Qualcomm Krait & MSM8960 @ AnandTech

[anexanhume]

How can you double on any architecture ALU count and not shaders, especially with USC ALUs like in Adrenos or SGX? If you meant TMUs for instance, yes it could very well be the case here also.

Yes, sorry my mistake.

 

[Ailuros]

Well if they doubled ALU amounts a 50% performance increase is quite low (marketing typically illustrates best case scenarios). iPad3 has a 543MP4@250MHz vs. a 554MP4@280MHz in iPad4 whereby the latter is nearly twice as fast in GLBenchmark2.5 than the latter.

 

[french toast]

Well this is really exciting news...im suprised many more geeks on this forum havnt got more stuck in to this topic since I have been away .. perhaps the disscusion has taken place on another thread.

Anyhow...if you look at the snapdragon 800 demo..I forget what it is called...the one with the giant dragon....it actually shows the adreno 330 frequency! It reads 450mhz. ..

Did we actually get the full details of adreno 320 uarch and frequencies?...if it turned out to be 300mhz then perhaps its just a 50% overclock with improved drivers and improved bandwidth via the lpddr3....not forgetting the increased krsit performance likely to slighly improve 3d also.

 

[ltcommander.data]

Well this is really exciting news...im suprised many more geeks on this forum havnt got more stuck in to this topic since I have been away .. perhaps the disscusion has taken place on another thread.

Anyhow...if you look at the snapdragon 800 demo..I forget what it is called...the one with the giant dragon....it actually shows the adreno 330 frequency! It reads 450mhz. ..

Did we actually get the full details of adreno 320 uarch and frequencies?...if it turned out to be 300mhz then perhaps its just a 50% overclock with improved drivers and improved bandwidth via the lpddr3....not forgetting the increased krsit performance likely to slighly improve 3d also.

Well if it's the same GPU architecture just frequency scaled shouldn't compute and graphics performance increase in sync rather than the 2x compute increase, 50% graphics increase claims? Unless they are somehow roping in the CPU improvements into some general "compute" description.

 

[french toast]

Well if it's the same GPU architecture just frequency scaled shouldn't compute and graphics performance increase in sync rather than the 2x compute increase, 50% graphics increase claims? Unless they are somehow roping in the CPU improvements into some general "compute" description.

Yes you raise a valid point...I couldnt think of a valid answer...but it would help if we hsd some details of 320 clock speed...does anyone have any?

 

[Helmore]

Yes you raise a valid point...I couldnt think of a valid answer...but it would help if we hsd some details of 320 clock speed...does anyone have any?

I thought the 320 had the same clocks as the 225, which has a max clock of 400 MHz AFAIK. I could be wrong and I'm not sure where I got it from. In other words, don't just take my word for it.

 

[french toast]

Lol thanks for the disclaimer.

Well that being the case, a 50mhz improvement with the same execution resources is not going to yeild the improvements advertised.

One rather interesting aspect im interested in is the batterylife of the new adreno 330 compared to the 225...krait v1 actually displayed very good power characteristics in anandtechs soc comparison article..whilst adreno 225 was average, qualcomms made some kind of claim that adreno 330 consumes half the power of previous generation...I assume adreno 225...pretty astounding achievement if correct.

 

[Helmore]

I thought they meant half the power of the Adreno 320, but you could be right as well. It would be quite a feat if that was the case. From what we know of some of AnandTech's tests, the Adreno 320 already suffers from thermal throttling in the Nexus 4. I hope that won't be the case with the Adreno 330.

 

[french toast]

Yes, im dissapointed in anandtech drawing too many conclusions from nexus device's. ....history tells us google purposely cripples its nexus device's to appease oems.

Saying that..their breakdowns and conclussions on medfield being competitive with arm has been very good in changing opinion on intels mobile solutions...me for one couldnt see intel matching or beating arm on power consumption for years..

Anyhow proof that google once again crippled nexus 4 (and I assume nexus 10 also) is with the benchmarks between it and the droid dna...even optimus g is better despite identicle software.
Batterylife on s4 pro is actually pretty good for dna..considering the massive 1080p screen, average battery and performance on offer...would have been nice for anand to compare a non nexus device s4 pro..like the dna..vs the motorola razr I ...4 thread latest socs from premium oems running identicle android...I suspect the adreno 320 and extra cores of the snapdragon would lower its power consumption and increase its performance figures vs the s4 8960 used.

Edit..seems ive got my wires crossed..I meant to say lenovo k900 which carries the dual core/4 thread @ 2.0 ghz medfield with I expect an sgx 545 @ around 500mhz..and 2gb ram...yes a performance/power comparison between the k900 and htc droid dna would have been a proper fair comparison and one I couldnt predict a clear winner .

 

[Arun]

From the recent TSMC Q4 CC:

High-k metal gate will surpass oxynitride. That's in the 28-nanometer. We have, as you know, actually four types. Three of those are high-k metal gate, and the earliest type that we introduced was the oxynitride, and indeed last year, the majority of the production was the oxynitride. But the more advanced version high-k metal gate will surpass oxynitride in the third quarter this year. And in the fourth quarter, it will even surpass oxynitride even more.

AFAIK Qualcomm is by far the primary 28LP customer (there are others though, e.g. Altera which uses both 28LP and 28HP). This very likely means that Qualcomm Krait-based products alone were the majority of the entire 28nm production at TSMC(!)

A few other noteworthy points:
- TSMC expects to triple 28nm production in 2013, with market share expected to slightly decline from nearly 100%(!) to nearly 90%(!). This does not include 32nm (e.g. Apple/AMD).
- TSMC expects 20nm production in 2014 to be greater than 28nm production was in 2012, and same for 2015 vs 2013. So potentially faster than Moore's Law but not by much.
- 16nm production is expected to be very small in 2015, which implies closer to a 18-24 month cadence than TSMC's previously claimed 12 months.

So based on that, the big question is... when is Qualcomm's 20nm roadmap going to leak?
Full transcript: http://www.tsmc.com/uploadfile/ir/quarterly/2012/43tP7/E/12Q4 Conf Call-Transcript-callstreet.pdf

 

[french toast]

From the recent TSMC Q4 CC:
AFAIK Qualcomm is by far the primary 28LP customer (there are others though, e.g. Altera which uses both 28LP and 28HP). This very likely means that Qualcomm Krait-based products alone were the majority of the entire 28nm production at TSMC(!)

A few other noteworthy points:
- TSMC expects to triple 28nm production in 2013, with market share expected to slightly decline from nearly 100%(!) to nearly 90%(!). This does not include 32nm (e.g. Apple/AMD).
- TSMC expects 20nm production in 2014 to be greater than 28nm production was in 2012, and same for 2015 vs 2013. So potentially faster than Moore's Law but not by much.
- 16nm production is expected to be very small in 2015, which implies closer to a 18-24 month cadence than TSMC's previously claimed 12 months.

So based on that, the big question is... when is Qualcomm's 20nm roadmap going to leak?
Full transcript: http://www.tsmc.com/uploadfile/ir/quarterly/2012/43tP7/E/12Q4 Conf Call-Transcript-callstreet.pdf

So...qualcomms snapdragon 800 will be shipping with this hkmg technology..how much of an improvement would this new hkmg improve efficiency in snapdragon 800 over snapdragon s4s oxynitride process?

 

[Ailuros]

Edit..seems ive got my wires crossed..I meant to say lenovo k900 which carries the dual core/4 thread @ 2.0 ghz medfield with I expect an sgx 545 @ around 500mhz..and 2gb ram...yes a performance/power comparison between the k900 and htc droid dna would have been a proper fair comparison and one I couldnt predict a clear winner .

It's OT here but the Lenovo K900 carries afaik a Clovertrail+ with a SGX544MP2@533MHz.

 

[almighty]

Problem with Adreno GPU's is that the drivers are crap...

I've had phones running the Adreno 200, 205 and the 220 and performance has been awful..... But 12 months ago Adreno released a new set of drivers that boosted performance by 2-3x

It's no use them showing off this new tech if its going to take them 12months+ to sort the drivers out.

 

[Jubei]

Qualcomm is supposedly releasing new drivers for Krait soon

 

[3dcgi]

- TSMC expects to triple 28nm production in 2013, with market share expected to slightly decline from nearly 100%(!) to nearly 90%(!). This does not include 32nm (e.g. Apple/AMD).

TSMC doesn't have 32nm. Or by Apple/AMD do you mean Samsung and Global Foundries?

 

[Alexko]

From the recent TSMC Q4 CC:
AFAIK Qualcomm is by far the primary 28LP customer (there are others though, e.g. Altera which uses both 28LP and 28HP). This very likely means that Qualcomm Krait-based products alone were the majority of the entire 28nm production at TSMC(!)

A few other noteworthy points:
- TSMC expects to triple 28nm production in 2013, with market share expected to slightly decline from nearly 100%(!) to nearly 90%(!). This does not include 32nm (e.g. Apple/AMD).
- TSMC expects 20nm production in 2014 to be greater than 28nm production was in 2012, and same for 2015 vs 2013. So potentially faster than Moore's Law but not by much.
- 16nm production is expected to be very small in 2015, which implies closer to a 18-24 month cadence than TSMC's previously claimed 12 months.

So based on that, the big question is... when is Qualcomm's 20nm roadmap going to leak?
Full transcript: http://www.tsmc.com/uploadfile/ir/quarterly/2012/43tP7/E/12Q4 Conf Call-Transcript-callstreet.pdf

Wasn't there some noise at some point about Qualcomm using GloFo for 28nm as a second source?

Edit: ah, there it is.

What should be noted, is that Globalfoundries and Qualcomm back in 2012 entered into a non-binding memorandum of understanding to collaborate on leading edge technologies, such as 45nm low power and 28nm LP with an intended collaboration on future advanced process nodes. Therefore, it is logical for Qualcomm to contract GF, which is why it was surprising that several weeks ago the rumours only pointed to Samsung and UMC.

http://www.xbitlabs.com/news/other/...rd_to_Make_28nm_Chips_at_Globalfoundries.html

But if TSMC had ~100% market share for the 28nm node in 2012, I guess this didn't materialize. Incidentally, that makes me worry a bit about GloFo's process and what it means for Kaveri, but that's another story.

 

[Exophase]

But if TSMC had ~100% market share for the 28nm node this year, I guess this didn't materialize. Incidentally, that makes me worry a bit about GloFo's process and what it means for Kaveri, but that's another story.

Rockchip's RK3188 was made on GlobalFoundries' 28nm process and technically first appeared in devices before 2013 began.

 

[Mariner]

I've not seen reviews or mention of any RK3188 devices being available as yet.

The Allwinner A31 (Quad-A7 @ 1GHz + PowerVR SGX544 MP2) is available in one or two tablets. Initial reviews seem to indicate battery life is disappointing and performance is underwhelming in comparison to quad-A9 devices which tend to be clocked much higher.

 

[Exophase]

I've not seen reviews or mention of any RK3188 devices being available as yet.

I don't know about reviews, but you can buy them.

http://android-sale.com/cube-u30gt-...ch-1920-1200-pixel-screen-android-tablet.html

AFAIK that's not the only one.

The Allwinner A31 (Quad-A7 @ 1GHz + PowerVR SGX544 MP2) is available in one or two tablets. Initial reviews seem to indicate battery life is disappointing and performance is underwhelming in comparison to quad-A9 devices which tend to be clocked much higher.

Without knowing anything about what battery these tablets are using, as well as the rest of the hardware, we can't draw any conclusions about the power consumption of the SoC. For instance, Archos Gamepad gets terrible battery life in no small part because it has a terrible battery.

But what exactly does A31 have to do with RK3188 or 28nm?

 

[Helmore]

But if TSMC had ~100% market share for the 28nm node this year, I guess this didn't materialize. Incidentally, that makes me worry a bit about GloFo's process and what it means for Kaveri, but that's another story.

I presume you meant to say that TSMC had nearly 100% market share during 2012, right? The way you're wording it makes it seem as if TSMC will have nearly 100% market share during 2013 as well....