Does anyone think we'll see ARM replace x86 in desktops?

[Shivansps]

Microsoft created a new software ecosystem with Windows Store Apps, and now with Universal apps, x86 and ARM are about the same, clear show of this is Raspberry PI 2 W10 IoT, for example right now i can make a ARM game for W10 using Unity 3d targeting Windows 10 Universal apps... and i can compile it for both.

So software in a few years its not gona be a problem, but ARM will never overtake x86 in desktops. Just because ARM is gopod for your phone does not mean is good for desktop...

 

[Exophase]

Microsoft created a new software ecosystem with Windows Store Apps, and now with Universal apps, x86 and ARM are about the same, clear show of this is Raspberry PI 2 W10 IoT, for example right now i can make a ARM game for W10 using Unity 3d targeting Windows 10 Universal apps... and i can compile it for both.

So software in a few years its not gona be a problem, but ARM will never overtake x86 in desktops. Just because ARM is gopod for your phone does not mean is good for desktop...

Developers can do this.. but will all of them? Supporting Win 10 Universal is kind of a pain for a lot of developers (I know I'm not that thrilled about having to compile in Visual Studio), and no matter how much it might seem like something is portable serious commercial developers can't really afford to release a binary for a platform without properly testing for it. Little things that are undefined in the language can end up biting you, for example I've seen discrepancies due to different shift behavior going from x86 to ARM. And even though it's often spoken of as obsolete today there is still a non-negligible amount of hand-written assembly or intrinsics floating around, including in some popular middleware.

Fat binaries have been the standard on Android for about as long as the NDK was around, and x86 got first class support here before they really became viable in the market. Despite that, a few years later a lot of the most popular apps still lacked compile targets for x86.

 

[VirtualLarry]

The age of the basic consumer needing a desktop PC for Facebook and Office is over. Now if you need a desktop you probably are doing something on there that you plan to get paid for.

Don't tell that to my clients...

 

[poofyhairguy]

Don't tell that to my clients...

 

[Thala]

The real problem for productivity on mobile today isn't software, but is a lack of good mouse and windowing support. That will come soon though at least to one mobile platform, as Remix desktop is showing what can be done in Android.

Agreed. Windowing support is a must for serious desktop work. Therefore i am convinced that keeping Windows RT alive would have been of significant strategic importance for Microsoft to prevent that they are overtaken left and right by others, who will support a windowed UI for ARM.

 

[Headcool]

I don't think ARM will replace any x86 systems in future. ARM lost already market share in the tablet space, while it could not gain any market share in servers and laptops. Even in places where it is a possible option like chromebooks, ARM has literally no market share.

I think at the moment the A9X is with no doubt the best ARM-SoC available. According to the anandtech review of the IPad Pro in terms of CPU performance the Intel Core m-5y71 is ~105% faster than the A9X at the same power constraints (passive cooled tablet). And that is only with autovectorized code. If you compare heavily vectorized code like Linpack the A9X will be times behind a Core m. Also note the Core m-5y71 was introduced almost a year before the A9X. Keep in mind that the die size of the A9X is significantly higher than that of the Core m, which means the latter could be sold at a lower price.

However I don't think that is problematic for ARM in the next couple years, because they can continue to grow in the IoT space.

 

[ksec]

You can't really judge / tell simply by looking at ARM / x86 from a technical perspective. What the question should be, are there any financial incentive to switch from x86 to ARM.

First you need to know where the market are. Desktop / Laptop PC sales are shrinking YoY. And this is despite the fact the China's sales has a YoY growth. If not things will be much worse. The replacement cycle for Desktop/Laptop are now 3+ to 5 years. Majority of uses cases are now business / work oriented, which means compatibility is priority.

Seconds there is cost, the ARM CPU / SoC cost different to Intel Atom ( on same performance level ) are relatively small compared to the total Laptop. And if you want performance, this is still an Intel only territory.

What about Chromebook? you may asked, well i don't see how that is different to a tablet + keyboard.

Even Intel are now focusing on Server CPU. Where there is growth and very healthy margins.

So I think we may see Tablet replacing many Home PCs before ARM wins the / any desktop market.

* Note: Tablet usage is actually increasing, despite the slowing sales. Which means people are passing their tablet to for others to use rather then disposing them like Phones.

 

[TeknoBug]

It will take a lot for ARM to replace x86, the chances are pretty slim.

 

[cytg111]

Replace? Nah.
Compete? Oki, maybe.

 

[Nothingness]

I think at the moment the A9X is with no doubt the best ARM-SoC available. According to the anandtech review of the IPad Pro in terms of CPU performance the Intel Core m-5y71 is ~105% faster than the A9X at the same power constraints (passive cooled tablet).

How often must that be said? AnandTech review favored Intel by using a compiler that's known to be heavily optimized for SPEC. On top of that they targeted 32-bit code on Intel and 64-bit on Apple.

 

[videogames101]

How often must that be said? AnandTech review favored Intel by using a compiler that's known to be heavily optimized for SPEC. On top of that they targeted 32-bit code on Intel and 64-bit on Apple.

Compilers are part of the performance equation. You can't discount the impact of a strong compiler.

Also to everyone talking about Microsoft support, remember that lots of "serious" desktop work is also done on *nix and not just windows.

 

[Exophase]

Compilers are part of the performance equation. You can't discount the impact of a strong compiler.

This is not about ICC being a strong compiler, it's about ICC being a compiler that performs exotic optimizations that heavily affect little more than a few SPEC subtests. The AT review speculates that the huge difference in libquantum performance is due to it being "easily vectorizable" and ICC being much better at that than LLVM, but in reality it's due to optimizations that change memory layout and reorder memory accesses at a very high level in an effectively bandwidth limited test. Apple employees developing LLVM are now looking at performing similar "heroic" optimizations and they too admit they'll only really benefit libquantum.

I have code that actually is very vectorizable but far from easily so, and the latest version of ICC actually performs worse with with it than GCC 4.9. Both generate pretty poor quality code that I can get 3+x better performance with using hand written assembly (32-bit x86 w/SSSE3, which is basically the baseline for Android)

Plus I can't find any way to get ICC outside of the $699 + yearly licensing fee structure. They used to offer free builds under various circumstances but no longer. Then they offered an Android NDK version for something like $89. That was replaced with a beta into some new "INDE" thing but now all the links on the site, which are very convoluted, seem to go in circles or are outright broken. So for the time being it looks like the only real options are very expensive.

 

[Headcool]

How often must that be said? AnandTech review favored Intel by using a compiler that's known to be heavily optimized for SPEC. On top of that they targeted 32-bit code on Intel and 64-bit on Apple.

Anandtech used the Intel Compiler because they wanted to compare how good the SoCs are and not how bad some compilers are. In the end the used Intels compiler for the Intel SoCs and Apples compiler for the Apple SoCs.

I also think anandtech tried 32 & 64 bit for both architectures and took for each architecture the faster one. That is a valid option for a benchmark that doesn't come near the 4GB memory border.

In the end no one can sugarcoat the A9X. Compared to an Intel Core M this SoC is pure shit and that is exactly why the A9X is a ~50$ SoC, while the Core M is sold at 280$.

 

[Exophase]

Intel has really done a great job with their benchmarketing, it's really hard to explain to people why it makes the tests meaningless...

 

[Nothingness]

Intel has really done a great job with their benchmarketing, it's really hard to explain to people why it makes the tests meaningless...

Indeed... I prefer to think that's because these people are fanboys rather than plain stupid.

 

[videogames101]

Indeed... I prefer to think that's because these people are fanboys rather than plain stupid.

Blame the test not the compiler.

If you think Intel/AMD both don't include compiler improvements in their quoted perf increases per generation, you don't know jack.

 

[Nothingness]

Blame the test not the compiler.

If you think Intel/AMD both don't include compiler improvements in their quoted perf increases per generation, you don't know jack.

I don't blame the compiler per se, I blame Anandtech for not doing a fair comparison so that we can get a good picture of where A9X stands.

I'm afraid SPECv6 (that should be called SPEC CPU2016) is already being actively worked on by Intel.

I have used icc many times over the years, and never got any speedup. My last tests show that for SPEC 403.gcc subtest, which is arguably the most interesting SPEC test, icc loses against gcc. icc is useless for me, and misleading when comparing CPU using different ISA.

 

[Arachnotronic]

In the end no one can sugarcoat the A9X. Compared to an Intel Core M this SoC is pure shit and that is exactly why the A9X is a ~50$ SoC, while the Core M is sold at 280$.

I don't think so. A9X is in many ways more sophisticated than Core m.

 

[jhu]

Blame the test not the compiler.

If you think Intel/AMD both don't include compiler improvements in their quoted perf increases per generation, you don't know jack.

There's a problem when the compiler optimizes for benchmarks though. We found some anomalies back in 2013 with Antutu (which is in part based on nbench) on 32-bit x86 code. In terms of real world code, I've not seen any significant speed improvements over gcc in any programs I care about (they're statistically significant, but not actual real world significant).

 

[Shaun_Brannen]

Does anyone think we'll see ARM replace x86 in desktops?

Seems to be a bit too forward-looking of a question to have a reasonable answer. ARM needs to get in desktops in the first place... if it does overtake x86, it won't happen overnight.

 

[Exophase]

Indeed... I prefer to think that's because these people are fanboys rather than plain stupid.

I don't think it's that simple either. SPEC has been around a long time and has been used a lot in marketing and academic research so it has a strong reputation. And important, authoritative people continue this. Most of the other CPU benchmarks used in mobile reviews are in a lot of ways worse than SPEC, but most of them are also not broken by ICC like SPEC is. There was AnTuTu, possibly others like some of XPRT.

So given SPEC's reputation, including the prestige that's attached to its high price, and the people who push it's only natural that others are going to take the scores as a legitimate holistic reflection of processor performance.

While people might understand and accept the role that the compiler plays in the benchmark's success they'll think that it should be counted as a natural win for the platform anyway, even if it's extremely abnormal for the compiler to show those kinds of gains in other programs, even if far from everyone is using that compiler. It's really hard to demonstrate these things convincingly.

Blame the test not the compiler.

If you think Intel/AMD both don't include compiler improvements in their quoted perf increases per generation, you don't know jack.

Everyone knows better than to trust the CPU manufacturer's word alone when talking about performance. When it comes to websites independently running benchmarks that are widely called the best in the industry it's a different problem.

We can talk about whether or not it's fair to criticize Intel for putting the sorts of optimizations in ICC that they have but that's really not even the point. I don't care who blames whom, I care about people being better educated on what the compiler is doing and what the implications are for benchmarks comparisons. Unfortunately there's a lot of ugly technical background involved so it's hard to really convince people.

 

[ed21x]

i actually see Atom SOC's overtaking ARM in the mobile space given time. Intel still has tons of space to optimize x86 processors in to <4.5 watt TDP

 

[poofyhairguy]

i actually see Atom SOC's overtaking ARM in the mobile space given time. Intel still has tons of space to optimize x86 processors in to <4.5 watt TDP

Why would Apple put Intel in the iPhone over a chip they build?

 

[ShintaiDK]

Why would Apple put Intel in the iPhone over a chip they build?

If it could sell the next iPhone, sure

I dont doubt for a second that Apple would sacrifice its own semi division any day if it could sell more phones. We already know Samsung does it.

But at this stage its purely theoretical. Since Intel would have to make a product much better to change to it from ARM. And other ARM companies isn't likely to overtake Apple either in terms of replacement there. But make no mistake, the goal is to sell the phones, not the chips.

 

[jhu]

While people might understand and accept the role that the compiler plays in the benchmark's success they'll think that it should be counted as a natural win for the platform anyway, even if it's extremely abnormal for the compiler to show those kinds of gains in other programs, even if far from everyone is using that compiler. It's really hard to demonstrate these things convincingly.

For Itanium code at least, icc does generally generate better code than gcc.