Intel Silvermont Architecture: Does This Atom Change It All?

[ojas]

Man that was a load of info. Thanks!

But doesn't what you're saying in the end sort of validate my point (3) i made previously? That in Core you have the area and power envelop to do both SMT and OoOE?

In Atom i don't think they have the space or the power envelop to do both yet, without sacrificing a core or clocks...and if i'm not wrong, adding an entire core is much better for performance (threaded and non-threaded) than SMT.

Obviously would have a higher power penalty, but i guess would post better efficiency numbers.

Maybe stuff changes with Broadwell. Most likely will. Skylake onward i'm pretty sure both the Core and Atom processes will be similar, if not identical. Heck, this could even be the case post-Broadwell.

 

[Rub3z]

Just signed up to comment... all this is kind of new to me, but this site has given me very informative and valuable articles on specs and tips for my PC, even if I don't quite understand half of the terms used.

But anyway, I just wanted to ask... are any of these chips, at 45nm and smaller, even visible to the naked eye? How powerful a microscope would you need to see, much less construct, such an intricate mechanism? This just astounds me.

 

[4745454b]

Each transistor can't be seen. But the CPU as a whole should be visible. And remember that you can always look up any terms you don't understand. We tend to use a lot of acronyms and short hand so we can write faster. Hard for someone new to follow unless they know the terms. That's where google comes in.

 

[InvalidError]

But doesn't what you're saying in the end sort of validate my point (3) i made previously? That in Core you have the area and power envelop to do both SMT and OoOE?

Sort-of. But I was not trying to prove or disprove anything - other than HT is perfectly viable at least on an academic/engineering level. As I mentioned though, the largely single-threaded real-world disagrees.

Some OoOE is a tremendous advantage over none whatsoever on any workload while SMT is a tremendous advantage over none whatsoever on threaded workloads. Cost-effective gains from OoOE are finite but universal while gains from SMT have no intrinsic limit but are conditional to availability of threaded work to do and availability of execution resources - no point in having four hardware threads with ~3.5 potential IPC each (~14 potential instructions per cycle) when the execution back-end is only 5-ports wide and could be filled just about all of the time with only two threads.

The original Atom only had one execution unit and the new one has only two, which does not require a particularly fancy OoOE circuitry to keep busy most of the time. If you can fairly easily keep everything busy with a single thread then SMT provides very little additional benefit.

Things become more complicated starting at three ports where the CPU starts burning through code much faster than dependencies can be resolved. That's why I think HT will come back once Atom hits quad-port and ends up almost always having at least one unused port on every cycle that could be leveraged by HT for a ~30% gain on threaded workloads.

As for Atom merging with normal parts, I doubt it. You do not want to bring Haswell's huge 192 instructions 6-ports OoOE architecture with similarly super-sized support structures to a line entirely dedicated to ultra-low-power SoC products. Standard desktop Haswell likely burns more power in its OoOE-related structures than a whole Atom SoC including IGP at full-load.

 

[timon_tablet]

If you wanted a really creditable Atom tablet, need to wait for Bay Trail Slides, will support SATA 2.0 (not 3.0) and USB 3.0. Also, 8 or 9 inches Windows 8 tablet would come soon. (Clover Trail Atom z2760 does not support SATA SSD and USB 3.0, but is a very bad eMMC SSD and USB 2.0 only)

Truly, Mr. Paul Otellini ought to retire. Up to now, the Windows 8 is not adequately active in tablet market, therein the problem in half actually showed from Intel Atom z2760. If with eMMC SSD, why we get to buy the expensive Atom z2760 Windows 8? iPad is cheaper with eMMC.

Even if the eMMC SSD Windows 8 Atom to drop down the price will be still inconclusive, the Android tablets are always with cheaper price and flooding, but the Windows x86 tablet's mainly advantageous conditions are not the lowest-end market, so, Atom in x86 Windows must support SATA SSD with better performance.

Kodak corp have ever had a lot of advanced digital image technologies, however Kodak only wanted to keep their film profit margin during faced to the raging waves of digital image market, sorry, people have seen Kodak finally speeded up died at himself. Today, could it be Kodak ghost would replay to Intel?

If others who - like AMD may hit in x86 mobile computing market, well then, Intel will have to do Atom to provide SATA 3.0 (no longer 2.0) and USB 3.0 with better performance. Regrettable, AMD is now very stagnant. In the current x86 mobile processor industry the competition is too lacking.

Recommend read:
(website tomshardware)
Samsung's ATIV Smart PC 500T: An Atom-Based Windows 8 Tablet
PCMark 7: A Look At Storage Performance
Benchmark Results: Photoshop, iTunes, And WinRAR

(website hothardware)
Intel Clover Trail Atom Z2760 Tablet Performance Preview
SunSpider, BrowserMark and PCMark 7

 

[portentous]

Great review. I am wondering if MS is considering dropping one of these into the Surface running Windows8 Pro instead of the RT. This will be very interesting.

 

[ceh4702]

How does it compare to an i-3 doing real tasks?
Play 1080p Flash Video smoothly while looking at your email and a web page?

 

[InvalidError]

How does it compare to an i-3 doing real tasks?

60% fewer execution units, likely close to half the clock, no HT... so that's about 1/6th of desktop i3's raw processing power but about double the previous Atom's single-threaded potential.

It is a nice step forward for Atom but it won't be replacing mainstream CPUs for things that require more heavy-lifting - at least not without GPGPU/hardware assist.

 

[jkflipflop98]

[citation][nom]Rub3z[/nom]Just signed up to comment... all this is kind of new to me, but this site has given me very informative and valuable articles on specs and tips for my PC, even if I don't quite understand half of the terms used.But anyway, I just wanted to ask... are any of these chips, at 45nm and smaller, even visible to the naked eye? How powerful a microscope would you need to see, much less construct, such an intricate mechanism? This just astounds me.[/citation]

We use a variety of optical and electron microscopes. At 20x magnification you can see the larger structures on the device like the markers used for lithography registration. At 100x you can begin to see the circuitry on the die. It takes 10,000x magnification to see the individual structures. We have equipment such as tunneling electron microscopes that can resolve down to the atomic level.

At that point, a wafer just appears to be a sea of vibrating beachballs.

 

[frenchi]

Time for Intel to purchase HTC. Imagine a HTC one with an Intel Inside.....with Verizon....

 

[ojas]

But doesn't what you're saying in the end sort of validate my point (3) i made previously? That in Core you have the area and power envelop to do both SMT and OoOE?

Sort-of. But I was not trying to prove or disprove anything - other than HT is perfectly viable at least on an academic/engineering level. As I mentioned though, the largely single-threaded real-world disagrees.

Some OoOE is a tremendous advantage over none whatsoever on any workload while SMT is a tremendous advantage over none whatsoever on threaded workloads. Cost-effective gains from OoOE are finite but universal while gains from SMT have no intrinsic limit but are conditional to availability of threaded work to do and availability of execution resources - no point in having four hardware threads with ~3.5 potential IPC each (~14 potential instructions per cycle) when the execution back-end is only 5-ports wide and could be filled just about all of the time with only two threads.

The original Atom only had one execution unit and the new one has only two, which does not require a particularly fancy OoOE circuitry to keep busy most of the time. If you can fairly easily keep everything busy with a single thread then SMT provides very little additional benefit.

Things become more complicated starting at three ports where the CPU starts burning through code much faster than dependencies can be resolved. That's why I think HT will come back once Atom hits quad-port and ends up almost always having at least one unused port on every cycle that could be leveraged by HT for a ~30% gain on threaded workloads.

As for Atom merging with normal parts, I doubt it. You do not want to bring Haswell's huge 192 instructions 6-ports OoOE architecture with similarly super-sized support structures to a line entirely dedicated to ultra-low-power SoC products. Standard desktop Haswell likely burns more power in its OoOE-related structures than a whole Atom SoC including IGP at full-load.

Read this today, great read, thanks!

Cleared stuff up!