Low-cost emerging market SoC/phone discussion

  • Thread starter Deleted member 13524
  • Start date
"Intel is quietly working with chinese tablet vendors."

Quietly? They've been announcing low-cost tablets with chinese brands since early 2014. Any popular chinese e-store has been selling Windows 8 tablets from Pipo, Ramos, etc. for quite a while.
Most of them were formally shown at Computex.

Well let's just say that the authors at fudzilla are "creative" :p I don't know where that "quietly" comes from either; every time I snoop around at chinese online retailers Intel's presense in all kinds of tablets is constantly increasing. Must be some sort of "common secret" or some insider joke only they laugh about....

I just wanted to add that Intel should be counted to the "low cost emerging market" this thread is about ;)
 
http://www.anandtech.com/show/8403/examining-huaweis-benchmark-optimizations-in-the-ascend-p7

Sorry not acceptable from anyone and not from Huaweii either. Keep up the good word Anandtech.



The problem is that K910T SoC in a supposedly high-end smartphone.
The 910 is actually very similar in performance to the Snapdragon 600 or the MT6592. It uses the same GPU as the MT6592 (Mali 450 MP4).

Come the reviews, everyone was going to compare benchmark results, and in China they love AnTuTu which scales really well with CPU frequencies.. so they bumped the clocks on the CPU cores in order to stay on top of last year's flagships.

I guess they had planned the 910 for their midranges and the 920 for the high-ends/flagships. The 920 probably got delayed so they had to settle with the inferior SoC for the Ascend P7.
I'm not saying this is excusable. I'm only guessing why they would feel the pressure to cheat on the benchmarks.
 
The problem is that K910T SoC in a supposedly high-end smartphone.
The 910 is actually very similar in performance to the Snapdragon 600 or the MT6592. It uses the same GPU as the MT6592 (Mali 450 MP4).

Come the reviews, everyone was going to compare benchmark results, and in China they love AnTuTu which scales really well with CPU frequencies.. so they bumped the clocks on the CPU cores in order to stay on top of last year's flagships.

I guess they had planned the 910 for their midranges and the 920 for the high-ends/flagships. The 920 probably got delayed so they had to settle with the inferior SoC for the Ascend P7.
I'm not saying this is excusable. I'm only guessing why they would feel the pressure to cheat on the benchmarks.

It's time everyone puts an end to it.

As for Antutu I honest wish there would be something in its place that would measure something worth mentioning. By the time you consider that an awkward octacore A7 (MTK6592) has a similar Antutu score with a quad A9 (R4) config it already speaks volumes how wortheless that thing really is.
 
By the time you consider that an awkward octacore A7 (MTK6592) has a similar Antutu score with a quad A9 (R4) config it already speaks volumes how wortheless that thing really is.


Antutu may be worthless, but the MT6592 is undoubtedly a spectacular chip for the price range of the phones it goes into. It's faster than any Snapdragon 400 in pretty much everything.
Nonetheless, SoC CPU/GPU speed is starting to make little difference in "user experience" performance.
Things are getting bottlenecked by mass storage performance mostly, which is why you'll see videos on the net showing a Moto G beating a Galaxy S5 to opening apps.
 
Antutu may be worthless, but the MT6592 is undoubtedly a spectacular chip for the price range of the phones it goes into. It's faster than any Snapdragon 400 in pretty much everything.
Nonetheless, SoC CPU/GPU speed is starting to make little difference in "user experience" performance.
Things are getting bottlenecked by mass storage performance mostly, which is why you'll see videos on the net showing a Moto G beating a Galaxy S5 to opening apps.

I expected a wee bit more from the 6592 to be honest, since I thought it would be a clean "little.LITTLE" implementation. Other than that the 6592 is also highly successful for tablets it seems, whereby the tablet intended MT8135 seems like a complete flop so far.

That said I very much look forward to upcoming quad A53 + quad A53 configs we're going to see soon from multiple sides http://www.gsmarena.com/htc_teases_the_desire_820_destined_to_launch_at_ifa_2014-news-9418.php
 
Not sure if this belongs here, but some Geekbench results from the Cortex A53 powered Snapdragon 410 compared against Cortex A7 based Snapdragon 400. Very nice performance increase! Not sure how Qualcomm has tweaked the Cortex A53, but based on these results it can be a potent core.

http://browser.primatelabs.com/geekbench3/compare/768801?baseline=753027

John poole of Primatelabs pointed out that the results are for the ARMv7 version of the binary, and likely the scores for the Snapdragon 410 will be higher with Aarch32 version (and possibly even higher in the 64-bit version).
 
I expected a wee bit more from the 6592 to be honest, since I thought it would be a clean "little.LITTLE" implementation. Other than that the 6592 is also highly successful for tablets it seems, whereby the tablet intended MT8135 seems like a complete flop so far.

That said I very much look forward to upcoming quad A53 + quad A53 configs we're going to see soon from multiple sides http://www.gsmarena.com/htc_teases_the_desire_820_destined_to_launch_at_ifa_2014-news-9418.php

So mediatek got the socket in the new 6 and 7 inch kindle fires, with the long await MT8135. Those form factors should sell a nice few million, given the attractive pricing.

https://developer.amazon.com/public...ications/01-device-and-feature-specifications
 
Last edited by a moderator:
Not sure if this belongs here, but some Geekbench results from the Cortex A53 powered Snapdragon 410 compared against Cortex A7 based Snapdragon 400. Very nice performance increase! Not sure how Qualcomm has tweaked the Cortex A53, but based on these results it can be a potent core.

http://browser.primatelabs.com/geekbench3/compare/768801?baseline=753027

Looks quite good indeed, though the overall result is probably a bit skewed due to the crypto benchmarks which all got a huge boost. Other than that it seems the memory pipeline got the most significant changes, though I wonder what memory this device used. I'm baffled what happened with the multicore sgemm/dgemm scores though the core count scaling for some reason is laughable. In any case this is really the SoC the MotoG2 should have used...
 
Looks quite good indeed, though the overall result is probably a bit skewed due to the crypto benchmarks which all got a huge boost. Other than that it seems the memory pipeline got the most significant changes, though I wonder what memory this device used. I'm baffled what happened with the multicore sgemm/dgemm scores though the core count scaling for some reason is laughable. In any case this is really the SoC the MotoG2 should have used...

I thought the boost for integer (>30%) is nice. FP is lagging in comparison ( only ~9% in multicore). But I'm not sure if "integer" is not only a summary of all the benchmarks below.
 
But I'm not sure if "integer" is not only a summary of all the benchmarks below.

It is.

You need to discount all the crypto hash subtests to get an idea about what real life performance is like.

If you do that, it's around 10% faster

Cheers
 
It is.

You need to discount all the crypto hash subtests to get an idea about what real life performance is like.

If you do that, it's around 10% faster
More accurately, without AES, Twofish, SHA1 and SHA2, it's 16% faster (versus 33% with them) :)
 
You need to discount all the crypto hash subtests to get an idea about what real life performance is like.

Why? I doubt the version of Geekbench is using ARMv8 crypto instructions (if they are the improvement is really weak), and those never applied to Twofish anyway.

I don't think they should be thrown out any more than any other single data point should unless it can be shown that they're really far off from a typical program workload.
 
I don't think they should be thrown out any more than any other single data point should unless it can be shown that they're really far off from a typical program workload.

They make up a third of the integer benchmarks of Geekbench. They are weighted at least three decimal orders of magnitude higher than they should in a performance benchmark.

Cheers
 
Last edited by a moderator:
More accurately, without AES, Twofish, SHA1 and SHA2, it's 16% faster (versus 33% with them) :)

True, specifically:

Bzip compress: +8.5%
Bzip decompress: +14.9%
JPEG compress: +27.0%
JPEG decompress: +16%
PNG compress: +15.9%
PNG decompress: +19.0%
Sobel: +18.5%
Lua: +21.5%
Dijkstra: +10.5%

Arithmetic mean: 16.9%
Geometric mean: 18.8%

Cheers
 
Back
Top