Qualcomm Krait & MSM8960 @ AnandTech

[mboeller]

nice snapdragon 800 benchmarks at engadget:

http://www.engadget.com/2013/06/18/qualcomm-snapdragon-800-mdp-benchmarks/

2.7 T-Rex HD offscreen: 27fps

2.5 Egypt HD offscreen: 69fps

Also the other benchmarks show a nice improvement.

 

[french toast]

nice snapdragon 800 benchmarks at engadget:

http://www.engadget.com/2013/06/18/qualcomm-snapdragon-800-mdp-benchmarks/

2.7 T-Rex HD offscreen: 27fps

2.5 Egypt HD offscreen: 69fps

Also the other benchmarks show a nice improvement.

Wow, the phone is actually more impressive considering it beats the tablet in some benchies and likely has less tdp headoom.

Nothing in an android smartphone is going to get anywhere near those numbers.

 

[Laurent06]

nice snapdragon 800 benchmarks at engadget:

http://www.engadget.com/2013/06/18/qualcomm-snapdragon-800-mdp-benchmarks/

2.7 T-Rex HD offscreen: 27fps

2.5 Egypt HD offscreen: 69fps

Also the other benchmarks show a nice improvement.

The Octane score lower than S4 Octa is odd. The problem is that it's probably the most CPU intensive benchmarks of the ones listed. Perhaps a different browser?

 

[Ailuros]

nice snapdragon 800 benchmarks at engadget:

http://www.engadget.com/2013/06/18/qualcomm-snapdragon-800-mdp-benchmarks/

2.7 T-Rex HD offscreen: 27fps

2.5 Egypt HD offscreen: 69fps

Also the other benchmarks show a nice improvement.

Tsktsktsk....who would had expected that the S800 would end up roughly 50% faster in GLB2.7 than Tegra4...oh wait that one is irrelevant

 

[Laurent06]

Here is the Anandtech coverage: http://www.anandtech.com/show/7082/...ce-preview-qualcomm-mobile-development-tablet

 

[Nebuchadnezzar]

This graph is pretty much what you have to know: http://images.anandtech.com/doci/7082/Performance400.png

By now I'm pretty sure Qualcomm is quoting per-core numbers, so about 1350mW top per core.

That's indeed quite less than the A15 in the 5410. I don't see how the ARM designs will want to compete against such a huge power advantage, the A15's are just too inefficient.

 

[RedVi]

What's the current gen competitior that is on 32nm HKMG but 'laptop only'? I can only think of Exynos 5210, but that's quite obviously in a tablet...

 

[mboeller]

What's the current gen competitior that is on 32nm HKMG but 'laptop only'? I can only think of Exynos 5210, but that's quite obviously in a tablet...

The Chromebook maybe?

 

[Exophase]

This graph is pretty much what you have to know: http://images.anandtech.com/doci/7082/Performance400.png

By now I'm pretty sure Qualcomm is quoting per-core numbers, so about 1350mW top per core.

That's indeed quite less than the A15 in the 5410. I don't see how the ARM designs will want to compete against such a huge power advantage, the A15's are just too inefficient.

These graphs are annoying, I wish Qualcomm didn't decide to copy Samsung in reporting DMIPS curves. Yeah they work as a proxy for clock speeds but only if you get the DMIPS/MHz right...

The far right end of the curve looks like it goes to about 8700-8800. The far right of the Krait 300 curve about 6100-6200. If the former is 2.3GHz and the latter is 1.7GHz that'd mean it went from about 3.6 DMIPS/MHz to about 3.8 DMIPS/MHz. Both of these numbers seem too high (higher than Cortex-A15, although with Dhrystone all sorts of stupid things are possible). It's possible that the Krait 300 number is for 1.9GHz, which puts it at a more reasonable ~3.25 DMIPS/MHz. This is also more consistent with the Krait 200 number, which would be aorund 3.17 DMIPS/MHz for 1.5GHz; I doubt that Qualcomm's improvements from Krait 200 to 300 resulted in a huge boost to Dhrystone.

If Krait 400 really does nothing to the uarch but makes the L2 faster then it should perform about the same at the same clock speed with Dhrystone. That would mean that the far right end is going past 2.3GHz - up to either 2.4 or 2.7GHz. That would be good news for Qualcomm, of course.

Also lolz at "more than HALF THE POWER".. good job Qualcomm, I'm pretty sure you meant less than. The (most likely) Exynos numbers also look dubious, maybe inflated due to other system design considerations that went into Chromebook, which it most likely is. 5250 is obviously not a laptop-only SoC, that denotation probably means they only tested it in a laptop.

 

[french toast]

These graphs are annoying, I wish Qualcomm didn't decide to copy Samsung in reporting DMIPS curves. Yeah they work as a proxy for clock speeds but only if you get the DMIPS/MHz right...

The far right end of the curve looks like it goes to about 8700-8800. The far right of the Krait 300 curve about 6100-6200. If the former is 2.3GHz and the latter is 1.7GHz that'd mean it went from about 3.6 DMIPS/MHz to about 3.8 DMIPS/MHz. Both of these numbers seem too high (higher than Cortex-A15, although with Dhrystone all sorts of stupid things are possible). It's possible that the Krait 300 number is for 1.9GHz, which puts it at a more reasonable ~3.25 DMIPS/MHz. This is also more consistent with the Krait 200 number, which would be aorund 3.17 DMIPS/MHz for 1.5GHz; I doubt that Qualcomm's improvements from Krait 200 to 300 resulted in a huge boost to Dhrystone.

If Krait 400 really does nothing to the uarch but makes the L2 faster then it should perform about the same at the same clock speed with Dhrystone. That would mean that the far right end is going past 2.3GHz - up to either 2.4 or 2.7GHz. That would be good news for Qualcomm, of course.

Also lolz at "more than HALF THE POWER".. good job Qualcomm, I'm pretty sure you meant less than. The (most likely) Exynos numbers also look dubious, maybe inflated due to other system design considerations that went into Chromebook, which it most likely is. 5250 is obviously not a laptop-only SoC, that denotation probably means they only tested it in a laptop.

What the heck has qualcomm been doing to that adreno 330?? The performance improvements are ridiculous!

Seriously..what have they done to it? :/

 

[ams]

What the heck has qualcomm been doing to that adreno 330?? The performance improvements are ridiculous!

Seriously..what have they done to it? :/

The Adreno 330 in S800 is pretty impressive. The SoC die size of S800 is much bigger than S600 (roughly 50% bigger IIRC), so Adreno 330 must have much larger die size area and much more shader execution units compared to Adreno 320. On a side note, it looks like Engadget made a mistake when reporting 3dmark Ice Storm scores for Adreno 330. They listed the "Extreme" score but compared it to the standard score for S600 and other devices.

 

[Ailuros]

I've seen occassional quirks with 3dmark here and there; so I'm not really surprised.

 

[Helmore]

The Adreno 330 in S800 is pretty impressive. The SoC die size of S800 is much bigger than S600 (roughly 50% bigger IIRC), so Adreno 330 must have much larger die size area and much more shader execution units compared to Adreno 320. On a side note, it looks like Engadget made a mistake when reporting 3dmark Ice Storm scores for Adreno 330. They listed the "Extreme" score but compared it to the standard score for S600 and other devices.

Where did you get those numbers and are you even comparing similar dies here? As far as I'm aware, the Snapdragon 800 SoC includes the baseband while for their Snapdragon 600 SoC the baseband is a separate chip. Of course the S800 is going to be bigger then, provided those numbers are accurate.

 

[ams]

Where did you get those numbers and are you even comparing similar dies here? As far as I'm aware, the Snapdragon 800 SoC includes the baseband while for their Snapdragon 600 SoC the baseband is a separate chip. Of course the S800 is going to be bigger then, provided those numbers are accurate.

I seem to recall that S600 die size is ~ 80 mm^2 while S800 die size is ~ 120 mm^2 (based on some leaked package size specs that were posted here a while ago), but I can't say for sure. Good point about the integrated modem in S800, but it is very unlikely that the integrated modem would take up all the extra die size area available in S800. Based on the benchmark results and the presumption that GPU clock operating frequencies haven't changed a huge amount, it is pretty clear that Adreno 330 is physically a much larger GPU than Adreno 320, with more shader execution units too.

 

[french toast]

I seem to recall that S600 die size is ~ 80 mm^2 while S800 die size is ~ 120 mm^2 (based on some leaked package size specs that were posted here a while ago), but I can't say for sure. Good point about the integrated modem in S800, but it is very unlikely that the integrated modem would take up all the extra die size area available in S800. Based on the benchmark results and the presumption that GPU clock operating frequencies haven't changed a huge amount, it is pretty clear that Adreno 330 is physically a much larger GPU than Adreno 320, with more shader execution units too.

Agree totally, ive noticed fill rate have just about doubled over the higher binned s600 in gs4...so wouldnt that mean double the TMUs? If thats the case then could we expect double the shaders also?

 

[Ailuros]

ALUs are usually relatively cheap; 320 has 8 TMUs and given on the past leaked roadmap the 330 should have the same amount too. Assuming it has 50% more ALUs/FLOPs compared to the 320 (note that the Galaxy S4 320 is most likely also clocked at 450MHz), it's obviously nowhere near adding 50% more die area to the entire SoC area estate.

By the way not to burst anyone's bubble but Anand's test are on a development platform and usually a best case scenario. Final devices may vary depending on the manufacturers power targets mostly.

 

[mboeller]

I seem to recall that S600 die size is ~ 80 mm^2 while S800 die size is ~ 120 mm^2 (based on some leaked package size specs that were posted here a while ago), but I can't say for sure. Good point about the integrated modem in S800, but it is very unlikely that the integrated modem would take up all the extra die size area available in S800. Based on the benchmark results and the presumption that GPU clock operating frequencies haven't changed a huge amount, it is pretty clear that Adreno 330 is physically a much larger GPU than Adreno 320, with more shader execution units too.

look at the Tegra4 LTE chip, the i500.

It was said that the i500 is only 40% as large as a normal baseband-chip. Despite that it has roughly an die-size of 1/4 of the Tegra4 itself, or maybe 20mm². Therefore a normal baseband-chip could be 50mm² large. Therefore all of the 40mm² size difference between S600 and S800 could be due to the LTE-system.

Link: http://www.engadget.com/2013/01/06/nvidia-tegra-4-official/

 

[mboeller]

By the way not to burst anyone's bubble but Anand's test are on a development platform and usually a best case scenario. Final devices may vary depending on the manufacturers power targets mostly.

Yes, but look at the engadget link. They had tested 2 development devises, one tablet and one smartphone. The smartphone should have been already quite restricted regarding power consumption with a 1500mAh battery.

 

[Jubei]

I dont see why the GPU results cant be done in a smartphone power envelope, its only a modest clock bump over the original 320 and as Ailuros mentioned the S4 is suspected of hitting the same frequency on LP node already

Is there any task on a smartphone that would need the CPU core/cores to hit 2,3 GHz? Im guessing that top number is only a burst speed before dropping down quite fast

 

[Ailuros]

I dont see why the GPU results cant be done in a smartphone power envelope, its only a modest clock bump over the original 320 and as Ailuros mentioned the S4 is suspected of hitting the same frequency on LP node already.

I'd think that for any manufacturer that considers to constraint power somewhat the first candidate to consider would be the CPU. Theoretically I'd say that you save a lot more power going down to say 2050MHz for the CPU in such a case then 45MHz for the GPU.

Is there any task on a smartphone that would need the CPU core/cores to hit 2,3 GHz? Im guessing that top number is only a burst speed before dropping down quite fast

You could ask the same question for Exynos5410 and its quad A15@1.6GHz or any other recent SoC. For that I'd say that Intel's "SDP" could end up being a far more accurate comparison point than real TDP:

http://www.fudzilla.com/home/item/31532-qualcomm-aims-at-25-to-3w-tdp-for-phones

Qualcomm feels that 45 degrees Celsius (113 degrees Fahrenheit) would be the maximum allowed temperature for SoCs such as the Snapdragon 600 or 800. Furthermore we learned that 2.5 to 3W would be the upper limit for Thermal dissipation of the whole SoC in a mobile phone, while the max thermal power of a tablet could go up to 5W.