The x86 Power Myth Busted: In-Depth Clover Trail Power Analysis(anandtech.com)
anandtech.com
The x86 Power Myth Busted: In-Depth Clover Trail Power Analysis
http://www.anandtech.com/print/6529
8 comments
The Intel version of the surface tablet is more than twice as fast (more like 3-4x faster depending on benchmark and workload -- things like SIMD or L3-cache-bound tasks are well over 10x faster) as the ARM version. ARM still has a way to go.
Apples to oranges. The Surface RT is an iPad competitor. The Surface Pro is a PC. You can make business arguments as to which is the more appropriate device and which makes the better tradeoffs, but there's no serious technical point to be made here.
Apples to oranges. The Surface RT is an iPad competitor. The Surface Pro is a PC. You can make business arguments as to which is the more appropriate device and which makes the better tradeoffs, but there's no serious technical point to be made here.
The Intel version of the Surface is an ultrabook-class i5 CPU. It is dozens of times faster than the ARM one. They're just not comparable at all. True, it has much less battery life, but also significantly more processing power... enough to run full scale Photoshop.
Intel has much to fear from ARM, but not from NVidia... especially big.little. Outside of NVidia-specific project Shield, it doesn't sound like Tegra 4 has any design wins. In fact, NVidia's big.little architecture is arguably a failure. Surface RT == NVidia Tegra 3 with big.little (btw: Surface RT doesn't use big.little features at all. Perhaps Microsoft is trying to keep SurfaceRT code portable across different ARM chips). In comparison, the iPad4 is an Apple custom ARM chip, and the Nexus 4 is Qualcomm Snapdragon. Current Nexus 7 is Tegra3, but the next one is rumored to be Qualcomm Snapdragon as well.
Basically, NVidia is weakening in the mobile sector. ARM overall is strong, but big.little is arguably a failure at this point.
Intel has much to fear from ARM, but not from NVidia... especially big.little. Outside of NVidia-specific project Shield, it doesn't sound like Tegra 4 has any design wins. In fact, NVidia's big.little architecture is arguably a failure. Surface RT == NVidia Tegra 3 with big.little (btw: Surface RT doesn't use big.little features at all. Perhaps Microsoft is trying to keep SurfaceRT code portable across different ARM chips). In comparison, the iPad4 is an Apple custom ARM chip, and the Nexus 4 is Qualcomm Snapdragon. Current Nexus 7 is Tegra3, but the next one is rumored to be Qualcomm Snapdragon as well.
Basically, NVidia is weakening in the mobile sector. ARM overall is strong, but big.little is arguably a failure at this point.
The Tegra 3 doesn't use big.LITTLE. It does have a 5th, smaller core - but that's not big.LITTLE.
Thanks for the clarification. I guess I got things mixed up.
Are there any big.LITTLE implementations out there? Wikipedia only mentions Samsung Exynos 5 Octa, which doesn't have any designs right now.
Are there any big.LITTLE implementations out there? Wikipedia only mentions Samsung Exynos 5 Octa, which doesn't have any designs right now.
I don't believe so, but willing to be proven wrong!
What about tablets like the Lenovo thinkpad tablet 2? Atom powered, gets 7h+ battery life, weight and power comparable to arm tabletsa, and can run x86 code.
Previous discussion: https://news.ycombinator.com/item?id=4964355
I'm not sure if this is a place to ask questions, but here it is: How can I compare ARM and x86 CPUs based on their specs, e.g. without actually testing them. RaspberryPi website says its "overall real world performance is something like a 300MHz Pentium 2" although the frequency of the chip is 700MHz. Can I use 3/7 as a rule of thumb?
No, different specific architectures can have more than an order of magnitude difference in performance per clock in ways that don't have anything to do with the instruction set used.
In this case, the Pentium 2 is an out of order 2-wide superscalar architecture and so is the A9 ARM processor talked about in this article, but the ARM 11 in the RasberryPi is a 1-wide in order processor, which mostly accounts for the wide disparity in performance. You'd actually normally expect the difference to be larger, but ARM instructions are a bit more powerful than x86 instructions so you'll tend to need less of them[1]. But x86 is more efficiently encoded, so you only need the same number of bytes of instructions to accomplish the same task for each.
The best way to compare different processors is by benchmarks. But remember to choose a benchmark that looks like what you're trying to do, because different architectures can excel in different situations.
[1] x86 has the advantage of having combined load/store and operate instructions, but ARM has 3-operand instructions, twice the registers instruction predication, and a free barrel roll when using it's ALU. Changes to the x86 instruction set since the Pentium 2 make things more fair.
In this case, the Pentium 2 is an out of order 2-wide superscalar architecture and so is the A9 ARM processor talked about in this article, but the ARM 11 in the RasberryPi is a 1-wide in order processor, which mostly accounts for the wide disparity in performance. You'd actually normally expect the difference to be larger, but ARM instructions are a bit more powerful than x86 instructions so you'll tend to need less of them[1]. But x86 is more efficiently encoded, so you only need the same number of bytes of instructions to accomplish the same task for each.
The best way to compare different processors is by benchmarks. But remember to choose a benchmark that looks like what you're trying to do, because different architectures can excel in different situations.
[1] x86 has the advantage of having combined load/store and operate instructions, but ARM has 3-operand instructions, twice the registers instruction predication, and a free barrel roll when using it's ALU. Changes to the x86 instruction set since the Pentium 2 make things more fair.
Look at benchmarks. It's not perfect, but that's the only reasonable way.
While this looks great, isn't there a fundamental issue that the test tablets were running different OSes? For instance, Microsoft's x86 code might have a decade more tuning done to it versus their ARM code.
Comparing the power consumption of a 40nm chip and a 32nm chip and trying to draw conclusions about what's the better /architecture/? Please. That's like playing a tennis match between Federer and Nadal, making Nadal play with a frying pan instead of a racquet, and then claiming Federer is the better player because he won using a racquet.
Cool data, though.
Cool data, though.
Intel had a lead in performance while they were working on fixing power consumption. Because it took them too long to fix that, now they lost the performance lead in both CPU, and especially in GPU performance. Apple gets credit all the time for having the "fastest chip", when it really only has the fastest GPU, but it seems people even forget mentioning that Intel has the weakest GPU, even when compared to Tegra 3.
I'm sorry, what's the evidence for having "lost the performance lead in CPU performance"? Atom continues to beat the A9, A15 and Krait in most benchmarks.
And the GPU point is just ridiculous on its face: the A6X in the iPad and the Clover Trail part tested here are using the same GPU. Apple is faster because it put more of them on the chip. That's not a architectural triumph, it's a transistor budget decision. Intel could do the same thing (hell, at 22nm it could do it much better -- Clover Trail is 32nm) with more or less the same cost (again, Intel probably has a cost advantage here because of its more advanced semiconductor process). That they chose not to may be a mistake, but it's an easily correctable one.
And the GPU point is just ridiculous on its face: the A6X in the iPad and the Clover Trail part tested here are using the same GPU. Apple is faster because it put more of them on the chip. That's not a architectural triumph, it's a transistor budget decision. Intel could do the same thing (hell, at 22nm it could do it much better -- Clover Trail is 32nm) with more or less the same cost (again, Intel probably has a cost advantage here because of its more advanced semiconductor process). That they chose not to may be a mistake, but it's an easily correctable one.
Intel wasn't "fixing" power consumption, but rather it simply wasn't a critical requirement: The CPU was often a small part of the overall platform power consumption, and people favoured performance over efficiency.
Obviously tablets and mobile devices change that equation, so Intel adapted. Intel is in a difficult situation with their products however in that one of their biggest concerns is competing with themselves too much.
Regarding the GPUs, Apple doesn't have the fastest mobile GPU - PowerVR from Imagination Technologies does. Intel can use a PowerVR SGX554MP4 if they wanted to integrate it. Is it better than an HD 4000? I honestly have no idea as I've never seen them actually compared (and note that the best mobile GPUs have miserable performance compared to discrete desktop GPUs, so don't assume that because the HD 4000 doesn't compete with a $400 gaming card that it can't compete with an Adreno or PowerVR).
EDIT: Just blew my mind to learn that the HD 4000 on the Intel SoC is a PowerVR GPU.
Obviously tablets and mobile devices change that equation, so Intel adapted. Intel is in a difficult situation with their products however in that one of their biggest concerns is competing with themselves too much.
Regarding the GPUs, Apple doesn't have the fastest mobile GPU - PowerVR from Imagination Technologies does. Intel can use a PowerVR SGX554MP4 if they wanted to integrate it. Is it better than an HD 4000? I honestly have no idea as I've never seen them actually compared (and note that the best mobile GPUs have miserable performance compared to discrete desktop GPUs, so don't assume that because the HD 4000 doesn't compete with a $400 gaming card that it can't compete with an Adreno or PowerVR).
EDIT: Just blew my mind to learn that the HD 4000 on the Intel SoC is a PowerVR GPU.
HD4000 is far more powerful than a PowerVR SGX554MP4, or whatever mobile chips are out there.
http://cdn.itproportal.com/photos/ibvssb_original.jpg
The HD4000 is also ~20% of a 166mm^2 Ivy Bridge on 22nm. If HD4000 were made on 32nm process, it'd be well over 60mm^2 alone.
The PowerVR SGX554's 8.8mm^2 at 40nm, or approximately 5.6mm^2 at 32nm.
The Intel HD4000 is about 12x larger than PowerVR SGX554. They just aren't in the same class... they're no where close to the same class.
--------------
BTW: the Clovertrail Atom uses PowerVR Graphics. Intel could have bought any of the PowerVR designs, but they chose a significantly weaker one for some reason.
--------------
With regards to your edit: The Atom doesn't have an HD4000 at all.
http://cdn.itproportal.com/photos/ibvssb_original.jpg
The HD4000 is also ~20% of a 166mm^2 Ivy Bridge on 22nm. If HD4000 were made on 32nm process, it'd be well over 60mm^2 alone.
The PowerVR SGX554's 8.8mm^2 at 40nm, or approximately 5.6mm^2 at 32nm.
The Intel HD4000 is about 12x larger than PowerVR SGX554. They just aren't in the same class... they're no where close to the same class.
--------------
BTW: the Clovertrail Atom uses PowerVR Graphics. Intel could have bought any of the PowerVR designs, but they chose a significantly weaker one for some reason.
--------------
With regards to your edit: The Atom doesn't have an HD4000 at all.
Indeed, I was more dovetailing off of the comparisons between the Surface RT and Surface Pro, where the latter has the HD 4000, and the GP's discussion about "Intel's" verus "Apple's" GPUs. As you mentioned in their nascent "mobile" chips they're using PowerVR (as I mentioned, apparently even the HD 4000 is licensed and derived from PowerVR), and they have all of the same capacity as Apple to choose whichever GPU they want.
Makes sense. Although for your last comment...
http://semiaccurate.com/2013/01/02/apples-silicon-design-cap...
For those without a subscription. That site is spreading a rumor that Apple has bought out a top-tier AMD grpahics team. Considering that Apple is now designing their own chips, and just headhunted an entire AMD (former ATI) team, we can see where Apple is going.
Apple does not want to be relying on commodity parts. They are positioning themselves to build a custom GPU a few years down the line.
So Intel currently can catch up to Apple GPU performance by just buying the latest PowerVR graphics. But will they be able to do that in the future, if Apple creates their own top-tier custom GPU?
http://semiaccurate.com/2013/01/02/apples-silicon-design-cap...
For those without a subscription. That site is spreading a rumor that Apple has bought out a top-tier AMD grpahics team. Considering that Apple is now designing their own chips, and just headhunted an entire AMD (former ATI) team, we can see where Apple is going.
Apple does not want to be relying on commodity parts. They are positioning themselves to build a custom GPU a few years down the line.
So Intel currently can catch up to Apple GPU performance by just buying the latest PowerVR graphics. But will they be able to do that in the future, if Apple creates their own top-tier custom GPU?
That wasn't my point. I know they use PowerVR GPU's, but they never use the latest ones. In fact they always use very old PowerVR GPU's. They had nothing last year that was even close to the other GPU's let alone the ones in iPads and the iPhone 5, and they will have nothing this year to compete with Apple's PowerVR Series 6 GPU. Their PowerVR drivers suck, too.
Why even bring the Intel HD 4000 in the discussion? We're talking about Atom here. Not a $200 17W TDP Core ULV Core i5 chip. If you want to compare HD4000 do it against Nvidia and AMD's own chips in laptops.
Why even bring the Intel HD 4000 in the discussion? We're talking about Atom here. Not a $200 17W TDP Core ULV Core i5 chip. If you want to compare HD4000 do it against Nvidia and AMD's own chips in laptops.
Note that the "latest" PVR cores are really just wider variants of the early ones. GPUs are mostly commoditized at this point, and short of increasing the core count and clock speed there's really not much that can be done architecturally to improve them. All of the "PVR SGX 5[345]5(MP[24])" are really just the same architecture.
That wasn't my point.
The GPU is almost a drop-and-play component on an SoC. Given that Intel could very rapidly incorporate any licensed GPU, such as the best of the best that Apple uses, what exactly is your point? You clearly are trying to hold the notion that Intel is so far behind, led off course by their energy quest that now they can't catch up on the GPU, yet they have exactly the same capacity as Apple does (actually let me amend that -- as the world's most advanced fab and chipmaker, they have dramatically more capacity than Apple does).
So why haven't they? Well thus far the only Atom-powered mobile devices have been pitched essentially to the third world. Intel is going through the mobile growing pains in the very low cost market, and when they decide to go for the mainstream I suspect that -- just like Apple -- they can decide to go for GPU performance as well.
I mentioned the HD 4000 given that you oddly decided to talk about "Intel's" and "Apple's" GPUs, and while you've quite rapidly shifted course to a more enlightened position now based upon reading other replies, I suspect you weren't quite aware of what the actual GPU sources were.
The GPU is almost a drop-and-play component on an SoC. Given that Intel could very rapidly incorporate any licensed GPU, such as the best of the best that Apple uses, what exactly is your point? You clearly are trying to hold the notion that Intel is so far behind, led off course by their energy quest that now they can't catch up on the GPU, yet they have exactly the same capacity as Apple does (actually let me amend that -- as the world's most advanced fab and chipmaker, they have dramatically more capacity than Apple does).
So why haven't they? Well thus far the only Atom-powered mobile devices have been pitched essentially to the third world. Intel is going through the mobile growing pains in the very low cost market, and when they decide to go for the mainstream I suspect that -- just like Apple -- they can decide to go for GPU performance as well.
I mentioned the HD 4000 given that you oddly decided to talk about "Intel's" and "Apple's" GPUs, and while you've quite rapidly shifted course to a more enlightened position now based upon reading other replies, I suspect you weren't quite aware of what the actual GPU sources were.
Intel never had any lead what so ever in this sector. Compar ing ARM chips in the mobile sector against previous generation Atoms is like comparing a laptop to a oct-core Xeon. They are different chips for different markets.
Unless you're talking about Z-class atoms (and almost no one actually does), the older Atoms (ie: D525 or N575) were netbookclass 10-Watt processors.
Intel hasn't even made an Atom on their most advanced process node yet. Intel's advanced factories are too busy selling Destkop and Laptop i5s or i7s to the masses. Clovertrail a 32nm design, while an Ivy Bridge (which came out months ago) is a 22nm design.
While yes, you're right. The Atom barely compares to the ARM chips... it also isn't Intel's best shot at the issue quite yet. Perhaps Intel doesn't have the smarts to execute a full scale attack at the mobile market, or maybe they're satisfied with their desktop / laptop sales... But whatever the reason, creating Atoms on older processes puts them at a self-inflicted disadvantage in the marketplace.
When Intel actually brings their A-game to the mobile market, I'll be watching.
Unless you're talking about Z-class atoms (and almost no one actually does), the older Atoms (ie: D525 or N575) were netbookclass 10-Watt processors.
Intel hasn't even made an Atom on their most advanced process node yet. Intel's advanced factories are too busy selling Destkop and Laptop i5s or i7s to the masses. Clovertrail a 32nm design, while an Ivy Bridge (which came out months ago) is a 22nm design.
While yes, you're right. The Atom barely compares to the ARM chips... it also isn't Intel's best shot at the issue quite yet. Perhaps Intel doesn't have the smarts to execute a full scale attack at the mobile market, or maybe they're satisfied with their desktop / laptop sales... But whatever the reason, creating Atoms on older processes puts them at a self-inflicted disadvantage in the marketplace.
When Intel actually brings their A-game to the mobile market, I'll be watching.
Can we get a 2012 added to the title?
It was written two months ago about chips that are still the newest on the market!
Tegra 3 is anything but new. It's a 15 months old SoC. Besides, not enough info about Cortex A15.
Tegra 3 is over a year old. Not sure why MS picked it for their tablet considering its age, but its an ancient chip by mobile standards. The Transformer Prime was released in the US in Jan 2012. It had a tegra 3.
Here's the thing, it's not just about speed or efficiency, it's also about cost and price. Part of ARM's advantage is that chips are cheap and they have no incentive to support a high margin business that Intel does.
Fundamentally, Intel doesn't want to switch everything over to Atom chips because it would ruin the economics of Intel's business. Right now they would much rather sell a i5 at $200 or an i7 at $300+. Compare that to an Atom chip at say $25 per chip. Intel would need to have 10x volume to make the same money.
So, while Intel wants to be competitive with ARM, the economics of that business just aren't as attractive as the PC/Server chip business.
Fundamentally, Intel doesn't want to switch everything over to Atom chips because it would ruin the economics of Intel's business. Right now they would much rather sell a i5 at $200 or an i7 at $300+. Compare that to an Atom chip at say $25 per chip. Intel would need to have 10x volume to make the same money.
So, while Intel wants to be competitive with ARM, the economics of that business just aren't as attractive as the PC/Server chip business.
Selling chips for mobile phones doesn't cannibalizes intel's business. Selling chips for tablets might, but at this point in time it would be very risky to think in such a way, and they know this very well.
So why hasn't they given their top game for this fight, i.e. offering 22nm mobile/tablet chips ? I think it's because their 22nm is expensive and not very fit to manufacture cheap chips. From what i read, their 22nm tech costs 30%-40% more than using TSMC 28nm (for manufacturing the same parts). Maybe their rising investment in fabs is due to this problem.
Another explanation would be that they are planning a coordinated attack: better atom and an LTE chip using a 14nm process. That combination might let them seriously decrease power over competitors. And doing that in bang has strategic value of creating pressure/fear over mobile handset makers and carriers to accept their new design, or be left out.
So why hasn't they given their top game for this fight, i.e. offering 22nm mobile/tablet chips ? I think it's because their 22nm is expensive and not very fit to manufacture cheap chips. From what i read, their 22nm tech costs 30%-40% more than using TSMC 28nm (for manufacturing the same parts). Maybe their rising investment in fabs is due to this problem.
Another explanation would be that they are planning a coordinated attack: better atom and an LTE chip using a 14nm process. That combination might let them seriously decrease power over competitors. And doing that in bang has strategic value of creating pressure/fear over mobile handset makers and carriers to accept their new design, or be left out.
ARM is also moving forward with a big.little architecture that duplicates the CPU cores for either high-energy or low-energy situations.
Intel still has a way to go…