It looks like Brian said Samsung was switching to the third option, HMP (global task switching in the visual), for their new SoC. Which would mean it could actually use all 8 cores at once. Wonder how power consumption would be during this, and if there'd ever be a use for it. I'm guessing it wouldn't be useful at all really. Seems like the most useful scenario would be running the A7 cores for normal lifting and powering up 1-2 A15 cores when you need the performance.
Not sure 4+4 HMP is really needed on a phone (more like 2+2 or 2+4), but from a general computational point of view this could be really useful: delegate threads of lower priority to the small cores if the big guys are busy, or tasks where you know they'll be finished soon, so that waking up a big one is inefficient. Intelligent task scheduling does not exactly become easier, though.
Looks like an HMP scheduler for big.LITTLE was added to the kernel for 3.10 via the Linaro Organization (funded by ARM and SoC manufacturers). Seems like it would be a beast of a problem to get right but the kernel is certainly the most logical place to put that decision making logic into.
It will be fantastic. I don't know how the heuristics work for the scheduler (how the scheduler makes its descisions will make or break these early efforts) but it should end up using mostly a7s with only bursty a15s. The exception will be games, of course. The battery life, ideally, will be better. Now that we've got a heterogenous aware scheduler we can look towards future designs where you have more A7s than A15s (or A53 and A57). That is probably the balance that would make most economic sense, but when sammy first released their devices the best we could do was use the cpufreq driver to migrate within PAIRS of A7/A15.
This should've been obvious. ARM is making its Mali GPU's very integrated with its Cortex CPU's, and it's only going to do more of that in the future. If you're looking for heterogeneous computing, you're not going to find it in a 3rd party CPU+Imagination GPU. You can expect it from Nvidia and Qualcomm, too, though, because they will make their own CPU's and GPU's.
You said final model is coming by end of Quarter 4, suppose I buy Samsung Galaxy Note 3 in first lot, then the final model of the Exynos 5420 can be updated on Note 3 via software update?
This is what I want to ask too. So early in its life, note 3 and the new note 10.1 won't have this, and it will be enabled by software? Or, god forbid, the next revision will have this, hardware-wise, anf the early adopters are screwed?
It shouldn't require any other than big.LITTLE. Even the CCI is intended more for the older, hypervisor driven core migration (but, I'd imagine, it would come in use when migrating tasks, but with since we are doing aSMP I'd expect you'd see less of that). With this the new scheduler the system should be seen as a sort of weird 8 core cpu mixed with a cluster.
Seems like you'd have either low performance required scenarios that need only a few low performance cores or high performance scenarios where you need lots of high performance ones.
Why would you ever want lots of low performance cores and few high performance cores?
For lower power/heat and smaller die(smaller is cheaper). If you looked at the videos you don't have all 4 A15 at peak all that much and with only 2 A15 you can clock them higher. Anyway, i wasn't hoping for such SoC, it's a known fact we'll get some, just curious about perf and power since it might be a very solid combination.
I thought Exynos 5420 was coming to Note 3 Intl version. End of Q4 is December and Note 3 is releasing this Sept 25. I suddenly don't want to buy Note 3 given that I want latest cpu. It is bad in the part of Samsung to release this news just at the verge of releasing flagship phone.
I'm confused too. The way I understood it from reading GSMarena is that GS4 has Exynos 5410 with Cluster Migration, Note3 has Exynos 5420 with CPU Migration and that another Exynos is coming later this year (possibly 5430) with Global Task Scheduling. -
I don't understand the need for HMP in a phone, or even a tablet. Core migration makes sense (the OS only sees 4 CPU cores, the SoC switches between the A7 or A15 as needed on a per-Core basis). IOW, the system would be using one of the following setups, depending on the load: - 4x A7 + 0x A15 - 3x A7 + 1x A15 - 2x A7 + 2x A15 - 1x A7 + 3x A15 - 0x A7 + 4x A15
That setup makes sense to me, and can be done all in hardware. HMP doesn't; sounds like a lot of extra complexity and scheduler "magic" to make it work, and lots of ways to make it not work.
The goal is to allow more flexible configurations, like 4xA7+2xA15 or the next Tegra with 4xA15 and 2xA7 for power reduction (instead of 5xA15).
This should allow for both better instant power and lower idle/low usage power, while keeping area under control (A7 is MUCH smaller than A15). Very powerful for the future.
I hope we will see more of those and I'm sure we'll find great new combinations for the best overall experience.
A SoC with 4xA7 + 2xA15 would be interesting. Or, the opposite of the Tegra setup: 4xA7 (main cluster) + 1x A15 (for when you actually need the raw speed). :)
In those situations, yeah, the HMP makes sense. But in a situation where the big and LITTLE sides have the same number of cores, CPU Migration makes more sense (IMO).
No, Exynos 5410 had a hardware bug limiting it to cluster migration (CCI-400 problem). The Exynos 5420 fixes this problem, however there was no software to take fully advantage of it, so the hardware would manage the migration and the scheduler would only see 4 virtual CPUs (each a pair of CPU, A7+A15). Linaro worked on the Linux Kernel scheduler for Heterogeneous Multi-Processing, which allows to show all 8 cores to the scheduler and let it decide what to run where (while caring for power/temp limits). This is a software patch, that will be distributed around the end of the year. There will be no need for newer hardware.
Hey ANAND admins, how hard would be to block this spam, considering it always begins with "Live my job" and always has the $82h and sitting in front of the new imac - I mean come on, what is the excuse to keep tolerating this? Like it is not bad enough there is no edit button... come on, what are you people being paid for?
"Hey ANAND admins, how hard would be to block this spam"
They would probably have to put in one of those irritating pictured phrases with jumbled letters to confirm you are human. The cure is worse than the disease LOL.
Generally speaking, any of us editors (Brian, me, Anand, Ryan, etc.) can mark something as spam and delete it, and we do it as soon as we see it. Of course, posting a reply to said spam ends up with your posts getting pushed to the bottom of the comment chain as a response to nothing. Considering this article has only been live for less than 12 hours and the spam is gone, I'd say we're doing okay. Just ignore the spammers and we'll take care of them within a few hours, or if you must you can email [email protected] and tell them which article has a spam comment.
And as a final note, if we mark a comment as spam, it will delete all comments everywhere by that user, so if you value your account and comments please don't spam.
Just to be clear, this is not a "Samsung Innovation" but a project that has been coordinated, and fulfilled by Linaro, and others. I don't expect Samsung to play this any other way than "Look what WE'VE done!" but it'd be nice if it wasn't portrayed as something Samsung did since all you need to support the work is to have a compliant big.LITTLE system and run the new linux scheduler.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
30 Comments
Back to Article
akmittal - Wednesday, September 11, 2013 - link
Now we can call This SoC true Octa core, Snapdragon 800 have competition now.MarcusMo - Wednesday, September 11, 2013 - link
I don't think it's correct to say this is true octa core since afaik only four cores can be active at once.garadante - Wednesday, September 11, 2013 - link
It looks like Brian said Samsung was switching to the third option, HMP (global task switching in the visual), for their new SoC. Which would mean it could actually use all 8 cores at once. Wonder how power consumption would be during this, and if there'd ever be a use for it. I'm guessing it wouldn't be useful at all really. Seems like the most useful scenario would be running the A7 cores for normal lifting and powering up 1-2 A15 cores when you need the performance.MrSpadge - Wednesday, September 11, 2013 - link
Not sure 4+4 HMP is really needed on a phone (more like 2+2 or 2+4), but from a general computational point of view this could be really useful: delegate threads of lower priority to the small cores if the big guys are busy, or tasks where you know they'll be finished soon, so that waking up a big one is inefficient. Intelligent task scheduling does not exactly become easier, though.ivanatpr - Friday, September 13, 2013 - link
Looks like an HMP scheduler for big.LITTLE was added to the kernel for 3.10 via the Linaro Organization (funded by ARM and SoC manufacturers). Seems like it would be a beast of a problem to get right but the kernel is certainly the most logical place to put that decision making logic into.tuxRoller - Wednesday, September 11, 2013 - link
It will be fantastic.I don't know how the heuristics work for the scheduler (how the scheduler makes its descisions will make or break these early efforts) but it should end up using mostly a7s with only bursty a15s. The exception will be games, of course.
The battery life, ideally, will be better.
Now that we've got a heterogenous aware scheduler we can look towards future designs where you have more A7s than A15s (or A53 and A57). That is probably the balance that would make most economic sense, but when sammy first released their devices the best we could do was use the cpufreq driver to migrate within PAIRS of A7/A15.
krazyfrog - Wednesday, September 11, 2013 - link
You should read the article and/or watch the videos, then. All eight cores are now working simultaneously.mjv.theory - Wednesday, September 11, 2013 - link
Watch the third video, all 8 cores can be active at the same timeKrysto - Wednesday, September 11, 2013 - link
This should've been obvious. ARM is making its Mali GPU's very integrated with its Cortex CPU's, and it's only going to do more of that in the future. If you're looking for heterogeneous computing, you're not going to find it in a 3rd party CPU+Imagination GPU. You can expect it from Nvidia and Qualcomm, too, though, because they will make their own CPU's and GPU's.SKRG - Wednesday, September 11, 2013 - link
You said final model is coming by end of Quarter 4, suppose I buy Samsung Galaxy Note 3 in first lot, then the final model of the Exynos 5420 can be updated on Note 3 via software update?Kenan_sadhu - Wednesday, September 11, 2013 - link
This is what I want to ask too. So early in its life, note 3 and the new note 10.1 won't have this, and it will be enabled by software? Or, god forbid, the next revision will have this, hardware-wise, anf the early adopters are screwed?tuxRoller - Wednesday, September 11, 2013 - link
It shouldn't require any other than big.LITTLE. Even the CCI is intended more for the older, hypervisor driven core migration (but, I'd imagine, it would come in use when migrating tasks, but with since we are doing aSMP I'd expect you'd see less of that). With this the new scheduler the system should be seen as a sort of weird 8 core cpu mixed with a cluster.jjj - Wednesday, September 11, 2013 - link
Makes me curious about how 2xA15 + 4xA7 would do.mmrezaie - Wednesday, September 11, 2013 - link
I was thinking about the same thing, but I suppose Snapdragon's solution is better when you can just play with the voltage on just four cores.HisDivineOrder - Wednesday, September 11, 2013 - link
Seems like you'd have either low performance required scenarios that need only a few low performance cores or high performance scenarios where you need lots of high performance ones.Why would you ever want lots of low performance cores and few high performance cores?
jjj - Wednesday, September 11, 2013 - link
For lower power/heat and smaller die(smaller is cheaper). If you looked at the videos you don't have all 4 A15 at peak all that much and with only 2 A15 you can clock them higher.Anyway, i wasn't hoping for such SoC, it's a known fact we'll get some, just curious about perf and power since it might be a very solid combination.
willstay - Wednesday, September 11, 2013 - link
I thought Exynos 5420 was coming to Note 3 Intl version. End of Q4 is December and Note 3 is releasing this Sept 25. I suddenly don't want to buy Note 3 given that I want latest cpu. It is bad in the part of Samsung to release this news just at the verge of releasing flagship phone.boris81 - Wednesday, September 11, 2013 - link
I'm confused too.The way I understood it from reading GSMarena is that GS4 has Exynos 5410 with Cluster Migration, Note3 has Exynos 5420 with CPU Migration and that another Exynos is coming later this year (possibly 5430) with Global Task Scheduling.
-
phoenix_rizzen - Wednesday, September 11, 2013 - link
I don't understand the need for HMP in a phone, or even a tablet. Core migration makes sense (the OS only sees 4 CPU cores, the SoC switches between the A7 or A15 as needed on a per-Core basis). IOW, the system would be using one of the following setups, depending on the load:- 4x A7 + 0x A15
- 3x A7 + 1x A15
- 2x A7 + 2x A15
- 1x A7 + 3x A15
- 0x A7 + 4x A15
That setup makes sense to me, and can be done all in hardware. HMP doesn't; sounds like a lot of extra complexity and scheduler "magic" to make it work, and lots of ways to make it not work.
frenchy_2001 - Wednesday, September 11, 2013 - link
The goal is to allow more flexible configurations, like 4xA7+2xA15 or the next Tegra with 4xA15 and 2xA7 for power reduction (instead of 5xA15).This should allow for both better instant power and lower idle/low usage power, while keeping area under control (A7 is MUCH smaller than A15). Very powerful for the future.
I hope we will see more of those and I'm sure we'll find great new combinations for the best overall experience.
phoenix_rizzen - Monday, September 16, 2013 - link
A SoC with 4xA7 + 2xA15 would be interesting. Or, the opposite of the Tegra setup: 4xA7 (main cluster) + 1x A15 (for when you actually need the raw speed). :)In those situations, yeah, the HMP makes sense. But in a situation where the big and LITTLE sides have the same number of cores, CPU Migration makes more sense (IMO).
frenchy_2001 - Wednesday, September 11, 2013 - link
No, Exynos 5410 had a hardware bug limiting it to cluster migration (CCI-400 problem).The Exynos 5420 fixes this problem, however there was no software to take fully advantage of it, so the hardware would manage the migration and the scheduler would only see 4 virtual CPUs (each a pair of CPU, A7+A15).
Linaro worked on the Linux Kernel scheduler for Heterogeneous Multi-Processing, which allows to show all 8 cores to the scheduler and let it decide what to run where (while caring for power/temp limits). This is a software patch, that will be distributed around the end of the year. There will be no need for newer hardware.
ddriver - Wednesday, September 11, 2013 - link
Hey ANAND admins, how hard would be to block this spam, considering it always begins with "Live my job" and always has the $82h and sitting in front of the new imac - I mean come on, what is the excuse to keep tolerating this? Like it is not bad enough there is no edit button... come on, what are you people being paid for?jljaynes - Wednesday, September 11, 2013 - link
If Anandtech pays me $82/hr I will sit at my computer and delete spam for them.wintermute000 - Thursday, September 12, 2013 - link
I'll do it for $80/hrretrospooty - Wednesday, September 11, 2013 - link
"Hey ANAND admins, how hard would be to block this spam"They would probably have to put in one of those irritating pictured phrases with jumbled letters to confirm you are human. The cure is worse than the disease LOL.
Roland00Address - Wednesday, September 11, 2013 - link
Can we get a button to report said post to the moderator like we have for the forums?JarredWalton - Wednesday, September 11, 2013 - link
Generally speaking, any of us editors (Brian, me, Anand, Ryan, etc.) can mark something as spam and delete it, and we do it as soon as we see it. Of course, posting a reply to said spam ends up with your posts getting pushed to the bottom of the comment chain as a response to nothing. Considering this article has only been live for less than 12 hours and the spam is gone, I'd say we're doing okay. Just ignore the spammers and we'll take care of them within a few hours, or if you must you can email [email protected] and tell them which article has a spam comment.And as a final note, if we mark a comment as spam, it will delete all comments everywhere by that user, so if you value your account and comments please don't spam.
tuxRoller - Wednesday, September 11, 2013 - link
Just to be clear, this is not a "Samsung Innovation" but a project that has been coordinated, and fulfilled by Linaro, and others.I don't expect Samsung to play this any other way than "Look what WE'VE done!" but it'd be nice if it wasn't portrayed as something Samsung did since all you need to support the work is to have a compliant big.LITTLE system and run the new linux scheduler.
everex11 - Monday, September 16, 2013 - link
like Diane replied I am alarmed that a mother able to profit $5803 in 1 month on the computer. my site......www.pick85.com/