skymont has 2% higher IPC than Raptor Cove and Lion Cove 14% over Redwood Cove. Reviewers more or less agreed and Redwood Cove has roughly the same IPC as Raptor Cove, which puts Lion Cove IPC roughly 12% ahead of Skymont. Now Zen 5c has almost the same IPC as Zen 5 so it does have a small lead before I multiply clock frequency.

The Core 285H Lion Cove/Skymont max frequency is 5.4/4.5 GHz and base is 2.9/2.7 GHz which if given a 12%/0% increase becomes 6.048/4.5 arbitrary units max and 3.248/2.7 arbitrary units which makes the E cores 26% and 17% slower max and base.

The Ryzen HX 370 Zen 5/Zen 5c max frequency is 5.1/3.3 GHz and base is 2.0/2.0 GHz. If we assume about equal IPC that makes the cloud cores 35% and 0% slower max and base.

I imagine a mobile chip running a many-core load will probably operate close to its base frequency and get closer to that 17% and 0% performance loss on the little core. That's probably a small edge for AMD. But we haven't talked at all about LPE cores which have no L3 cache and on Arrow Lake might even be older Crestmont cores. Tasks scheduled on those will be a lot slower. On Panther Lake those will at least be the same microarchitecture as the E cores but it'll still be a hit to performance.

Lastly you mentioned the threads sticking to only P cores or only E cores. I don't understand that and I don't have a mixed Intel processor to test but that is how Intel and reviewers said the Windows scheduler is intended to work on Alder Lake at least. I don't much like that idea myself.

SquaredSimilarly OEMs showed working Meteor Lake and later Lunar Lake computers months before those chips launched. Assuming a December launch date, Intel should share Panther Lake running by June.

I forgot: Intel demoed Panther Lake running at all 2 months ago at CES.