The Intel Atom Thread

[Sweepr]

Talking about modems...

https://benchlife.info/intel-xmm-7480-lte-modem-announce-with-450mbps-download-02232016

 

[dark zero]

The big problem will be if they find an OS to support it besides Linux.

 

[Sweepr]

AnandTech reviewed the Voyo V3, Cherry Trail-based (Atom x7-Z8700) mini-PC, available for $215 (as configured).

AnandTech: Voyo V3 Review - A Fanless Intel Atom x7-Z8700 (Cherry Trail) mini-PC

Intel transitioned their Bay Trail-T Atom lineup (targeting affordable 2-in-1s, tablets and Compute Stick form factors) to 14nm with the introduction of Cherry Trail-T. The Atom x5 and x7 SoCs coming under this family have four Airmont cores and Broadwell-class Intel HD Graphics. We have already seen the x7-Z8700 in action in the Microsoft Surface 3 and the x5-Z8300 in the Cherry Trail Compute Stick. Due to the success of UCFF (ultra-compact form factor) PCs, many vendors (including no-name Asian brands) have resorted to making small computers by using these tablet platforms with minor modifications. One such vendor is Voyo, and their V3 mini-PC is a unique take on the Atom x7-Z8700 platform compared to traditional tablets / affordable 2-in-1s.

Traditional UCFF PCs have stayed true to Intel's reference designs for such PCs. In particular, the NUC, Mini-Lake and Compute Stick reference platforms have enabled vendors to quickly bring their own variants into the market. Obviously, vendors such as ASRock and Zotac do have custom boards, but, they are usually few in number compared to the number of UCFF PCs in the current market. Since the introduction of Bay Trail-T, we have seen a rise in the number of systems based on reference boards for tablets / 2-in-1s. Most of them just put a case around such a board (with soldered DRAM, eMMC storage etc.) and market it as a PC.

In most cases, we ignore review requests for these types of PCs - after all, they have nothing unqiue to offer and are held back by the abysmal eMMC storage sub-system and skimpy DRAM. So, when GearBest offered to send us a review sample of the Voyo V3 mini-PC equipped with the high-end Atom x7-Z8700, 4GB of RAM and a bonafide 128GB M.2 SSD, we were mildly interested. Some of the advertised aspects such as Windows 10 being pre-installed, USB Type-C support, '5G Wi-Fi' etc. seemed too good to be true for the price, but, we decided it was interesting enough to put through our rigorous test process for low power computing systems.

- Performance & Power Consumption

The fanless Voyo V3, with its top-end Cherry Trail-T SoC, 128 GB M.2 SATA SSD and 4GB of RAM, outperforms passively cooled Braswell-based mini-PCs in almost all benchmarks
The M.2 SATA SSD is miles ahead of eMMC-based storage subsystems found in most other Cherry Trail PCs in this price range. The RAM capacity is also quite large for this price range.
The industrial design of the unit, with its curved edges, metal chassis and glass top, is quite pleasing
The unit is perfect for a secondary 1080p HTPC where HD audio bitstreaming is not a must - Kodi 16.0 works flawless with hardware video decoding for all major codecs

www.anandtech.com/show/10100/voyo-v3-review-a-fanless-intel-atom-x7z8700-cherry-trail-minipc


Overall one of the best Cherry Trail designs yet, faster / lower power consumption than Bay Trail and even Braswell competitors.

 

[VirtualLarry]

Speaking of Cherry Trail mini-PCs, I've got a "Wintel W8 Pro" coming in, with Win10 (supposedly 64-bit according to their specs, but the images on their site show 32-bit), 2GB / 32GB. Nothing major, basically kind of like a compute stick (I think it's a quad-core, though), but with more bulk, more I/Os, and a similar price. It was $81 + ship.

 

[TeknoBug]

Speaking of Cherry Trail mini-PCs, I've got a "Wintel W8 Pro" coming in, with Win10 (supposedly 64-bit according to their specs, but the images on their site show 32-bit), 2GB / 32GB. Nothing major, basically kind of like a compute stick (I think it's a quad-core, though), but with more bulk, more I/Os, and a similar price. It was $81 + ship.

Yeah I seen those, they look like the Roku box and yes 32bit because of the 2GB ram/32GB storage and Windows is free on devices like those. Good price you got that for, they're $150 here.

 

[IntelUser2000]

At last I found a reason for why a 16EU Gen 8 is barely 2x fast as a 4EU Gen 7 unit.

As soon as the GPU starts getting loaded, it moves between 200 - 300 MHz while keeping the power consumption of the whole system around 10 W.

So, I went back to Notebookcheck and started looking at reviews for how well clocks are kept.

Here's the conclusion:
-Bay Trail's 4EU iGPU can clock between 600-800MHz in testing. The devices can reach top speed of 800MHz fairly easily
-Cherry Trail's 16EU iGPU stands between 350-400MHz. It's almost half of Bay Trail, and even far off from the max Turbo clock of 600MHz

I wonder if that's because Intel decided that the 14nm yields are too low and they needed a big GPU but clocking low to keep costs down? Maybe lack of perf/watt for Gen 8 part had to do with it as well.

Looking at the future, 18EU for Goldmont isn't a worry if realistic frequency can reach at least 600MHz. +50% clock and +30% performance that Gen 9 brought over Gen 8 would be enough for Goldmont to reach 2x the performance of Cherry Trail graphics. If they can clock it at Bay Trail's frequency, Goldmont can be potentially be 2.5x the performance. The ARM camp has iGPUs that won't be caught up even with that level of performance but it will be able to reach mid-end.

 

[dark zero]

Mmm.. Goldmont might be real deal if done properly... They might finally reach Celeron Core levels?

 

[IntelUser2000]

Mmm.. Goldmont might be real deal if done properly... They might finally reach Celeron Core levels?

In order for Goldmont to be class leading compared to ARM:
2.5x GPU
5x GPU

That won't happen. Meaning no design wins and continuing poor sales in the Android camp.

Too much performance means they start killing Core chips. So for Intel the only logical choice is to make it not competitive in the Android-ARM camp. They continue to sell Core chips, but Atom will still be a big advancement meaning they sell value desktops.

Remember they "promised" 2x performance of Core i3 over Atom? That means artificial crippling. If we want to see that kind of gap but Atoms being class leading against ARM parts that means a 15W U part needs Iris Pro(read: Pro) performance in a HD part not Iris.

Maybe a CPU with 50% higher IPC on Core.

That won't happen.

 

[Abwx]

I wonder if that's because Intel decided that the 14nm yields are too low and they needed a big GPU but clocking low to keep costs down? Maybe lack of perf/watt for Gen 8 part had to do with it as well.

With 16U they can double the perfs while doubling the perf/watt simultaneously, and it was indeed badly needed since the GPU consume about as much as a Beema GPU (about 4W) and is still 20% less performing.

 

[Arachnotronic]

At last I found a reason for why a 16EU Gen 8 is barely 2x fast as a 4EU Gen 7 unit.

So, I went back to Notebookcheck and started looking at reviews for how well clocks are kept.

Here's the conclusion:
-Bay Trail's 4EU iGPU can clock between 600-800MHz in testing. The devices can reach top speed of 800MHz fairly easily
-Cherry Trail's 16EU iGPU stands between 350-400MHz. It's almost half of Bay Trail, and even far off from the max Turbo clock of 600MHz

I wonder if that's because Intel decided that the 14nm yields are too low and they needed a big GPU but clocking low to keep costs down? Maybe lack of perf/watt for Gen 8 part had to do with it as well.

Looking at the future, 18EU for Goldmont isn't a worry if realistic frequency can reach at least 600MHz. +50% clock and +30% performance that Gen 9 brought over Gen 8 would be enough for Goldmont to reach 2x the performance of Cherry Trail graphics. If they can clock it at Bay Trail's frequency, Goldmont can be potentially be 2.5x the performance. The ARM camp has iGPUs that won't be caught up even with that level of performance but it will be able to reach mid-end.

I dunno man, +50% clock seems like a stretch to me. The ARM guys have Intel beaten in low-power GPUs, plain and simple.

 

[NTMBK]

I dunno man, +50% clock seems like a stretch to me. The ARM guys have Intel beaten in low-power GPUs, plain and simple.

If only Intel could just pick up the phone and licence some Mali cores...

 

[Arachnotronic]

If only Intel could just pick up the phone and licence some Mali cores...

They did for SoFIA 3G-R and SoFIA LTE, lol.

These guys need to just license ARM CPU & GPU IP and build some chips in their fabs, if those fabs are as great for mobile chips as they claim. If those fabs aren't as great, then maybe somebody at Intel needs to have the nerve to decide to build their chips on leading-edge TSMC process.

Or they could stop wasting everybody's time and just exit the market.

 

[Shivansps]

We really need another ARM core builder?

 

[VirtualLarry]

Remember they "promised" 2x performance of Core i3 over Atom? That means artificial crippling.

Maybe a CPU with 50% higher IPC on Core.

That won't happen.

Too bad Intel painted themselves into a corner like that in mobile, with the arbitrary and artificial performance split between Core and Atom.

Sure, perhaps it's good for margins in the short term... but if it (that policy of performance separation) costs them entire markets?

Face facts, Intel, ARM CPUs are cheap as chips, so to speak. If Intel wants to compete with ARM CPUs and SoCs on their own turf, they're going to have to change their cost structure to be able to lower gross margins and still be able to compete.

 

[Face2Face]

At last I found a reason for why a 16EU Gen 8 is barely 2x fast as a 4EU Gen 7 unit.

So, I went back to Notebookcheck and started looking at reviews for how well clocks are kept.

Here's the conclusion:
-Bay Trail's 4EU iGPU can clock between 600-800MHz in testing. The devices can reach top speed of 800MHz fairly easily

-Cherry Trail's 16EU iGPU stands between 350-400MHz. It's almost half of Bay Trail, and even far off from the max Turbo clock of 600MHz

I wonder if that's because Intel decided that the 14nm yields are too low and they needed a big GPU but clocking low to keep costs down? Maybe lack of perf/watt for Gen 8 part had to do with it as well.

Looking at the future, 18EU for Goldmont isn't a worry if realistic frequency can reach at least 600MHz. +50% clock and +30% performance that Gen 9 brought over Gen 8 would be enough for Goldmont to reach 2x the performance of Cherry Trail graphics. If they can clock it at Bay Trail's frequency, Goldmont can be potentially be 2.5x the performance. The ARM camp has iGPUs that won't be caught up even with that level of performance but it will be able to reach mid-end.

I've tested a lot of games with many different Bay Trail devices and I've never seen the GPU boost beyond 620MHz in any of them. Even in games where the CPU usage was extremely low, so power limiting didn't come into play. The LP GEN 8 16EU GPU can be two times as fast as the LP GEN 7 4EU GPU In Bay Trail; however, the performance with not be sustained due to thermal throttling. I've tested this with my Surface 3.

If you watch this video I made, you'll see when starting out the CPU and GPU clocks are nearly at their maximum clock speeds. As the temps start to climb the clocks start to drop off, to the point where the CPU sits at it's base clock and the GPU hovers around 400Mhz. I would think if the surface 3 had active cooling it could then boost the GPU and CPU for a much longer period of time; of course, power limiting does come into play as well.

If you take a game like Crysis 3 , you'll see the GPU maintain at 400MHz, but the CPU clock speeds end up dropping below it's base clock due to thermal throttling and potentially power limiting.

Also keep in mind the GPU is a lot bigger in Cherry Trail, so as it heats up, so does the entire SoC. Never mind the fact that Cherry Trail CPUs have less surface area than Bay Trail, so extracting heat from the CPU is a tougher task.

 

[dark zero]

Too bad Intel painted themselves into a corner like that in mobile, with the arbitrary and artificial performance split between Core and Atom.

Sure, perhaps it's good for margins in the short term... but if it (that policy of performance separation) costs them entire markets?

Face facts, Intel, ARM CPUs are cheap as chips, so to speak. If Intel wants to compete with ARM CPUs and SoCs on their own turf, they're going to have to change their cost structure to be able to lower gross margins and still be able to compete.

Another fact is that Intel is not situable for Android after those years. Yeah, they better go with Windows Mobile 10 since it sucks hard on ARM.

 

[IntelUser2000]

I've tested a lot of games with many different Bay Trail devices and I've never seen the GPU boost beyond 620MHz in any of them.

600MHz is still way higher than what any Cherry Trail devices can do. That's a 50% clock speed difference.

In one of Intel investor slides they show a Broxton die as being smaller than Bay Trail, probably about the 80mm2 range that Cherry Trail is: https://lazure2.files.wordpress.com/2014/01/image20.png

What might have happened:
1. Cherry Trail was originally 8EU, but they got it quickly that it won't perform enough because 14nm would have yield issues. Then they doubled it to 16EU but low clocks
2. They knew from the beginning that Cherry Trail was nearly a straight shrink of BT without aligning the chip for 14nm's characteristics. That means the gains are entire dependent on the process(an ever diminishing one)
3. Broxton is similar in size because a significant increase is due to clock increase, not area investment

Sure, perhaps it's good for margins in the short term... but if it (that policy of performance separation) costs them entire markets?

Sure, they can sacrifice their ENTIRE Core mobile and make $20-30 Atom CPUs quite performant. Here's the issue: That does not guarantee a marketshare gain. In order for a loss of Core to be worth it, they would need 70%+ market of Smartphone/Tablet CPUs. That won't happen. So the way they are doing it now might have been a safer one.

I dunno man, +50% clock seems like a stretch to me. The ARM guys have Intel beaten in low-power GPUs, plain and simple.

That doesn't excuse Intel's poor execution. I am just suggesting why Goldmont can be more potent than we speculate, even though we all know it won't be top of the line anyway. 50% clock is possible if the yield issue with 14nm is really that bad, and I think its not unreasonable considering we are talking bottom of the barrel frequencies(for Intel).

 

[frozentundra123456]

If they have yield issues at 14 nm (a given initially, but dont know now), but anyway, why would they double the size of the igp? Wouldnt that make the yield issues *worse*?

 

[Exophase]

Here's the conclusion:
-Bay Trail's 4EU iGPU can clock between 600-800MHz in testing. The devices can reach top speed of 800MHz fairly easily
-Cherry Trail's 16EU iGPU stands between 350-400MHz. It's almost half of Bay Trail, and even far off from the max Turbo clock of 600MHz

I wonder if that's because Intel decided that the 14nm yields are too low and they needed a big GPU but clocking low to keep costs down? Maybe lack of perf/watt for Gen 8 part had to do with it as well.

Accomplishing the same amount of work with more functional units at a lower clock speed yields better efficiency.. to an extent. So long as you can lower the voltage and don't pass the tipping point vs the increased leakage.

It has the other bonus of allowing higher scaling in less TDP constrained platforms like set-top units with fans. That and doing well with short run benchmarks.

We've seen the same trend with most other mobile SoCs (at least the higher end ones) and if anything Intel has been very conservative with GPU die area on their mobile parts. It's a good use of the extra transistors enabled by 14nm. The 2x performance improvement they managed to get is actually a pretty big deal, and something they'd need both this big increase in area/lowered clocks plus the improvement per transistor brought by 14nm. I don't think this is a sign of process or design weakness at all.

Looking at the future, 18EU for Goldmont isn't a worry if realistic frequency can reach at least 600MHz. +50% clock and +30% performance that Gen 9 brought over Gen 8 would be enough for Goldmont to reach 2x the performance of Cherry Trail graphics. If they can clock it at Bay Trail's frequency, Goldmont can be potentially be 2.5x the performance. The ARM camp has iGPUs that won't be caught up even with that level of performance but it will be able to reach mid-end.

I doubt they're going to be able to increase both functional units and clock by that huge amount while still on 14nm, not for the same power consumption anyway.

 

[IntelUser2000]

If they have yield issues at 14 nm (a given initially, but dont know now), but anyway, why would they double the size of the igp? Wouldnt that make the yield issues *worse*?

Not necessarily. It depends on how the schmoo plot of the device turns out.

A shmoo plot: http://www.realworldtech.com/includes/images/articles/isscc06-tulsa-fig-5.gif

Shmoo plot is basically a relation of frequency versus voltage and whether a device operates at those parameters or not. So at high enough frequency and low enough voltage it stops working. While a combination of uarch + process might not have defects at an "average frequency" it might have a high failure at high enough frequencies.

That's why over time you get slightly faster CPUs and GPUs. Like for example the 7850K/7870K/7890K. You might remember that originally CT was promised at 2.7GHz. It came out at 2.4. And recently they released X7-8750 chip with 2.56GHz frequency.

You aren't talking about big dies like a 300mm2 -E chip and trying to double that. Going from what may have been a 50mm2 to a 80mm2 won't be that much of a deal.

I doubt they're going to be able to increase both functional units and clock by that huge amount while still on 14nm, not for the same power consumption anyway.

We'll see. Yield issues at 14nm isn't in denial by anyone. The only question is to what degree and in what area? There's a plenty of variation that can happen based on how well the project is executed. Also maybe getting cheap crap dies is what allowed them to reach contra revenue promises?

I am not hoping much for Goldmont either. The problem is that Goldmont is delayed too. Delayed CPUs ALWAYS underwhelm.

 

[Exophase]

We'll see. Yield issues at 14nm isn't in denial by anyone. The only question is to what degree and in what area? There's a plenty of variation that can happen based on how well the project is executed. Also maybe getting cheap crap dies is what allowed them to reach contra revenue promises?

I am not hoping much for Goldmont either. The problem is that Goldmont is delayed too. Delayed CPUs ALWAYS underwhelm.

I doubt that the yield issues were so severe in Q1 2015 when x7-8500 came out that they can now have 50% higher clocks in the same power envelope.

We're not even talking about max clock since these GPUs can do 600MHz but not sustained under real loads. Looking at x7-8550 like you mentioned, there's no increase in max GPU clock. Maybe we can revisit reviews of some product using this newly stepped chip but I doubt we'll see a big increase in average clock speed in similar form factor devices.

Broadwell parts from the same launch period as x7-8700 could clock the Gen 8 EU at up to 1.1GHz. Cherry Trail likely loses some of that headroom due to more SoC focused process stages, a more perf/W and area friendly implementation, and less aggressive binning. But probably not enough to take it all the way down to 600MHz, let alone ~300-400MHz. Obviously that doesn't say anything about V/MHz but what I'm really saying is that Cherry Trail is already working in the lower part of the design's dynamic range. So I doubt there's this huge amount of slack to frequency uplift to be gained with yields. And I doubt Gen 9 does that much to change that.

 

[VirtualLarry]

Sure, they can sacrifice their ENTIRE Core mobile and make $20-30 Atom CPUs quite performant. Here's the issue: That does not guarantee a marketshare gain. In order for a loss of Core to be worth it, they would need 70%+ market of Smartphone/Tablet CPUs. That won't happen. So the way they are doing it now might have been a safer one.

Then Intel should pull out of mobile, full-stop.

What I mean by this is, they need to stop "sandbagging" in mobile, and get their top tech and process and design bits into their mobile lineup, if they have any hope of dominating the market.

Or, maybe, they're in a holding pattern for mobile, and want to wait for some of the many ARM vendors to go belly-up, so that margins for mobile SoCs go up, and then they will attempt to enter the mobile market "a second time", and compete with whomever the surviving companies are at that time.

 

[dark zero]

Then Intel should pull out of mobile, full-stop.

They need to be only available on Windows mobile, in Android, they still suck hard...

 

[IntelUser2000]

I doubt that the yield issues were so severe in Q1 2015 when x7-8500 came out that they can now have 50% higher clocks in the same power envelope.

I don't think we can say for a certainty whether the yield issue was serious enough to affect it that much. These are seriously complicated stuff here. Only figure we know about yield is that ambiguous graph comparing to 22nm with no y-axis label. Is that a comparison a) at the same die size as the 22nm one? b) different die size from the 22nm? c) perhaps they have other problems like reaching certain frequency/voltage curve.

Intel is clearly behind ARM competition so we have two possibilities:
-14nm process really sucks and they were totally lying to us
-They are not capable of putting out a competent uarch. Whether due to internal conflicts, politics, disorganization, there's only one result apparent to the outsiders

I am just trying to guess what's so wrong that they are so behind even with a process difference that shouldn't be that much either way.

Then Intel should pull out of mobile, full-stop.

Perhaps. Perhaps they thought that having some mobile presence actually offered them benefits on the PC. Would we have very affordable, very low power PCs without mobile advancements?

Either way, we've got no control of what they are doing. They are clearly in the trouble mid to long term no doubt about that, but then again, companies everywhere are starting to face such issues.

 

[beginner99]

They need to be only available on Windows mobile, in Android, they still suck hard...

Exactly.

The rumored Surface phone could be this. x86 based with Windows 10. Plug-it into a monitor and with continuum you have a normal PC. They could also make a dock that offers more performance by somehow cooling the phone and/or by having additional hardware included like a GPU. The cooling part being the most important. If you can go from 2W TDP to 15W you have a winner. Core m phone?