First mediatek phone hits glbenchmark

[tangey]

Mediatek has been making big noises in China with its MT6573 Soc, that includes an unknown SGX core. Recent press announcements from them have shown that the chip is shipping in big volume, and having only started shipping in August, they are predicting 10M by the end of this calendar year. there has been press reports that mediatek has had to look for more production capability from its foundry partners as demand for the chip has exceeded expectation.

It was thought by some that this chip was using SGX520, IMG's smallest and lowest performing GLes2.0 compliant core, although I saw no info to lead me to believe that.

The Lenovo a60 phone has just been posted to Glbenchmark, and it confirms that the Soc in fact contains SGX531.
http://www.glbenchmark.com/phonedetails.jsp?benchmark=glpro21&D=Lenovo+A60&testgroup=overall

This core was initially licensed to Philips NXP division 3 years ago for STB (then sold to trident) and until recently this was the only known use of this core. At the time, much was made of its major unique feature, that being a 128-bit bus.This core has recently also been revealed to be in an Renesas chip destined for In-Car navigation solutions, the R-Car-E1 chip. (the chip is from a line of chips that Renesas inherited from NEC)

Interesting to note that this low cost chip is slightly outperforming the Iphone 3GS on graphics tests, and significantly outperforming it on the "pro" tests.

 

[Deleted member 13524]

Seems to be replacing the MT6516 in chinese dual-sim knock-offs.

It's quite a big upgrade, since the older one had a 460MHz ARM9 + 280MHz ARM7, and this one has a 650MHz ARM11 + 400MHz ARM9 + SGX531. The new chip should provide a much more decent user experience. My brother owns one of those HD2 knock-offs with a MT6516, it's painfully slow in everything.


Regarding the score, it's really weird that the wider bus isn't helping the SoC in the offline 720p scores, when comparing to 3GS.


Any thoughts on how they split the tasks in Android between the ARM11 and ARM9? These don't support heterogeneous computing off-the-box, right?

 

[Lazy8s]

Good question.

Switching focus, pixel fill possibly indicates 200 MHz for the GPU? It's got 1 TMU.

 

[Ailuros]

Regarding the score, it's really weird that the wider bus isn't helping the SoC in the offline 720p scores, when comparing to 3GS.

As Lazy8s said SGX531 = 1 TMU vs. 3GS/SGX535 = 2 TMUs. If those two should have comparable frequencies the latter has twice the fillrate.

 

[Arun]

I could be wrong (maybe someone from IMG could clarify) but the main benefit of the 128-bit bus would be for chips with quite a lot of bandwidth (e.g. STB) for UI interfaces using uncompressed 32-bit textures. It just won't help a low-end handheld as much. So why did Mediatek go for it? Maybe because they wanted to use it in one of their TV chips as well (no idea if they did so in the end) and the cost was basically identical. Or maybe it also supports a relatively newer bus standard better (e.g. AMBA3 AXI), I don't know.

BTW, Mediatek revealed their next-gen mainstream 3G baseband will support up to 14.4Mbps. I'm much more interested in their longer-term roadmap and how much they might use the Coresonic DSP they licensed though. They've got clear LTE ambitions (they already licensed the NTT Docomo modem a long time ago, not sure if they'll use it in the end) and combined with Rogue they might be able to target the high-end more effectively than many people expect.

 

[tangey]

The marketing for SGX531 at the time stated:-

"The core features an upgraded 128-bit internal bus architecture and related enhancements to maximize performance when integrated in the latest SoCs. This makes POWERVR SGX531 an ideal solution for next generation SoCs, including those targeting 1080p HD resolution displays and high frame rate applications."

 

[Ailuros]

I doubt Arun is wrong with his point about STB UI interfaces with uncompressed 32bit textures (amongst others).

 

[Arun]

I doubt Arun is wrong with his point about STB UI interfaces with uncompressed 32bit textures (amongst others).

FWIW, I got that tidbit from Kristof at MWC09, it's just been long enough that I'm not sure I remember it perfectly. And obviously it might not be the only consideration.

 

[tangey]

I doubt Arun is wrong with his point about STB UI interfaces with uncompressed 32bit textures (amongst others).

I wasn't suggesting he was wrong, merely adding their PR speak from the time. The first use of it, like I said in the original post, was indeed in STB chips from NXP (now trident)

Also if this anandtech article is correct, then the sigma chip used in this media player also contains SGX531 (original speculation was SGX530).
http://www.anandtech.com/show/4132/acryan-enters-us-market-with-innovative-media-players/3

That would make 4 instances of SGX531 known to date, the NXP chips, the Sigma Chips, The renesas in-car chip, and the mediatek chip.

 

[Rys]

Arun's right; it's effectively our smallest core tuned for 1080p UI use cases (available in a few configs of bus and cache, depending on target). It's not actually a 530 derivative though, so it's not quite that core with 128-bit, although if you squint a bit it mostly looks like it :smile:

 

[Ailuros]

Sorry for the bump, but I didn't want to open another thread for it:

http://www.itproportal.com/2012/09/28/mediatek-to-launch-8-core-mt6599-soc-with-lte-in-q1-2013/

Octacore A7? Is it just me or is that a rather weird design decision?

Other claimed features include the use of Imagination Technologies’ PowerVR SGX544 GPU (the same as the A5X on the Apple iPad 3 and a notch above the Apple iPhone 5 and the A6),...

A notch above for what?
​

 

[Exophase]

Maybe it really means 4 Cortex-A15 + 4 Cortex-A7 cores.

 

[Ailuros]

Maybe it really means 4 Cortex-A15 + 4 Cortex-A7 cores.

While it sounds more reasonable than 8*A7, it still sounds like overkill unless I'm missing something.

 

[Exophase]

While it sounds more reasonable than 8*A7, it still sounds like overkill unless I'm missing something.

The only other alternative which would work would be 8 Cortex-A15s, which is the least reasonable option of all. 8 Cortex-A9's can't be done, at least not with coherency using standard ARM IP.

Of course they could be playing games with what "core" means, counting embedded and non-user accessible cores. Or the 8-core reports could just be mistaken.

Four Cortex-A7s is already kind of weird, outside of embedded applications that really benefit from having four cores for realtime purposes.

 

[mczak]

Four Cortex-A7s is already kind of weird, outside of embedded applications that really benefit from having four cores for realtime purposes.

Oh I could see some reasons why you'd want 4-core Cortex-A7:
- up to 4 can share a singe cpu cluster (hence l2 etc. is shared) and a single core is only about 0.5mm² on 28nm according to arm hence the die size difference between 2 and 4 cores is very very small - I believe (the numbers are not easily comparable) 4 Cortex-A7 are smaller than 2 Cortex-A9.
- on the power front 4 Cortex-A7 might require less power than 2 Cortex-A9, while providing higher raw power if you've got software that can use that many cores, and even if you don't (which admittedly is very likely for smartphone usage) it will at least provide better perf/power.
- in some ways it is more modern than Cortex-A9. In particular it comes with 128bit AMBA4 interface instead of dual 64bit AMBA3 which might be important if you for instance had a whole family of different chips and you'd want to standardize some things. It also comes with features the Cortex A9 lacks (like virtualization support, and I could see that being potentially useful on a smartphone, and some other minor features, for instance VFPv4).

There are chips out there in smartphones which prefer dual-core Cortex-A5 over single-core Cortex-A9 (in some newer lowend devices) too, so quad-core Cortex-A7 doesn't look that odd. Granted 4 cores of anything doesn't sound too useful in a smartphone right now, but that didn't stop companies from producing quad-core chips until now neither...

I agree though 8 cores don't make sense.

edit: actually qualcomm announced quad-core lowend chips (msm8225q) which is supposedly quad-core Cortex-A5. I'd consider that definitely more odd than quad-core Cortex-A7...

 

[Exophase]

Yeah, I wrestled with that comment for a while. I can see where it might work out better but I'm kind of on the fence on it. The actual comparisons (between A7 and A9) are going to be hard w/o clock speeds, at any rate. The decision probably also has a lot to do with licensing costs.

I actually hadn't heard of any dual core Cortex-A5 chips, just single core. Much less quad core.

 

[Ailuros]

Even weirder scenario (since marketing and relevant messages are more than often awkward), how about someone just said octacore SoC and they're not only counting CPU but also GPU cores?

 

[shyam]

Mediatek is going to take the low-mid end market with TI's exit.

MT6588/6589,6583 spec.


Qcom has dual core Cortex A5's (msm8225,8625),with few devices shipping already (HTC Desire X,Huawei G330d).

 

[shyam]

Can't edit

 

[Ailuros]

Mediatek is going to take the low-mid end market with TI's exit.

MT6588/6589,6583 spec.


Qcom has dual core Cortex A5's (msm8225,8625),with few devices shipping already (HTC Desire X,Huawei G330d).

Is that a quadA9 in the MT6588 or something else? The GPU smells like a SGX544MP1@320MHz