Samsung 8-Core Exynos 5 architecture @ ISSCC'13

pxc

Extremely [H]
Joined
Oct 22, 2000
Messages
33,063
Anandtech has coverage: http://www.anandtech.com/show/6768/samsung-details-exynos-5-octa-architecture-power-at-isscc-13

It's a big.LITTLE design, with 4 cores of each in 2 CPUs. The power efficient LITTLE quad core processor is designed to run most tasks, and the big cores use much more power for heavier computational tasks. The OS determines when to switch over from one CPU to another. Both cores are based on the same ARM v7 ISA.

The 4 ARM A7 LITTLE cores are very low power, maxing out under a Watt @ 1.2GHz full load (28nm process). But the 4 big A15 cores use a lot of power, up to almost 6W at full tilt. Ouch, won't see this in a phone. :p
 
The OS determines when to switch over from one CPU to another.

Minor error here. The OS actually cannot see the A15 quad nor the A7 quad. All it sees is a quad-core ARM v7. The SOC itself makes the decision on which quad to use based on power demands. This is transparent to the OS. This was explained by ARM but poorly explained (if not an outright error) by the Anandtech writer who wrote that piece.

But the 4 big A15 cores use a lot of power, up to almost 6W at full tilt. Ouch, won't see this in a phone.

The power envelope can and will be adjusted. We see this in SOCs like Tegra 3 and Exynos 4 when used in tablets and phones. The 6W figure provided was for 1.8ghz plus. Don't be surprised to see a 1.2-1.4ghz phone version.

The writer also goes on to explain that both SOCs can be used in parallel, something that both ARM and Samsung have previously debunked. They claim that the GPU has been switched from ARM Mali to PowerVR, but provide no source for this (and this was a rumor that had been questioned but not officially debunked in the past). They also go on to claim that it will likely be used in the Galaxy S IV. Samsung just recently announced Exynos5 Octa, with both Samsung and ARM claiming that the Big.Little architecture will likely debut in early 2014, or late 2013 at the earliest. Samsung announced Exynos5 -dual and -quad early last year and Anand tried to start rumors that it would be in the S III and Note II, but those got quad-core Exynos 4 chips. Samsung typically announces a product, samples it for new products 6 months later, and then has shipping products about 1 year after announcement. To expect this in the Galaxy S IV at this point is VERY premature.

I'm open to eating crow on this, but when it comes to predicting Samsung's hardware over the last two years, I've done a far better job than Anand has. Don't believe me? Look me up over at XDA (username jaykresge).
 
I'm kinda doubtful myself that it'll be out anytime real soon. Quad cores just made it into phones not that long ago and now we're getting 8 cores? Don't get me wrong I'm always for moar coars but a quad rips thru anything you can throw at it now. Not sure the battery life hit would be worth it and not sure you'll see a benefit. He'll 8 core chips hardly get you much boost over quad cores in a desktop much less a phone.
 
I'm kinda doubtful myself that it'll be out anytime real soon. Quad cores just made it into phones not that long ago and now we're getting 8 cores? Don't get me wrong I'm always for moar coars but a quad rips thru anything you can throw at it now. Not sure the battery life hit would be worth it and not sure you'll see a benefit. He'll 8 core chips hardly get you much boost over quad cores in a desktop much less a phone.

It's not a real 8-core. You won't have 8-cores running on the phone. They simply put a small, low-power quad-core next to the main chip, and the SOC determines which of the two to use depending on the load.

To give a better example; the Ivy-Bridge based Core-i5 is quad-core on the desktop. Imagine next year Haswell is advertised as an 8-core, but it simply this year's IB i5 paired with a quad-core Atom, and the CPU used is switched based on the demand of the system in an effort to save power. As in the case of Big.Little, we're not seeing an octa-core in action. We're simply seeing a high-powered and a low-powered quad-core alternate based on the demands of the system.
 
Yeah I get that and I do get how it could work well but I'm just thinking that with today's current quad core mobile procs being more than enough to crush anything out there you might throw at it, I don't see them rushing this to market anytime soon. But just a guess on my part.
 
Yeah I get that and I do get how it could work well but I'm just thinking that with today's current quad core mobile procs being more than enough to crush anything out there you might throw at it, I don't see them rushing this to market anytime soon. But just a guess on my part.

Its not about crushing performance its about battery life.
 
Its not about crushing performance its about battery life.

It's an attempt at addressing both, actually.

What happens is that the scheduler sends the threads to either A15 or the A7 depending on the task. For example, if you're gaming then you want all the goodies turned up so the A15s will handle that workload (and tone down clock speed to fit within a certain TDP). Likewise, if you're just browsing the web and aren't putting much stress on the SoC, the A7s offer a perf-per-watt that's simply unmatched.

I do think this approach will fare much better on tablets than it will in phones. As tablets start to chew up even more laptop and basic PC computing needs for a good portion of users, those same users will require cores that can handle both heavy and light tasks.

Not everyone is sold on this approach, though. Qualcomm isn't jumping on the big.LITTLE bandwagon, but that may be because it's still a bit too early to do so. It's also important to keep in mind that ARM's cores are designed to be tinkered with. They make a good-but-still-basic core that can fit as many customer needs as possible with the implication that the core architecture can be tweaked to better serve a specific purpose.
 
It's an attempt at addressing both, actually.

He was referring to top-end performance, in which case he is still correct. Big.Little adds no extra performance (unless you enable all 8-cores, which ARM now says can be done for specific devices, but not on-demand for most devices). Even in that cases, four A7s add marginally little power to the mix. A quad-A15 at 1.5ghz won't be any slower than a Big.Little at 1.5ghz, it just won't offer the perf-per-watt for lower demanding tasks.

Not everyone is sold on this approach, though. Qualcomm isn't jumping on the big.LITTLE bandwagon, but that may be because it's still a bit too early to do so.

ARM uses ASMP, which has advantages and disadvantages when compared to SMP, used in the Cortex series. SMP defeated ASMP in the desktop space a long time ago. A popular SMP vs. ASMP debate was in regards to the disaster that was the Sega Saturn, which used the ASMP approach.

ASMP offers better power scaling (talking about power consumption, not performance), but worse multi-threaded performance. Big.Little is an attempt by ARM to have their cake and eat it too, the higher multi-threaded performance of SMP, but having power scaling closer to ASMP.
 
He was referring to top-end performance, in which case he is still correct. Big.Little adds no extra performance (unless you enable all 8-cores, which ARM now says can be done for specific devices, but not on-demand for most devices). Even in that cases, four A7s add marginally little power to the mix. A quad-A15 at 1.5ghz won't be any slower than a Big.Little at 1.5ghz, it just won't offer the perf-per-watt for lower demanding tasks.



ARM uses ASMP, which has advantages and disadvantages when compared to SMP, used in the Cortex series. SMP defeated ASMP in the desktop space a long time ago. A popular SMP vs. ASMP debate was in regards to the disaster that was the Sega Saturn, which used the ASMP approach.

ASMP offers better power scaling (talking about power consumption, not performance), but worse multi-threaded performance. Big.Little is an attempt by ARM to have their cake and eat it too, the higher multi-threaded performance of SMP, but having power scaling closer to ASMP.

Everything in the Saturn was poorly slapped together. The second process was added to the Saturn at last minute after Sega saw the specs of the Playstation.
As a result Sega's dev kits initially didn't fully support all the poorly implement second processor. The lead to third partiest have to hack together their own libraries.
It's more than just the fact that it was ASMP.
In the last year of the Saturn when they(Sega) figured out the thing they were able to achieve a lot with it(See the unreleased Shenmue footage on youtube).
 
Everything in the Saturn was poorly slapped together. The second process was added to the Saturn at last minute after Sega saw the specs of the Playstation.
As a result Sega's dev kits initially didn't fully support all the poorly implement second processor. The lead to third partiest have to hack together their own libraries.
It's more than just the fact that it was ASMP.
In the last year of the Saturn when they(Sega) figured out the thing they were able to achieve a lot with it(See the unreleased Shenmue footage on youtube).

Regardless, the primary reason developers hated ASMP was the inability for multiple cores to access the cache simultaneously. This led to an alternating approach that wasn't ideal. It's not "true" multi-threading as with today's SMP-based cores. And that is a failing long term for Snapdragon if they don't switch to SMP in the next generation or two.

For today's smartphones, it's fine. Even the most demanding user won't find too much fault with it. But as devices begin to converge even further (think Asus Padphone, but Windows 9 running on phone/tablet/desktop), SMP will matter more.
 
Back
Top