Microsoft Windows 10 on ARM

[Exist50]

Regarding K12 as some have mentioned, I don't think this has much relevance for it. For better or worse, AMD seem to have scrapped their K12 plans for the near future, or are at least keeping it on the back burner.

 

[knutinh]

There are also many fine points in the ISA that makes simulation either slow or inaccurate (e.g. FP NaN).

Quirks like that are allready difficult on one platform (e.g. Windows on x86 using a single compiler). Even switching to Linux (using the same hardware), default memory philosophy could expose nasty assumptions made by the programmers on the basis of "works for me". An emulator that lacks the context of a developer would have a really hard time doing the "right" tradeoff between accuracy and speed in such cases.

What if (say) Intel does SIMD floating-point very efficiently for rounding, while ARM does it very efficiently for truncating? For many applications, the difference might be irrelevant, but for some it might be the difference of a crash or not (branching on floating point values is dangerous).

My impression is that the ARM instruction set it more functionally "complete" than Intel (although occasionally performance-wise nonsensical). So perhaps it is easier to emulate a restricted instruction set by a less-restricted instruction set than the other way around?

I don't know how low-level typical win32 applications are (heck, I don't know how a linker works). If the win32 API offers high-level abstractions and developers rely heavily on common libraries, perhaps the problem is simpler than suggested above. Rewrite low-level Microsoft libraries, some common dependency libraries, and many/most applications will spend most of its time in native ARM code anyways, and those who chose to do their own low-level coding will suffer, but those are a minority?

-k

 

[deasd]

Regarding K12 as some have mentioned, I don't think this has much relevance for it. For better or worse, AMD seem to have scrapped their K12 plans for the near future, or are at least keeping it on the back burner.

As far as I know, K12 is very much similar to Zen and developed in parallel, so I don't agree 'scrap' or 'dead' statement.

If Microsoft doesn't get ready, there's nothing much that AMD or other manufacturers like Samsung could do. Reason why MS choose Snapdragon first I think it's just because they already cooperated in Windows Phone.

 

[dark zero]

Geekbench 4 results....

(the Q6600 doesn't have AES)

Q6600 (LGA 775 C2Q):
https://browser.primatelabs.com/v4/cpu/1273759
(ST 1497, MT 4125)

Snapdragon 820 @ 1.59 Ghz (as MSM 8996):
https://browser.primatelabs.com/v4/cpu/1284320
(ST 1681, MT 3974)

Snapdragon 820 @ 2.19 Ghz (as MSM 8996):
https://browser.primatelabs.com/v4/cpu/1283841
(ST 1822, MT 4096)

J3455 (Goldmont atom quad core):
https://browser.primatelabs.com/v4/cpu/1279939
(ST 1495, MT 4068)

And considering that the old Kryo is replaced by Kryo 200 next Q2 2017...the situation would be totally dramatic.

 

[Lodix]

And considering that the old Kryo is replaced by Kryo 200 next Q2 2017...the situation would be totally dramatic.

Forget about kryo, stock ARM Cortex A72/73 are much better not just in performance but in power consumption. This Atom are in a 10W TDP and can't compete with a 3.5W full loaded CPU like Kirin 950/960 with a disadvantage in process node.

 

[Exophase]

Wouldn't you have to explicitly change symbols resolution in the linker to prevent interworking from properly functioning? IIRC symbols get their lower bit set to 1 for T2 code while ARM functions have it set to 0. It's been a long time since I did any 32-bit ARM dev

They don't do something like that to outright prevent it from working, they just say that the kernel won't necessarily properly handle state transitions from ARM mode and therefore you could get crashes on syscalls or interrupts. But from all testing I've seen this wouldn't appear to be an actual problem. It may be that failing to handle it properly caused too much of a headache.

 

[Face2Face]

I'd be interested to see how this emulation handles older graphic's APIs (DX9 & DX10). Would be pretty cool to have a SD 820/821 stick or box and fire up Rocket League at a friends house on the cheap. The Adreno 530 is nearly 3 time faster than Intel's HD 400 (Braswell) in GFX Bench 3.0.

 

[dark zero]

Forget about kryo, stock ARM Cortex A72/73 are much better not just in performance but in power consumption. This Atom are in a 10W TDP and can't compete with a 3.5W full loaded CPU like Kirin 950/960 with a disadvantage in process node.

Don't forget the Snapdragon 650 and 652... despite they are at 28 nm, the 2 cores at 1.8 are enough strong to compete against the Core 2 Duo E6600.

Here is from the Xiaomi Redmi Note 3 Pro (SD 650)
http://browser.primatelabs.com/v4/cpu/222924

And here is from the Core 2 Duo E6700 at 2.66 Ghz
http://browser.primatelabs.com/v4/cpu/1216571

Literally at the same level. Even Windows 10 could run fluently on those chips. And those 4 remaining Pentium 4 tier chips, are still useful for the most basic tasks.

I'd be interested to see how this emulation handles older graphic's APIs (DX9 & DX10). Would be pretty cool to have a SD 820/821 stick or box and fire up Rocket League at a friends house on the cheap. The Adreno 530 is nearly 3 time faster than Intel's HD 400 (Braswell) in GFX Bench 3.0.

Even Adreno 510 is some (%) better than ther HD 400, enough to some advanced emulation. So it could run LOL at 30 fps on HD on low resolution.

 

[ashetos]

How do they make peripherals work when all the drivers are x86?

 

[willfr]

This is good news because Intel has what, 85%-90% of the PC market? I like Intel but that's not a good thing. We need this competition. And didn't Intel kill Atom for Windows? These ARM chips will fill that gap nicely (I'm always surprised to see how well Atom chips sold, so ARM Windows machines could sell very well)

 

[lefty2]

I'm guessing that the applications shown in the video are "pre-cached". In other words, the application has already been run once, which allows the x86 to arm code to be translated and stored in the cache, the second time it's run the application runs much faster and that's how it looks like everything is running very snappy on the snapdragon (excuse the pun).

 

[deasd]

Don't forget the Snapdragon 650 and 652... despite they are at 28 nm, the 2 cores at 1.8 are enough strong to compete against the Core 2 Duo E6600.

Here is from the Xiaomi Redmi Note 3 Pro (SD 650)
http://browser.primatelabs.com/v4/cpu/222924
And here is from the Core 2 Duo E6700 at 2.66 Ghz
http://browser.primatelabs.com/v4/cpu/1216571

Literally at the same level. Even Windows 10 could run fluently on those chips. And those 4 remaining Pentium 4 tier chips, are still useful for the most basic tasks.

Well..... I don't know how to start this topic, I still cannot figure out how this tiny chip can catch up with Sandybridge in Geekbench. From what I know the 2.3Ghz A72&73 still feel terrible sluggish when meet flash on the web.
What I doubt is that newest GB has absolutely no optimization for old hardware, and yes I agree ARM has quite large leaping room these years, but still something looks too incredible especially when A72 come out, almost nothing change in the architecture but the GB result skyrocketing.

I think the A72 is still similar to A15 no matter what GB result they could give.

As for GB, I'm considering to ditch this weird bench from my reference of smartphone/tablet/or any other equipment purchase, until I can understand what mechanism this bench hold.

 

[Nothingness]

I'm guessing that the applications shown in the video are "pre-cached". In other words, the application has already been run once, which allows the x86 to arm code to be translated and stored in the cache, the second time it's run the application runs much faster and that's how it looks like everything is running very snappy on the snapdragon (excuse the pun).

That's not how dynamic translation works: each block of instructions is translated when run for the first time (or after having been run several times if you're doing translation based on profiling), so you don't cache all of a program by running it once.

 

[Nothingness]

What I doubt is that newest GB has absolutely no optimization for old hardware, and yes I agree ARM has quite large leaping room these years, but still something looks too incredible especially when A72 come out, almost nothing change in the architecture but the GB result skyrocketing.

I think the A72 is still similar to A15 no matter what GB result they could give.

You have to be kidding. A15 and A72 are vastly different. This can be seen on other benches than Geekbench (since you don't trust it). Or perhaps there's some conspiracy of all sites and benchmarks to show A72 much better. Yes that must be coming from the Chinese CIA.

 

[deasd]

I'm guessing that the applications shown in the video are "pre-cached". In other words, the application has already been run once, which allows the x86 to arm code to be translated and stored in the cache, the second time it's run the application runs much faster and that's how it looks like everything is running very snappy on the snapdragon (excuse the pun).

Yes, A15 and A72 are different, from 3 decode 8 dispatch to 3 decode 5 dispatch, it's become even narrower. A73 is even a 2 decode design. I believe ARM just cutting down the core to reach higher efficiency/clock, or another point of view, they realized wide core like A15 doesn't make sense in low power design.
People believe A72 could reach Sandybridge performance, I'll let it go, I won't intend to change people's mind until I find more evidence.
What's bad is there's very few trustworthy bench on the web that work cross-platform except geekbench, if you have some please let me know.

 

[witeken]

nvm

 

[dbcoopernz]

Apparently someone has compiled 7-Zip and PuTTY for ARM64 Win 10:

https://mspoweruser.com/windows-arm64-gets-first-compiled-apps/

 

[dark zero]

Yes, A15 and A72 are different, from 3 decode 8 dispatch to 3 decode 5 dispatch, it's become even narrower. A73 is even a 2 decode design. I believe ARM just cutting down the core to reach higher efficiency/clock, or another point of view, they realized wide core like A15 doesn't make sense in low power design.
People believe A72 could reach Sandybridge performance, I'll let it go, I won't intend to change people's mind until I find more evidence.
What's bad is there's very few trustworthy bench on the web that work cross-platform except geekbench, if you have some please let me know.

Not sure if you are serious or you are trolling.

We never expect to see SB performance in at least 2 or 3 generations more, but now the A72 are on Core Duo performance (the 65 nm one), and Atom recently managed to get that performance too, but at higher power cost.

Even more I see the current chips on the following tier considering the same clock rate:
- A7 = Pentium 3
- A15 = Pentium 4
- A53 = Pentium D (low clock only)
- A57 = Netburst Pentium D (due overheating)
- A72 = Core Duo (Yonah 65nm)
- A73 = Core 2 Duo/Quad (45nm)

And Apple A10 Fusion are near Sandy Bridge tier right now.

So the only chip that could run Windows 10 without problems besides the Kryo could be the ones who has at least 2 ARM A72/73 Cores, like the Snapdragon 650, MTK Helio X20 or the Kirin 950.

 

[deasd]

Not sure if you are serious or you are trolling.

We never expect to see SB performance in at least 2 or 3 generations more, but now the A72 are on Core Duo performance (the 65 nm one), and Atom recently managed to get that performance too, but at higher power cost.

Even more I see the current chips on the following tier considering the same clock rate:
- A7 = Pentium 3
- A15 = Pentium 4
- A53 = Pentium D (low clock only)
- A57 = Netburst Pentium D (due overheating)
- A72 = Core Duo (Yonah 65nm)
- A73 = Core 2 Duo/Quad (45nm)

Not troll. If you look into GB database you'll find MSM8996 already match i3-2310m in overall result:

http://browser.primatelabs.com/v4/cpu/search?utf8=✓&q=i3+2310m
http://browser.primatelabs.com/v4/cpu/search?utf8=✓&q=MSM8996

GB result gives a much weird and weird impression to me when time goes by, that's why I tried hard to look for more reliable benchmark or any other good test approach to compare between multiple platform.

for additional: Apple A10 already outperform i3-6100u here:
http://browser.primatelabs.com/v4/cpu/search?utf8=✓&q=i3+6100u
http://browser.primatelabs.com/v4/cpu/search?utf8=✓&q=iphone+7

 

[William Gaatjes]

It could be that the PC becomes the console in the future.

https://www.engadget.com/2016/08/17/microsoft-aaron-greenberg-qa-project-scorpio-vr/



In fact, Microsoft is already Merging Windows 10 with Xbox One to some degree already (example below):


http://www.xbox.com/en-US/windows-10

Well, that is what i had in mind some time ago. A console that can be used as a real pc.
I think what they really mean is that there will be no more "game" console.
When(or if) the hbm2 based apu's soc become available and available for a reasonable price, it is very likely, that Microsoft would just ask a large motherboard designer to create a motherboard with specifications that Microsoft finds interesting. On the other hand, it may be in the near future that computation power needed cannot be performed by a single chip simply because of the amount of power dissipation, way too much thermal throtteling. The chip would be impossible to cool with a low cost cooling solution and use to its full potential at the same time.
That is something that i am curious about. Because with that in mind, we would continue to have the gpu and cpu as separate entities if we want the absolute highest amount of calculation power. And that would indeed mean that Microsoft may try for a pc alike solution. But i would not be surprised if it would be kind of like the home theater pc we have now. Only way more slick designed.

Then again on a side note, i remember cooling solutions based on industrial created nanoscale sized diamonds. A very thin diamond layer on the chip would help to get the heat out of the chip to an external closed (water or oil) cooler. But to get the nanolayer of diamonds on the silicon is still a problem. The process required can damage the silicon although progress has been made. Who knows what thee future will bring.

 

[dark zero]

Not troll. If you look into GB database you'll find MSM8996 already match i3-2310m in overall result:

http://browser.primatelabs.com/v4/cpu/search?utf8=✓&q=i3+2310m
http://browser.primatelabs.com/v4/cpu/search?utf8=✓&q=MSM8996

GB result gives a much weird and weird impression to me when time goes by, that's why I tried hard to look for more reliable benchmark or any other good test approach to compare between multiple platform.

for additional: Apple A10 already outperform i3-6100u here:
http://browser.primatelabs.com/v4/cpu/search?utf8=✓&q=i3+6100u
http://browser.primatelabs.com/v4/cpu/search?utf8=✓&q=iphone+7

Other benchmark which are trusting is Antutu and is pretty much the same result.. ARM is catching x86 up. And big time.