4 ms·

>hybrid cores What is the benefit of this for a desktop processor? I understand for mobile devices that have to conserve battery, but Intel is also making a d

>hybrid cores What is the benefit of this for a desktop processor? I understand for mobile devices that have to conserve battery, but Intel is also making a desktop version with 8 of each core type. Most modern CPUs have the ability to turn off all but one core and throttle clock speeds way down when idle already.

rincebrain 5y ago
I would presume because it turns out, in many cases, you have some tasks continuously running that don't require that much CPU, but a nonzero amount, and this way you can shove them on the LITTLE cores while the big core(s) go to town on single big tasks.
adrian_b 5y ago
According to Intel's claims, the Gracemont cores have more than a half of the throughput for multi-threaded tasks of a big core in about a quarter of the area and a quarter of the power consumption. That means that a Gracemont thread is supposed to have about the same throughput as a SMT thread on a big core. Therefore an 8+8 cores Alder Lake with 24 hybrid threads should have about the same MT performance with a CPU having 12 big cores with 24 SMT threads, while having the area and the power consumption of only 10 big cores. So Alder Lake is expected to have a similar MT performance with the 12-core Ryzen 9 5900X, while having a much better ST performance. Of course that remains to be seen. Early next year AMD should counter with Zen 3+ CPUs having a much larger cache memory and probably a little higher clock frequency. Those will probably have much better MT performance than Alder Lake, but Alder Lake should retain the best ST performance until Zen 4, late next year.
aseipp 5y ago
It's not a matter of absolute numbers, but a question of die area, to some degree. You can either spend X amount of area on a "big core" that can clock all the way down to idle (where it will spend most time) and all the way up to its max perf, or you can spend a fraction of X on a smaller core with a much, much smaller relative surface area, and get the same performance at idle or near-idle tasks. It just won't be able to scale "up" to max perf. But the net gain is you've saved a significant amount of area, and that translates to better TDP and thermals. The goal is to reap this benefit while also having "big cores" for the throttle-up phase, so you need both. There's a lot of talk about "background tasks" that do light, but necessary processing in asymmetric multicore setups like this; for example on macOS Spotlight performs background indexing continuously, this is a great fit for a "small core" like in the M1. There's no amount of idle clocking you can do to match the perf-per-watt here. Similarly, there is also talk of "dark silicon" and "increasing heterogenous systems", like hybrid designs. But these are all the same problem, really: modern designs are not about having the most powerful components, but the correct blend of components to maximize the work-per-watt you can achieve. You can't just throw "more transistors" at every single design anymore and get free wins on all fronts 1 year later when transistors shrink.

rincebrain 5y ago

I would presume because it turns out, in many cases, you have some tasks continuously running that don't require that much CPU, but a nonzero amount, and this way you can shove them on the LITTLE cores while the big core(s) go to town on single big tasks.

adrian_b 5y ago

According to Intel's claims, the Gracemont cores have more than a half of the throughput for multi-threaded tasks of a big core in about a quarter of the area and a quarter of the power consumption. That means that a Gracemont thread is supposed to have about the same throughput as a SMT thread on a big core. Therefore an 8+8 cores Alder Lake with 24 hybrid threads should have about the same MT performance with a CPU having 12 big cores with 24 SMT threads, while having the area and the power consumption of only 10 big cores. So Alder Lake is expected to have a similar MT performance with the 12-core Ryzen 9 5900X, while having a much better ST performance. Of course that remains to be seen. Early next year AMD should counter with Zen 3+ CPUs having a much larger cache memory and probably a little higher clock frequency. Those will probably have much better MT performance than Alder Lake, but Alder Lake should retain the best ST performance until Zen 4, late next year.

aseipp 5y ago

It's not a matter of absolute numbers, but a question of die area, to some degree. You can either spend X amount of area on a "big core" that can clock all the way down to idle (where it will spend most time) and all the way up to its max perf, or you can spend a fraction of X on a smaller core with a much, much smaller relative surface area, and get the same performance at idle or near-idle tasks. It just won't be able to scale "up" to max perf. But the net gain is you've saved a significant amount of area, and that translates to better TDP and thermals. The goal is to reap this benefit while also having "big cores" for the throttle-up phase, so you need both. There's a lot of talk about "background tasks" that do light, but necessary processing in asymmetric multicore setups like this; for example on macOS Spotlight performs background indexing continuously, this is a great fit for a "small core" like in the M1. There's no amount of idle clocking you can do to match the perf-per-watt here. Similarly, there is also talk of "dark silicon" and "increasing heterogenous systems", like hybrid designs. But these are all the same problem, really: modern designs are not about having the most powerful components, but the correct blend of components to maximize the work-per-watt you can achieve. You can't just throw "more transistors" at every single design anymore and get free wins on all fronts 1 year later when transistors shrink.