Samsung Chromebook (XE303) Review: Testing ARM's Cortex A15
by Anand Lal Shimpi on October 31, 2012 9:00 AM ESTPerformance: Atom vs. ARM's Cortex A15
In our iPhone 5 review I included this crude diagram comparing the high level micro-architecture layouts of the current crop of mobile CPU cores. While most of the cores looked fairly similar, the one to really stand out is ARM's Cortex A15. A three issue, OoO core, the Cortex A15 was designed to put ARM in a completely new performance category.
For 2013, the Cortex A15 is expected to be the ARM CPU core of choice for the next wave of smartphone and tablet SoCs. NVIDIA's Wayne should integrate multiple Cortex A15s as well as competing solutions from Samsung and TI. Samsung's Exynos 5 Dual, found in the new Chromebook, integrates two ARM Cortex A15 cores running at 1.7GHz. As our first experience with a Cortex A15, I wanted to get a good idea for how it would compare to Intel's Atom. And now me comparing to the older Chromebook 500 makes sense. The Atom N570 in the older Chromebook is the closest approximation we have to the currently shipping Atom based mobile SoCs. There are a number of enhancements to the newer chips (particularly when it comes to power consumption), but the base core is very similar. It's clear that the Sandy Bridge Celeron based Chromebook is faster than this new Cortex A15 design, but how about the old dual-core Atom model?
As Chrome OS is built around the Chrome browser, our tests had to be largely JavaScript based unfortunately. The good news is that even given the nature of the benchmarks, we're able to get a good feel for performance between the two SoC platforms. Both systems were running the latest version of Chrome OS at the time of publishing.
| Samsung Chromebook Performance Comparison | ||||||
| SunSpider 0.9.1 | BrowserMark | RIABench Focus Tests | Kraken | |||
| Atom N570 1.66GHz | 1034.3 ms | 152780 | 1968 ms | 14229.5 ms | ||
| Exynos 5 Dual 1.7GHz | 690.5 ms | 217031 | 1192 ms | 9733.2 ms | ||
The Cortex A15 is fast. Across the board we're seeing a 40 - 65% increase in performance over a dual-core Atom. Although it's not clear how performance will be impacted as companies work to stick Cortex A15 based SoCs in smartphones with tighter power/thermal budgets, in notebooks (and perhaps even tablets) the Cortex A15 looks capable of delivering a good 1 - 2 generation boost over Intel's original Atom core.
The IE10 browser tests tend to agree with our JavaScript performance tests, although the CSS Maze Solver benchmark shows a huge advantage for ARM over Intel's Atom here.
| Samsung Chromebook Performance Comparison | ||||||
| IE10 Bubbles Test | IE10 Fishbowl | IE10 Maze Solver | ||||
| Atom N570 1.66GHz | 11 fps | 5 fps | 45 seconds | |||
| Exynos 5 Dual 1.7GHz | 17 fps | 8 fps | 17 seconds | |||
GPU performance is an even bigger advantage for the Exynos 5 Dual over Intel's old Atom N570 (GMA-3150 GPU). I ran three different webGL tests, each of which showed just how bad the old Atom GPU core was.
| Samsung Chromebook GPU Performance Comparison | ||||||
| WebGL Solar System | WebGL Cubes (500) | WebGL Aquarium (50) | ||||
| Atom N570 1.66GHz | 2 fps | 10 fps | 2 fps | |||
| Exynos 5 Dual 1.7GHz | 22 fps | 28 fps | 38 fps | |||
This comparison isn't really all that fair as the newer Atom cores use Imagination GPUs, although even then they are using relatively underpowered solutions compared to what Samsung is shipping on the Exynos 5 Dual.
The more relevant conclusions here apply to the CPU comparison. Next year Intel is expected to introduce its first new Atom core since the platform's introduction five years ago. The new architecture will bring an Out of Order execution core as well as a tangible performance increase. The question is whether or not this will be enough to fend off advances from Cortex A15 based designs.
The new Chromebook (left) vs. the old Atom based Chromebook (right)
In our Surface review I looked at Clovertrail Windows 8 tablet performance and put it a good 40%+ faster than NVIDIA's Tegra 3. If ARM's Cortex A15 is able to outperform Clovertrail by a similar margin, it could make the next generation of Windows RT tablets even more attractive. Keep in mind that we're looking at an older Atom platform here and not Clovertrail, so the performance deltas could shrink a bit.
Facebook
Twitter
RSS
149 Comments
View All Comments
Krysto - Thursday, November 1, 2012 - link
Tablets and phones don't boot-up instantaneously. That's waking from idle mode that you're talking about, which basically just means turning on the screen.Krysto - Wednesday, October 31, 2012 - link
When Cortex A7 arrives in big.Little I expect the idle times to drop dramatically in A15 set-ups, and I doubt Intel can keep up with that. Not to mention they still won't keep up in graphics department.Anand, can you stop plugging Surface into Android and Chromebook reviews? Why would I want to hear how Surface might do in the future with this chip that Chromebook already has? Seriously.
Exophase - Wednesday, October 31, 2012 - link
Anand's Cortex-A15 diagram looks really weird vs an official one.. http://regmedia.co.uk/2011/10/20/arm_a15_pipeline_...Three instructions are decoded and dispatched to 5 clusters of 8 execution queues (8 slots each) per cycle, then each cycle each execution queue can issue a cycle to its associated execution unit (note this diagram is using issue/dispatch opposite from how it's often used). There's no execution relationship like given in Anand's diagram, the clusters are: NEON (2x), load/store (2x), simple ALU (2x), multiply, and branch.
Anand Lal Shimpi - Wednesday, October 31, 2012 - link
I based my diagram on a lot of the content in here:http://www.arm.com/files/pdf/AT-Exploring_the_Desi...
Slide 20 introduces the independent issue queues, although subsequent slides make it all a bit more ambiguous (and more like the block diagram you included). I didn't include "issue" width for each of the arrows in my diagram to keep thing simple but I can see that it can definitely give the wrong impression.
Let me see about doing a better pass on the A15 once I'm back from my current trip.
Take care,
Anand
madmilk - Wednesday, October 31, 2012 - link
More proof that ARM ISA isn't magically more efficient than x86 in perf/watt.My i7-3720QM Mac can complete Kraken in 2100ms, with a load power consumption of 30W and idle power of 12W (display on at mid-brightness, discrete GPU off). Ivy has very good power gating, so I'll just assume the CPU is consuming 20W. This results in 42 Joules to complete one run of Kraken. The A15 on the other hand, with a differential of 4 watts (again assuming perfect gating at idle) takes 38.8 Joules. In server usage, perf/watt is king. If the best ARM can deliver is 10% better perf/watt at the expense of 5x per-core performance, they're not going to make a dent against Xeon. I imagine V8 could be optimized more for ARM, but the gains there are not nearly as big as before.
Some other observations -
I ran 8 instances of Kraken at once to make use of all the cores and HT, while reducing the effect of perf/watt-killing Turbo Boost. Power consumption was about 65W total, or approximately 55W going into the CPU. Each Kraken finished in about 3700ms, so what we have is 55W*3.7s/8 = 25 joules per Kraken. Much better, but obviously an apples-to-oranges comparison against the A15 which should probably have two instances run on it for optimal efficiency.
Also, shame on Intel/Samsung/Google for the Chromebook 500. 12W idling for a netbook is just ridiculous.
I don't think we can pull the 32nm vs 22nm fab card here either, because Intel has always been ahead in this respect and the gap is widening.
lemonadesoda - Wednesday, October 31, 2012 - link
The new chromebook is a near-winner! While it is targetting impulse purchase, low-budget buyers, and 2nd world economies, they should consider a 1st world premium buyer audience. If this had a better quality screen with "retina" pixel density, I would buy one in a heartbeat. I don't want to game... I want to read my PDFs without having to zoom to make the text legible... and with the SD card reader... it would be a perfect image display device for the SD cards once it comes out of the camera. 1366x768 is very 2009, not 2012.ol1bit - Wednesday, October 31, 2012 - link
I don't understand why Chrome OS continues at all?Just slapping JellyBean on this hardware would have much more functionality.
andrewaggb - Wednesday, October 31, 2012 - link
Only google can answer this question....Kinda like Firefox OS... or whatever they are calling it. Just not sure what the point is. Maybe they figured linux needs just a few more radically different distributions :-)
ssddaydream - Wednesday, October 31, 2012 - link
Please, will somebody ban this clown?ShaneMickey, that is...
ssddaydream - Thursday, November 1, 2012 - link
...there will be support that would allow AndroidOS, Linux, or Windows RT to be installed and properly functioning on this device.