Mediatek MOAR COARS!

[DrMrLordX]

Dude I ran multiple programs under DOS (well, tried to anyway) using DOS shell tricks and other nonsense. I had my awesome .mod/.st3/.mtm player in the background while logging in to my favorite BBSes, so l337.

Still not enough to keep ten cores busy, but whatever.

 

[Exophase]

Dude I ran multiple programs under DOS (well, tried to anyway) using DOS shell tricks and other nonsense. I had my awesome .mod/.st3/.mtm player in the background while logging in to my favorite BBSes, so l337.

Still not enough to keep ten cores busy, but whatever.

Worked great I'm sure, until more than one of those programs wanted the timer interrupt.

 

[lopri]

I do not know why people think this is a marketing. I mean, yes, the OEMs may hope they sell better but I have seen zero evidence that this so-called "marketing" actually works. It's the usual tech-press and "analysts" who float these "moar cores sell in certain part of the globe" meme. Last I've heard iPhones sell quite well in China despite having only two cores.

I suspect this push towards more cores have nothing to do with performance nor marketing, but something more asinine.

 

[imported_ats]

I like moar coars!!!

While it may be true that it's hard to write software to use more than four cores, it's not hard to run a bunch of programs that don't use more than four at the same time. It's not DOS anymore - can actually run more than one task at a time.

One task, one core. Is that asking for so much?

;-)

Even running a bunch of programs, its hard to actually use more than a couple cores. You pretty much need heavy computation background programs of which there are very very few outside of content creation (rendering) or engineering. In the real world, most of those computationally intensive background programs are farmed out to compute farms anyways. So outside of server workloads, 4 cores is basically more than enough for 99.99% of all desktop users. All going to more cores does is slow down the core you are actually using.

 

[Exophase]

I do not know why people think this is a marketing. I mean, yes, the OEMs may hope they sell better but I have seen zero evidence that this so-called "marketing" actually works. It's the usual tech-press and "analysts" who float these "moar cores sell in certain part of the globe" meme. Last I've heard iPhones sell quite well in China despite having only two cores.

I suspect this push towards more cores have nothing to do with performance nor marketing, but something more asinine.

If no one cared about AnTuTu scores for marketing Intel wouldn't have so blatantly broken the benchmark (and just in time for a "third party" to publish a report about how amazing Intel SoCs are because AnTuTu)

Incidentally, if the slides have been any indication AnTuTu has been MediaTek's most cherished benchmark too.

 

[jhu]

If no one cared about AnTuTu scores for marketing Intel wouldn't have so blatantly broken the benchmark (and just in time for a "third party" to publish a report about how amazing Intel SoCs are because AnTuTu)

Incidentally, if the slides have been any indication AnTuTu has been MediaTek's most cherished benchmark too.

I think the only ones who truly care about AnTuTu and these other useless benchmarks are the CPU and device manufacturers.

 

[lopri]

What about counter-evidence I exampled above? iPhones? Krait-based Snapdragons? Or Sandy Bridge v. Bulldozer? There just isn't an evidence that consumers prefer mediocre many-cores over competently designed SOCs with fewer cores, in spite of the prevailing narrative (which I believe was manufactured out of thin air).

I guess that is where Mediatek et. al come in. They don't have custom cores and hadn't been able to build an SOC using ARM's big cores (A15/A57) until recently. All they could muster up were A7s, A9s, and A53s. They probably thought their products were not standing out enough (duh) and looked for ways to differentiate. That, I think, may have been one driving force behind this little.LITTLE and now big.Medium.LITTLE many-core phenomenon. These fast-growing SOC vendors desire to be recognized in the market but they are not there yet with the top dogs, so their chosen tactic for the time being is packing as many of ARM's LITTLE cores and trying to persuade consumers that their chips are different and worth paying for.

In other words, my view is that this many-core marketing story is a self-fulfilling prophesy originated from within the industry, and has little to do with consumers being none the wiser.

 

[Exophase]

I think the only ones who truly care about AnTuTu and these other useless benchmarks are the CPU and device manufacturers.

I've heard stories otherwise where device OEMs care. That's all that really matters here.

What about counter-evidence I exampled above? iPhones? Krait-based Snapdragons? Or Sandy Bridge v. Bulldozer? There just isn't an evidence that consumers prefer mediocre many-cores over competently designed SOCs with fewer cores, in spite of the prevailing narrative (which I believe was manufactured out of thin air).

That's a false dichotomy. It's not about preferring mediocre many-cores over good SoCs with fewer cores, it's about preferring the SoC which has the same cores but more. SoCs used only in iOS devices don't really play in the same market because people are divided by many other preferences that differentiate them from Android devices. And of course as I said before the desktop/laptop chips have nothing to do with this.

 

[lopri]

So consumers care for more cores on smartphones but not on desktops/laptops?

 

[ClockHound]

Even running a bunch of programs, its hard to actually use more than a couple cores. You pretty much need heavy computation background programs of which there are very very few outside of content creation (rendering) or engineering. In the real world, most of those computationally intensive background programs are farmed out to compute farms anyways. So outside of server workloads, 4 cores is basically more than enough for 99.99% of all desktop users. All going to more cores does is slow down the core you are actually using.

Content creation is my real world - that's why moar coars matter moar. And as has been proven time and time again in various online polls that I've sponsored, my needs are far important than your hypthetical 99.99%ers.

Oh..wait this is about mobile cores. Why would I care? Consumers. They'll buy anything they're told to. Let them eat iPads!

:biggrin:

 

[lopri]

Well, in any case we will learn what consumers think by how this new 10-core SOC performs in the market. My prediction is that it will tank because consumers do not go blindly after moar cores.

 

[Exophase]

So consumers care for more cores on smartphones but not on desktops/laptops?

Again, on mobile those extra cores means a small extra price for the SoC. The choice is between more cores or a minor cost savings.

If the choices were 4 core i7 for $250 or 8 core i7 for $270 do you think no one would care about the 8 core version?

 

[imported_ats]

Again, on mobile those extra cores means a small extra price for the SoC. The choice is between more cores or a minor cost savings.

If the choices were 4 core i7 for $250 or 8 core i7 for $270 do you think no one would care about the 8 core version?

No that isn't the choice. The choice is between lesser faster better performing cores or slower worst performance more cores. Literally every major smart phone SOC using 4+4 bl would be better off with 2+2. The end user performance would increase, it would use lower energy on average, etc.

Same with your i7, even though Intel has much much better power management, 99.99% of users would be much better off with the lower core count but higher frequency cores.

 

[Ruiner1]

I've heard stories otherwise where device OEMs care. That's all that really matters here.

And device OEMs care mostly because the press seems to report this crap. Whatever gets headlines is considered to be high priority. It seems that Anandtech is one of the few who try to put the effort into understanding what they are actually talking about.

 

[scannall]

I like moar coars!!!

While it may be true that it's hard to write software to use more than four cores, it's not hard to run a bunch of programs that don't use more than four at the same time. It's not DOS anymore - can actually run more than one task at a time.

One task, one core. Is that asking for so much?

;-)

Yeah, it is. Your computer/phone/tablet etc has a lot of processes running at the same time anyway. In the mobile space, race to sleep is a big deal. Get the job done, then sleep. If you are using a bunch of small, high frequency cores then it will take longer for them to get into the sleep state. And use more energy getting there because of the high frequency.

We've had multi-tasking for far longer than we've had more than one core.

It's a bad design choice for mobile devices.

 

[Fjodor2001]

Intel cores are big (much, much bigger than Cortex-A53), putting 8 or even 6 of them on a mainstream consumer chip results in a much larger and more expensive die, which is not a cheap value proposition for something most users don't benefit from. So why bother?

If they could put 4 cores in the Q6600 on 65 nm, they ought to be able to put 8 cores on 14 nm! The transistor density has increased by a factor 16x or so going from 65->14 nm!

Intel already sells a class of server CPUs that are the same die as the desktop ones.. the statement basically boils down to "Intel doesn't want to sell 8 core chips because Intel doesn't want to cannibalize their 8 core desktop chips." The fact is that they DO sell 8 core chips for desktop enthusiasts, they just charge a lot of money for them, because they can.

The current 8 core desktop CPUs are enthusiast class, not mainstream, and thus priced after that, which also is the reason they are sold in very small quantities. So not comparable to e.g. Q6600 on 65 nm.

That would be the same if they made an 8 core desktop chip that ran on the same socket that the mainstream desktop chips run on, only it'd be substantially more expensive because it'd have to absorb the costs of making yet another die for a product few people want.

It all depends on the price. If priced by die area, I'm quite sure a lot of people would buy an 8 core mainstream desktop CPU.

Honestly, on 14 nm we should have a product mix looking something like this by now:

Pentium: 2 cores / threads
i3: 4 cores / 4 threads
i5: 4 cores / 8 threads (or perhaps 6 cores)
i7: 8 cores / 16 threads

Most people would likely by the Core i3 & i5, just as now. But I'm quite sure enough people would get the i7 for it to be justified.

 

[SirCanealot]

People still need to try to keep in mind that processors that aren't being used are going to be turned off anyway. I could build a 10,000 core processor that doesn't use any more power than a quad core if those extra cores are never used.

With BIG.little we're going to see a lot of experimentation over the next few years and I think it will take some time before we see usage patterns and how specific setups function. All of these setups have their pros and cons.

If a company wants to try putting out a 10 core processor -- why not? It's only costing them on silicone budget; you're not paying extra for a phone with extra cores. If the software is build correctly, it's still going to work pretty well.

And I have the Snapdragon 800 Note 3. You'd be surprised at how often the phone uses all 4 cores and how laggy it is when I force it to 2 core mode to save battery. Any sort of multitasking, especially involving Chrome (that being the memory hog that it is) will plug in all 4 cores quite easily. If you think that quad core in your phone isn't cool, root and disable two of the cores and you'll see what it's like

 

[DrMrLordX]

Worked great I'm sure, until more than one of those programs wanted the timer interrupt.

It usually worked. Maybe I'd get the occasional static from the audio program running in the "background", but that was almost never. Doing anything more intensive than loading up the term program would usually lock up the 'puter though.

I think the main reason why it worked, is that the shell was an extension of the audio player program, rather than a separate shell/menu program. But that's ancient history.

Back on topic: If the consumer market keeps getting kludged up with multicore madness, it seems to me that some of these phone SoCs could be cut down a little and used as microserver CPUs. The variable clockspeeds could be a headache, but it might not be that big of a deal if the scheduler knows how to deal with the problem. You'd just have to cut out some of the SoC features that wouldn't necessarily be useful in a microserver.

 

[Fjodor2001]

People still need to try to keep in mind that processors that aren't being used are going to be turned off anyway. I could build a 10,000 core processor that doesn't use any more power than a quad core if those extra cores are never used.

With BIG.little we're going to see a lot of experimentation over the next few years and I think it will take some time before we see usage patterns and how specific setups function. All of these setups have their pros and cons.

If a company wants to try putting out a 10 core processor -- why not? It's only costing them on silicone budget; you're not paying extra for a phone with extra cores. If the software is build correctly, it's still going to work pretty well.

And I have the Snapdragon 800 Note 3. You'd be surprised at how often the phone uses all 4 cores and how laggy it is when I force it to 2 core mode to save battery. Any sort of multitasking, especially involving Chrome (that being the memory hog that it is) will plug in all 4 cores quite easily. If you think that quad core in your phone isn't cool, root and disable two of the cores and you'll see what it's like

+1

Yeah really, if Intel is not putting more CPU cores on the mainstream dies soon, the die area will be ridiculously small as node shrinkage continues. It's been 4 node shrinks since the last time we saw a core count increase...

 

[ShintaiDK]

+1

Yeah really, if Intel is not putting more CPU cores on the mainstream dies soon, the die area will be ridiculously small as node shrinkage continues. It's been 4 node shrinks since the last time we saw a core count increase...

And? If software doesnt use it its just useless idle cores. Then it doesnt matter how small they are. Phone SoCs are an utter joke there. And do I even have to mention TDP?

 

[Fjodor2001]

Depends on the workload. It would be more useful than wasting ever more die area on the iGPU at least.

There are for sure enough people that would get an 8 core desktop CPU for it to be justified on desktop if available in the mainstream lineup.

TDP is not a problem on desktop. At least not on 14 nm, if the iGPU is not too bloated.

 

[mohit9206]

iPhone 6 still uses dual core chip and probably still be dual core in 6S yet it still apparently blows away any Mediatek chip.
So what we need is more powerful and efficient cores not MOAR cores.

 

[monstercameron]

iPhone 6 still uses dual core chip and probably still be dual core in 6S yet it still apparently blows away any Mediatek chip.
So what we need is more powerful and efficient cores not MOAR cores.

In what metric?

 

[Fjodor2001]

iPhone 6 still uses dual core chip and probably still be dual core in 6S yet it still apparently blows away any Mediatek chip.
So what we need is more powerful and efficient cores not MOAR cores.

Sure if that would be possible, it would be preferred. But in reality inreasing ST performance has proven to be ever more difficult.

Especially on desktop where we're getting the yearly 6-7% increase, which is ridicolously low. On mobile we're srarting too hit the ST performance wall too. That's why increasing core count and MT performance is the only realistic way to improve performance in any meaningful way.

 

[monstercameron]

Sure if that would be possible, it would be preferred. But in reality inreasing ST performance has proven to be ever more difficult.



Especially on desktop where we're getting the yearly 6-7% increase, which is ridicolously low. On mobile we're srarting too hit the ST performance wall too. That's why increasing core count and MT performance is the only realistic way to improve performance in any meaningful way.

Also note that raising clock speeds might not be as feasible as it it on the desktop, higher clocks means higher voltage more heat and more power.