3 ms·

Still, it's a "Core" core and "Atom" core together, leading to all kinds of unexpected consanguinities. As with M1 benchmarks, what really suffers is the 20%-80

by bitL 4y ago

Still, it's a "Core" core and "Atom" core together, leading to all kinds of unexpected consanguinities. As with M1 benchmarks, what really suffers is the 20%-80% load area, together with the lack of AVX-512 which 11th gen had that might be a deal breaker to some. Also, having 8 efficiency cores is just silly, if you just want to run background tasks with minimal power consumption, how do 8 cores lead to that?

jeffbee 4y ago
The only thing I can figure out to do on these hybrids is to constrain system junk to the e-cores using a systemd directive, and put all my user session processes on the p-cores in the same fashion. Also necessary to boot with interrupts on the p-cores only, unless you want your nvme and ethernet device performance to be terrible. I'm not sure the complexity is worth it.
kllrnohj 4y ago
There's still 8P cores, so no the 20-80% load area is not suffering. > Also, having 8 efficiency cores is just silly, if you just want to run background tasks with minimal power consumption, how do 8 cores lead to that? Who said efficient cores must only be for low power background tasks? These efficiency cores are small, very small, so Intel can pack a lot of them into the die. And they provide far far bigger returns on multithreaded workloads than SMT does. As in, given a choice between 2 more P cores (so 10P cores total) or these 8 E cores (so 8p+8e), it's a no brainer to go with the 8 E cores. You get larger multithreading returns with a similar die size cost. And the background power efficiency is just a freebie bonus.
- bitL 4y ago
  Ok, so given you can place 10 P-cores vs 40 E-cores on the same die area by your math, why don't they ship a 40 E-core CPU then? If those little cores are so capable in multi-threading (better than 2P anyway), that would be a big win, no?
  - kllrnohj 4y ago
    Yes exactly https://www.anandtech.com/show/17259/intel-discloses-multigeneration-xeon-scalable-roadmap-new-ecore-only-xeons-in-2024 https://www.anandtech.com/show/17259/intel-discloses-multige...

by bitL 4y ago

jeffbee 4y ago

The only thing I can figure out to do on these hybrids is to constrain system junk to the e-cores using a systemd directive, and put all my user session processes on the p-cores in the same fashion. Also necessary to boot with interrupts on the p-cores only, unless you want your nvme and ethernet device performance to be terrible. I'm not sure the complexity is worth it.

kllrnohj 4y ago

There's still 8P cores, so no the 20-80% load area is not suffering. > Also, having 8 efficiency cores is just silly, if you just want to run background tasks with minimal power consumption, how do 8 cores lead to that? Who said efficient cores must only be for low power background tasks? These efficiency cores are small, very small, so Intel can pack a lot of them into the die. And they provide far far bigger returns on multithreaded workloads than SMT does. As in, given a choice between 2 more P cores (so 10P cores total) or these 8 E cores (so 8p+8e), it's a no brainer to go with the 8 E cores. You get larger multithreading returns with a similar die size cost. And the background power efficiency is just a freebie bonus.

bitL 4y ago
Ok, so given you can place 10 P-cores vs 40 E-cores on the same die area by your math, why don't they ship a 40 E-core CPU then? If those little cores are so capable in multi-threading (better than 2P anyway), that would be a big win, no?
- kllrnohj 4y ago
  Yes exactly https://www.anandtech.com/show/17259/intel-discloses-multigeneration-xeon-scalable-roadmap-new-ecore-only-xeons-in-2024 https://www.anandtech.com/show/17259/intel-discloses-multige...