Intel's smartphone platforms

http://files.shareholder.com/downlo...9b220d3/Hermann_Eul_Computex_2013_keynote.pdf

Intel says Bay Trail is 2x over Clover Trail, not CT+. Odd.

 

In TouchXPRT 2013, which I think is CPU-only. CPU-wise what's different between CT and CT+, 1.8GHz to 2GHz? I wonder how well threaded TouchXPRT is.

Speaking of which, I wonder if TouchXPRT is using ICC for the Win 8 version, and/or Intel's math libraries (big functions of TouchXPRT are image enhance and photo export which could easily be swayed by quality of SIMD.. no idea what "podcast" does). There has to be a reason why Intel would be compelled to showcase a much less well-known Windows benchmark over an Android one. Although at least these scores look less ridiculous than the AnTuTu ones.

Then again, judging by power consumption here:

http://www.anandtech.com/show/6536/arm-vs-x86-the-real-showdown/5

It looks like the Snapdragon S4 takes about as long as the Atom in most tests - at least for photo enhance, photo export, and podcast. And yet, who else could competitor B be? Are there any other ARM SoCs in WinRT products outside of Tegra 3 and Snapdragon S4Pro? Maybe the hugely different video test is what's screwing it up. That one has definitely got to be a software issue. I wonder why AT gave scores here for WebXPRT and not TouchXPRT. I guess they're probably in another review.

So this new slide deck repeats the 50% improvement for Merrifield claim, and looks especially odd when paired next to a claim of 100% improvement for BayTrail-T. Is it really certain that they're talking about a 4C SoC? If so it must be clocking pretty low, but I don't see why they need to go so conservative with clocks - surely it should be capable of boosting pretty high so it can win those benchmarks by the biggest amounts possible, er, so they can have good burst performance for loading webpages and apps.

For me the GPU comparison vs Clovertrail is more interesting (or should I use the word "telling") than the CPU one. The best Egypt HD score I can find for an SGX540 Intel device is here: http://gfxbench.com/device.jsp?benchmark=gfx27&D=Asus+ME371MG I thought this might be worse than Clovertrail since it's just Lexington, but an actual Clovertrail score here looks worse: http://www.anandtech.com/show/6872/...ndows-tablets-compared-using-gldxbenchmark-27

4x this score would be about 28-34fps which isn't that impressive for late 2013 tablets. It's also close to what CloverTrail+ already does in phones (http://gfxbench.com/device.jsp?benchmark=gfx27&D=Lenovo+K900), a score also in line with Intel's original CT+ projections. No wonder they don't want to make that comparison.

It's going to be really funny if Merrifield has a faster GPU than Bay Trail-T.

 

Would also like to point out that this makes Anand's presumption about Bay Trail-T's GPU performance (roughly A6X-level) sound way off.

 

LOL

Clovertrail is in DX2.7 at 163 frames / 2.9 fps. Now Intel is using GLB2.5 or Ice Storm for a reason since it's times lighter than 2.7. In GLB2.5 offscreen Clovertrail is at 926 frames / 8.2 fps the entire enchilada times 3x = 2778 frames / 24.6 fps.

Apple iPad4 in GLB2.5 is at 6149 frames / 54.4 fps; if the above speculative math should be for real than the iPad4 is ~2.2x faster than Bay Trail and Lord knows how the picture looks like in GLB2.7, since if the difference there would be larger than 3x times I don't think any marketing department would chose to quote GLB2.5 results instead

Jokes aside there's something not adding up here; they most likely compared against Clovertrail+ and forgot (whether on purpose or not) the "+"
Anything else doesn't make much sense for a tablet design.

And to make it even more complicated:

Z2520 = SGX544MP2@300MHz
Z2560 = SGX544MP2@400MHz
Z2580 = SGX544MP2@533MHz

 

Not sure if this was posted yet

http://www.eetasia.com/ART_88006860...2760,9950138804,2013-06-11,EEOL,ARTICLE_ALERT

 

Can't say what benchmark they picked, but the CPU results look very odd. Do I need to repost a Geekbench link comparing K900 to any other ARM phone? They obviously did a poor job at comparing performance or they had an agenda

 

Or Intel's technical marketing is still trying to figure out what is what in the SFF mobile market....uhmm wait that sounds familiar

 

They actually said 4x faster than CloverTrail specifically in GLB2.5 Egypt HD offscreen (page 28). Or 32.8fps, just barely faster than CloverTrail+.

At the bottom of the slide they list the configuration they tested, which is a CT tablet in Windows 8. So they definitely didn't just forget to write in the +. I doubt they would have wanted to compare with CT+ anyway since it isn't available for Windows tablets. And there's no way a GPU characterized as 1/4th of HD4000 would have been able to get ~130fps in GLB2.5. A full HD4000 in a ULV IB doesn't even perform that well (108fps for Surface Pro). So what we see here is something that performs better than 1/4th of HD4000, but not that much. Even Nexus 10 will beat this.

Note that 8.2fps is already pretty generous, since Anand gets 7fps with the same test configuration Intel gave. So it's possible the BayTrail-T (Z3770) projected score is only 28fps. It's possible that it performs worse than CloverTrail+! Whatever the case, it won't look good vs other tablets. You may think it doesn't make sense for a tablet design, but it doesn't make any less sense than Medfield or CloverTrail's anemic SGX540 did in 2012.

 

You missed the Acer W510 review.

http://www.anandtech.com/show/6522/the-clover-trail-atom-z2760-review-acers-w510-tested

Obviously competitor A and B is Tegra 3 and Snapdragon. It isn't the first time they quoted Anandtech numbers.

I do not think its coincidence the slide deck titled "Not All Cores Are Created Equal" back with Silvermont reveal are using dual core Silvermont vs. quad core competition at 1W. It's Smartphones.

-Desktop 4th Gen Core with GT2, and Laptop 4th Gen Core with GT3
-Clover Trail with SGX545 versus Clover Trail+ with SGX544MP2

It wouldn't be the first time.

EDIT: Just seen the part where Intel's Tom Kilroy talks about Merrifield. He never mentions quad core, they only say that for Tablet Bay Trail.

 

So when I said "I guess they're probably in another review" you must have thought I meant "they must not be in another review" e_e I assumed it was there, I just didn't really care enough to go dig it up at the moment.

The reason I mentioned it is because earlier in this thread the claim was made that the Merrifield SoC used in the comparison had four cores.

The desktop vs laptop situation isn't comparable. A laptop uses the IGP for better perf/W. The desktop can just be paired with a discrete card, so it can stand to have something weaker. That said, there are desktop Haswells with GT3, even one with GT3e.

It happened with CT vs CT+, however that's more of a matter of timing, with CT using the same GPU as the previous Saltwell cores. This time they're coming out at closer times so there isn't much reason why that should cause the GPUs to be different.

But the real reason why this is strange is because it makes you wonder why Intel chose the Gen 7 core for BayTrail-T in the first place. I assumed that Merrifield got IMG because it had better perf/W but BayTrail-T got Gen 7 because it had better perf within an acceptable tablet power budget. This would not appear to be the case. I guess there are other reasons too - Intel would want a DX11 part for Windows 8 even if it's too weak to run DX11 games worth a damn. Maybe it takes up less space. And it saves them money in royalties. But really maybe the GPU is just not that good and this is a starting point for evolving it down the road. Or worse, Intel is forcing a technically inferior in-house solution to satisfy their pride.

Whatever the case may be, Temash will easily put it in its place GPU-wise.

 

Geez I wonder how high royalties for Intel are in the end considering how antiquated the GPU and video decoding/encoding IP is when they integrate it into devices.

Because IMG's driver team was too dumb to deliver a WHQL DX10.1 windows driver for SGX545 maybe? Last time I checked it comes with an OpenGL3.1 and a DX9L3 driver and that for any device they've used the 545 so far in. I don't blame Intel one bit in that regard.

IMHO it's more expensive to remove DX11 for Intel from Gen7 than to leave it inside. Besides if you have it already why bother removing it?

Intel was way too conservative from any of their starting point (pick and chose when you want to start counting) with GPUs in their SFF mobile SoCs, so there's no surprise here really either. Irrelevant of what they'll use in Merrifield in the end I wouldn't in your place expect anything that is comparable with competing solutions in the future. As for how it'll compare to their own tablet solutions who cares?

 

3-4x will make it competitive with Hondo. While that's slower than Temash, it'll be far closer than they are now against the 3.9W version. I don't think they are worried directly competing against AMD with graphics, especially their advantage will be everywhere else.

It's not just that. Bay Trail will go into embedded devices, and servers which often use Linux. Moving to GenX means much better support.

Strangely they are not boasting about graphics on Merrifield. I'm guessing the gains aren't anything to boast about.

 

Posted on RWT:
http://www.realworldtech.com/forum/?threadid=134968&curpostid=134968

It seems that Intel missed the mark looking at those early results, nice improvements against Satwell but it doesn't align with Intel early claims.

 

I was wondering why float seemed so low first but finally realized the android float results aren't really comparable to the windows ones it seems, you can see that there is a huge difference for the Acer ICONIA W3-810 vs. LENOVO Lenovo K900_ROW (both Clovertrail). I think this is the uninitialized float problem which might become denorms showing up as mentioned by jfpoole here: http://forums.anandtech.com/showthread.php?t=2330027&page=2. Plus it's running 32bit windows which might mean x87 fpu instead of sse(?).

So just comparing against the Iconia the improvement is quite solid. Something like a 40% IPC improvement (and of course multithread scaling is better). At least IFF assuming it really runs at that clock and doesn't turbo up (would be really fail otherwise), and that indeed can reach those 2Ghz+ clocks. Not-quite-Kabini/Temash (or Cortex-A15 for that matter) but not that far off neither per clock so if the clock and power consumption side of things hold up it looks quite ok. So probably quite competitive but not the total arm-killer intel claims it to be . The 50% improvement claimed over Clovertrail doesn't seem far off however. Of course that only addresses the cpu side of things, gpu might fare worse (compared to competition not predecessor), however feature-wise the gpu will definitely be top notch slaughtering all the DX9L3 stuff out there with full DX11 support. But in any case, those development platform benches tell you only half the story since all the meat is in perf/w which you just don't see in those benches.

 

Sorry if this has been brought up before, but has Intel showed any roadmap of integrating iris pro into future mobile tablet/smartphone chips? I know they can bring the TDP down quite far by lowering the core count(like the HD 4000), but I am unsure of how eDRAM comes into play outside of cost and space. How would a small amount of eDRAM(lets say 16MB) coupled with a bus width appropriate for a balanced system impact battery life?

 

That I believe was without turbo.

Newer results: http://www.realworldtech.com/forum/?threadid=134968&curpostid=135358 (http://browser.primatelabs.com/geekbench2/compare/2188036/2226937)

 

Also, apparently, the Geekbench image processing FP benchmarks are stock full of denorms, which on Intel processors causes a trap (because correctness matters) while ARM cpus doesn't (flushing to zero/largest).

Mobile computing desperately need a credible benchmarking suite.

Cheers

 

Geekbench as run on ARM Android doesn't use FTZ. It's just that ARM CPU handle subnormals (new terminology as per IEEE754-2008 ) in hardware rather than relying on microcode as Intel CPU.

If you want to compare Intel and ARM CPU you can recompute FP score ignoring the two tests the process subnormals, sharpen and blur. I get 1050 for Silvermont, 1130 for Nexus 10, 1080 for S4 QC, 1220 for S4 A15, 830 for RAZRi. So FP doesn't seem to have improved that much, if the frequency indeed is 2.4GHz.

EDIT: forgot to say the numbers above are geo-mean of single threaded FP tests (without sharpen and blur).

 

Dumb question: couldn't reviewers just approximate display brightness on devices they want to compare and then with full batteries run on all of the exact same routine consisting of web browsing, video and some 3D amongst others and then note how much battery has been left?