Qualcomm Krait & MSM8960 @ AnandTech

QCOM going for a tablet push: http://fudzilla.com/home/item/31957-qualcomm-about-to-distance-itself-from-intel

 

According to Anandtech, the One's CDMA versions use a different baseband processor, the MDM9615.

 

No. In all recent Qcom products, only x6xx and x9xx parts support CDMA (e.g. 8660 8960 8974 9600 9615). So MDM9215 will not support CDMA.

 

http://www.fudzilla.com/home/item/32259-new-kindle-fire-hd-sports-2560x1600-display

That's quite a noticeable foot in the tablet market door for Qualcomm; congratulations. Let's see what Amazon exactly plans for it.

 

http://gfxbench.com/device.jsp?benchmark=gfx27&D=Amazon KFAPWA
(2560x1600)

http://gfxbench.com/device.jsp?benchmark=gfx27&D=Amazon KFTHWI
(1920x1200)

 

Despite that there had been rumors for the MT8135 contending for it at this stage the S800 sounds like a better choice, albeit I wouldn't necessarily bet that it was cheaper than the former.

 

I have to admit I'm a surprised by this, I really didn't expect Amazon to put such high-end hardware in a Kindle tablet. I wonder what kind of price point they'll manage to hit with this.

 

It's been suggested that they will have a lower end product, that might use the mediatek chip, perhaps in new versions of the non "fire" kindles.

 

I've been trying to determine why the Galaxy S4 LTE Advanced performs significantly better than other 2.3 GHz S800 based phones in graphics tests, and the reason may be that the Adreno 330's clock rate isn't topping out at 450 MHz in these initial devices after all. Apparently, some variants may already be specced up to 550 MHz, though I can't be sure the Galaxy phone gets its boost for that reason or that it reaches that top clock rate (although, Galaxy flagships have featured higher clocked GPUs in the past, like the 450 MHz Adreno 320 version of the S600 last time around.)

http://anandtech.com/show/7276/xiaomi-announces-the-mi3-tegra-4-and-msm8974ab-inside

What I find a little interesting is that 500 MHz was the target clock rate for the Adreno 420 from the leaked Qualcomm roadmap, so Qualcomm apparently starts from a target GPU clock for each platform and then ranges upward for better (binned?) variants.

Another edit: After some more research, I see people have been overclocking the GPU to at least 600 MHz, but I see throttling becomes an issue in some benchmarks even at the default clocks on S800 devices like the Galaxy S4 LTE-A.

 

Beat me to it...this has caught me by surprise..in a nice way galaxy note 3 has been benched at 68fps on gfx bench so I wonder whether this carries the enhanced AB s800?

Xiaomi mi3 at least will...I really want this phone

 

Note 3 is running 4.3 though

Got this from Brian Klug, its the cheaper bin of 8974 (2.15 ghz)

http://pastebin.com/VhyujVGb

 

Good point about 4.3...likely contains new adreno drivers.
Right..so there are what three s800 bins? Same as with s600?

-Krait 400 @ 2.15 adreno 330 @ 450mhz lpddr 3 @ 800mhz/12.8 gbs (1600*32*2/8)

-Krait 400 @ 2.26 adreno 330 @ 450 mhz lpddr 3 @ 800mhz/12.8gb/s.

-Krait 400 @ 2.3 adreno 330 @ 550 mhz (+22%) lpddr 3 @ 933 mhz/14.9 gb/s (+16 %) ( MSM8974AB) -hexagon dsp also gets a boost to 465mhz. (+45%)

Adreno 330 @ 550mhz and dual channel lpddr3 @ 933mhz (1866 data rate) ...in a smartphone...yikes!

Edit..trying to find exact clock speeds of various s800 bins.

 

There are only 2 CPU bins for Snapdragon 800, but for the fluffies in marketing they must have rounded the freq number up.

So the lower binned version, which Sony uses is listed as a 2.2 GHz, but the sources reveal this as actually 2.15 GHz.
The higher binned version is listed as 2.3 GHz, but is actually 2.265 GHz. The Galaxy Note 3 according to Samsung's press release is using the higher binned CPU version 2.265 GHz (2.3). For the Snapdragon 600, the 1.7 GHz version was actually a 1.728 GHz, for the spec obsessed.

The GPU is more interesting, but there were no clues in the kernel sources this time, unlike for Snapdragon 600, in which there was a file called 8064boardgpu.c or similar, I forget!. This explicitly mentioned that if the SoC was identified as a 8064ab, then its GPU would be clocked at 450 MHz, instead of s600s default 400 MHz.

I mentioned to @Nerdtalker that I suspected that a Snapdragon 8974ab existed, mainly because of the unreleased Galaxy S4 s800 version listed in GLBenchmark 2.7, that was a clear 10%+ faster than any other S800 device, despite having been tested on Android 4.2.2, so not having the latest OpenGL 3.0 Qualcomm drivers, that come with 4.3, although I doubt that OpenGL ES 3.0 drivers will boost a OpenGL ES 2.0 benchmark, but they're probably the most up to date, however there was a big v14.0 driver update recently made available for 4.2.2 devices from Qualcomm, that already boosted results considerably.

Of course Samsung could have been overclocking their own device for the lol, but now Xiaomi have confirmed that there is a Snapdragon 800 - 8974ab variant, that uses the top binned CPU part @ 2.265 MHz (2.3) and a monster 100 MHz GPU overclock, from 450 Mhz (default) to 550 MHz. I do wonder if Xiaomi typoed the 550 MHz and meant 500 MHz, because >20% GPU boost is massive in a mobile SoC, and the benchmarks are only showing a ~10% gain.

From this I suspect that the Galaxy Note 3 is also using the 8974ab variant, which tallies with its first GLB 2.7 benchmark of 26 FPS (Offscreen), matching the previously mentioned GS4 S800, and faster than the other S800 devices tested so far.

What I'm most excited about is how well TSMC's 28nm HPM process is working in terms of performance / leakage, as this process node will be our bedrock well into 2014. As an example, just look at the voltage difference at 1.7 GHz between a Krait 300 built on the 28nm LP process, and a Krait 400 built on 28nm HPM, I know they are not the same core, but they must be quite similar... Also I chose the same PVS (bin) level of 5, as this obviously affects voltage.

28nm LP @ 1.7 GHz = 1075mv
28nm HPM @ 1.7 GHz = 900mv

If this holds true for other SoCs built on the HPM process, then I'd be very happy. It also means that Tegra 5 may actually stand a good chance of achieving its performance goals, despite no available die shrink.

TSMC 28nm LP
static struct acpu_level tbl_PVS5_1700MHz[] __initdata = {
{ 1, { 384000, PLL_8, 0, 0x00 }, L2(0), 875000 },
{ 1, { 486000, HFPLL, 2, 0x24 }, L2(5), 875000 },
{ 1, { 594000, HFPLL, 1, 0x16 }, L2(5), 875000 },
{ 1, { 702000, HFPLL, 1, 0x1A }, L2(5), 875000 },
{ 1, { 810000, HFPLL, 1, 0x1E }, L2(5), 887500 },
{ 1, { 918000, HFPLL, 1, 0x22 }, L2(5), 900000 },
{ 1, { 1026000, HFPLL, 1, 0x26 }, L2(5), 925000 },
{ 1, { 1134000, HFPLL, 1, 0x2A }, L2(14), 937500 },
{ 1, { 1242000, HFPLL, 1, 0x2E }, L2(14), 950000 },
{ 1, { 1350000, HFPLL, 1, 0x32 }, L2(14), 962500 },
{ 1, { 1458000, HFPLL, 1, 0x36 }, L2(14), 987500 },
{ 1, { 1566000, HFPLL, 1, 0x3A }, L2(14), 1012500 },
{ 1, { 1674000, HFPLL, 1, 0x3E }, L2(14), 1050000 },
{ 1, { 1728000, HFPLL, 1, 0x40 }, L2(14), 1075000 },
{ 0, { 0 } }

TSMC 28nm HPM

static struct acpu_level acpu_freq_tbl_2p3g_pvs5[] __initdata = {
{ 1, { 300000, PLL_0, 0, 0 }, L2(0), 750000, 72 },
{ 0, { 345600, HFPLL, 2, 36 }, L2(1), 750000, 83 },
{ 1, { 422400, HFPLL, 2, 44 }, L2(2), 750000, 101 },
{ 0, { 499200, HFPLL, 2, 52 }, L2(2), 750000, 120 },
{ 0, { 576000, HFPLL, 1, 30 }, L2(3), 750000, 139 },
{ 1, { 652800, HFPLL, 1, 34 }, L2(3), 760000, 159 },
{ 1, { 729600, HFPLL, 1, 38 }, L2(4), 770000, 180 },
{ 0, { 806400, HFPLL, 1, 42 }, L2(4), 780000, 200 },
{ 1, { 883200, HFPLL, 1, 46 }, L2(4), 790000, 221 },
{ 1, { 960000, HFPLL, 1, 50 }, L2(9), 800000, 242 },
{ 1, { 1036800, HFPLL, 1, 54 }, L2(10), 810000, 264 },
{ 0, { 1113600, HFPLL, 1, 58 }, L2(10), 820000, 287 },
{ 1, { 1190400, HFPLL, 1, 62 }, L2(10), 830000, 308 },
{ 1, { 1267200, HFPLL, 1, 66 }, L2(13), 840000, 333 },
{ 0, { 1344000, HFPLL, 1, 70 }, L2(14), 850000, 356 },
{ 0, { 1420800, HFPLL, 1, 74 }, L2(15), 860000, 380 },
{ 1, { 1497600, HFPLL, 1, 78 }, L2(16), 870000, 404 },
{ 1, { 1574400, HFPLL, 1, 82 }, L2(17), 880000, 430 },
{ 0, { 1651200, HFPLL, 1, 86 }, L2(17), 890000, 456 },
{ 1, { 1728000, HFPLL, 1, 90 }, L2(18), 900000, 482 },
{ 0, { 1804800, HFPLL, 1, 94 }, L2(18), 910000, 510 },
{ 0, { 1881600, HFPLL, 1, 98 }, L2(18), 920000, 538 },
{ 1, { 1958400, HFPLL, 1, 102 }, L2(19), 930000, 565 },
{ 0, { 2035200, HFPLL, 1, 106 }, L2(19), 940000, 596 },
{ 0, { 2112000, HFPLL, 1, 110 }, L2(19), 955000, 627 },
{ 0, { 2188800, HFPLL, 1, 114 }, L2(19), 965000, 659 },
{ 1, { 2265600, HFPLL, 1, 118 }, L2(19), 975000, 691 },
{ 0, { 0 } }
};

 

Yes that makes sense, but I would question the restriction to only 2 bins...by that meaning lg g2 which carries the - higher than xperia z ultra - binned 2.26ghz krait soc....would also have the same adreno 330 @ 550mhz...judging by results I have seen this seems unlikely.

Regarding galaxy note 3 - I would think judging by samsung having first dibs on s600 A/B soc dthat they would also have early access to the highest binned version.
-adding weight to this theory is the announcement of the exynos 5420 supporting 933mhz lpddr3.....matching the new s800 MSM8974AB SOC... and also aligning up quite nicely with the announced 3gb lpddr3 sticks @ 933mhz at around the same time and presumably to be used for both galaxy note 3 versions which also carry the same 3gb memory...memory speed has no been confirmed on either gn3 version of course, but im pretty sure only samsung has released 3gb lpddr3 memory ...due to the difficulty in its un even alignment of 2+1gb ..samsung solved this problem with some internal trickery that prevents performance degredation in such un balanced memory set ups as seen in pcs.

Also unless nexus 7-2 carries an underclocked s600 @ 1.5ghz... (possible) then there was 3 bins of s600...
-1.7ghz cpu 400mhz adreno 320
-1.9ghz cpu 450mhz adreno 320
-1.5ghz krait 300 adreno 320 @ 400 mhz..this was badged "s4 pro"- despite carrying krait 300?

 

During Qualcomm's Uplinq livestream, they mentioned that we have have "multi-core GPUs running upto 600 MHz" As you mentioned the Adreno 420, scheduled for launch in 2014.?. was specced at 500 MHz on the leaked roadmap, so an AB variant could easily be > 550 MHz.

I do suspect that Xiaomi made a typo with 550, and meant 500 MHz, as 20% is a huge boost, and the 8974ab benchmarks are 'only' around 10% faster, which means that if the 550 MHz figure is correct, they are being bottlenecked somewhere, which begs the question, why they wouldn't just clock it at 500 MHz. I'm assuming that GLB 2.7 scales linearly with frequency, which it does seem to. However as per my previous post the TSMC's HPM process seems to be performing well, so maybe they found more headroom in terms of frequency than originally expected.

 

Perhaps the answer to that +22% frequency only +10% 2.7 score is the memory frequency/bandwidth I mentioned before your comment.

For reference I dug up the press release although they mention they can clock the new 20nm lpddr3 "up to 2133" data rate..that doesnt stop them using 1866 data rate.
http://www.samsung.com/global/busin...am=&newsId=12979&page=&searchType=&rdoPeriod=

 

The LG G2 uses the top CPU bin @ 2.3 GHz, but it doesn't appear to an AB SoC. The AB is a GPU / Memory boost, so it seems there are 3 versions

2.15 (2.2) GHz CPU & 450 MHz GPU
2.265 (2.3) GHz CPU & 450 MHz GPU
&
2.265 (2.3) GHz CPU & 550 MHz GPU

We know that Qualcomm loves to confuse us, I don't know if all these variants are just done to binning, or if there are slightly different TDPs at work.

 

That would be interesting to found out, the truth will out soon enough, once a GNote 3 gets rooted, we'll be able to pull the GPU clockspeed. By the way the RAM in the Galaxy Note 3 isn't unbalanced like in a PC with 2GB & 1GB sticks. They are using 6 x 20nm class - 4Gb LPDDR3 (500MB) chips, probably in a PoP format, with 3 chips / 1.5 GB per memory channel

 

Yes I think we have a winner.
In regards to the note 3...samsung announced a little while ago it has stsrted mass production of new 10nm emmc 5.0 nand flash...said to boost performance by 10x over previous generation (edit micro sd cards). ..it will be interesting to see IO performance of gn3 to see if this has materilised.

http://www.tomshardware.com/news/eMMC-PRO-embedded-memory-eMMC-5.0-JEDEC-Samsung,23700.html

 

Yes sorry you are correct...I was going off of memory when this was announced and was referring to the symmetrical stacking they used to get even performance with an odd amount of ram...in this case 6x 512mb -3 on each side like you point out.

Edit: here are some benchmarks for note 3 that match mdp reference platform...
http://www.pianetacellulare.it/post...e-3-test-benchmark-confermano-potenza-e-v.php