Qualcomm Krait & MSM8960 @ AnandTech

[french toast]

Interesting. .htc is going to push a patch for halti in its htc one.
http://m.theinquirer.net/inquirer/news/2257651/htc-one-will-get-opengl-es-30-support-video

Obviously htc must have some yet to be revealed game optimisation deals else the api would be useless.

Backs up what I was saying a month or so ago about halti making an appearance sooner than many think.

Also in the same demo.. (fortress fire) it appears to be the same demo as used at CES for s800..but running at what appears to be full speed and fully featured...meaning s600 was shown at CES?..
..that demo had adreno clocked at 450mhz..could be that it was a 320 after all backing up what anand was saying. .or could still be s800?.

Either way s800 is obviously capable of cooking up better graphical treats than fortress fire then if it sails on s600.

Confusing is the demo guy banging on about geometry shaders being used in halti..I thought halti passed on that feature with it still only being a DX10 feature?? Or does he point to adreno carrying geometry shading on top of halti?

Either hes getting his wires crossed or adreno is more capable than we thought. ..id bet on the former.

 

[Deleted member 13524]

[16.8 fps (GalaxyS4) / 14.2 fps (Nexus4)] * 400MHz = 473MHz.

It's possible that they're all clocked at 400MHz and the difference comes only from the 50% higher memory bandwidth in the S4.

FWIW, the people at xda are reporting that custom kernels can set the Adreno 320 in the HTC One up to 487MHz, from its original 400MHz limit.

 

[Ailuros]

What about the difference in memory bandwidth? The fact that there actually might be a small revision in the GPU (you never know)? Different drivers?

Legitimate points; counter-point would be that the S600 CPU for the S4 is coincindentially clocked at 1890MHz. When your power budget allows you to increase CPU frequency by N persentage what speaks against a say ~20% frequency increase for the GPU also?

It's possible that they're all clocked at 400MHz and the difference comes only from the 50% higher memory bandwidth in the S4.

FWIW, the people at xda are reporting that custom kernels can set the Adreno 320 in the HTC One up to 487MHz, from its original 400MHz limit.

The 2nd paragraph is a highly interesting tidbit

 

[Turbotab]

The biggest weakness of GLbenchmark, it doesn't / cannot filter for overclocked GPUs. A lot of the Android results are skewed because of this, at least on iOS we know GPU overclocking isn't a worry

If we take Anandtech's freezer run of the Nexus 4, as the best possible result for a stock Nexus 4, as it stops the device being throttled artificially early, due to strangely low thermal limits, you get 30 FPS in Egypt HD offscreen. The HTC one scores 34 FPS on the same test, even after multiple runs its score is stable, so we know they fixed their early thermal throttling issue.

From the HTC One kernel source we know that it's GPU is clocked at 400 MHz, as per the Nexus 4, so there are more than likely undisclosed uArch changes to the 'speed enhanced' Adreno 320s in the Snapdragon 600s.

Interestingly, the kernel sources also revealed that if a SoC is identified as 8064ab, then the GPU's max clock is bumped to 450 MHz, however the 8064ab is referred to as an APQ8064ab in other parts of the source, so maybe Qualcomm are just boosting their reference devices

Drivers are a big factor, it appears that the HTC One is running the oldest drivers "AU 4.1", compared to "AU 6.0" for N4 and "AU 14.0" for the Samsung SGH-I337M / GS4. On Qualcomm's dev site, the latest drivers are for 4.2.x ROMs

An Experia Z (S4 Pro), with the latest OpenGL ES 3.0 V@14.0 drivers, briefly outscored the HTC One on 2.7 T-Rex HD offscreen (The result has now disappeared?) The Xperia Z went 756 Frames to 846 frames (higher is better), ~12% improvement, HTC One currently scores 826.

There is definitely a perf increase on GLB with the latest drivers, but until HTC update The One to 4.2.2 we aren't playing on a level field. IMO the Galaxy S4's GPU is still clocked at 400 MHz, and once we've equalized the driver environment, The One & GS4 will score almost the same on GLB 2.7, perhaps with a slight edge to the GS4, due to its superior memory bandwidth, anyway we'll soon find out!

 

[Ailuros]

The biggest weakness of GLbenchmark, it doesn't / cannot filter for overclocked GPUs. A lot of the Android results are skewed because of this, at least on iOS we know GPU overclocking isn't a worry

Make out of that a petition to Kishonti to try to get behind that and I'll sign it instantly :smile:

 

[french toast]

@turbotab..
Very interesting find, this would fit in with my suspicion that the CES fortress fire demo (advertised as s800 seen running @450mhz) could in fact be just s600 in demo mode.

I also think this is a simple driver issue and s600/s4pro both run at 400mhz, demo mode was obviously to imitate s800 performance as they might not have had stable silicon to show.

 

[Helmore]

@turbotab..
Very interesting find, this would fit in with my suspicion that the CES fortress fire demo (advertised as s800 seen running @450mhz) could in fact be just s600 in demo mode.

I also think this is a simple driver issue and s600/s4pro both run at 400mhz, demo mode was obviously to imitate s800 performance as they might not have had stable silicon to show.

I don't know why they would do that when they obviously had stable Snapdragon 800 chips on hand. At least, I'm not aware of S600 being capable of decoding 4K video, which is something they demoed, along with Dolby Headphone for S800 and the new baseband in the S800 chipset. They have also been sampling S800 chips since the start of the year, something they wouldn't be able to do without stable S800 chips.

 

[french toast]

I don't know why they would do that when they obviously had stable Snapdragon 800 chips on hand. At least, I'm not aware of S600 being capable of decoding 4K video, which is something they demoed, along with Dolby Headphone for S800 and the new baseband in the S800 chipset. They have also been sampling S800 chips since the start of the year, something they wouldn't be able to do without stable S800 chips.

Thats very true indeed, didnt think of that part that would definately peg 320 at 400mhz then, especially considering 330 is based off 320, and 320 can only be pushed to 480mhz max, HKMG+higher binning would help, but i cant see s800 going over 500mhz, not with batterylife concerns.

So 320=400mhz 330=450mhz (as leaked and seen in fortress fire demo) any ideas then on just what qualcomm have added to adreno 330 to make up the claimed performance improvements? (2* compute, 50% graphics).

I have not the deep knowledge to even hypothesis.

 

[Nebuchadnezzar]

The Galaxy S4 kernel sources are out, I'm not too familiar with the Qualcomm driver hierarchies, so if somebody can point out where to check the GPU frequency, I'll post it. If not, this is a heads-up for others.

 

[french toast]

The Galaxy S4 kernel sources are out, I'm not too familiar with the Qualcomm driver hierarchies, so if somebody can point out where to check the GPU frequency, I'll post it. If not, this is a heads-up for others.

Thanks.

Gsmarena has redone its benchmark suite for GT-i9505 with final shipping firmware for those whos interested:
http://www.gsmarena.com/samsung_galaxy_s4-review-914p6.php
Highlights included the first proper optimised browser for krait (outside of windows phone) with a scorching 810ms..

Geekbench 3227
GL benchmark 2.5 1080p offcreen : 40fps
Linpack multithread 788
PI 132.

Some find most mobile benchmarks on here a pointless exercise, still they are at least a reference point between SOC generations if nothing else.

Edit, here is a comparison on GL bench of the two best galaxy s4 devices, one with s600 and one with exynos 5410.
http://gfxbench.com/compare.jsp?D1=...337M&D3=Apple+iPad+4&D4=Apple+iPhone+5&cols=4

They are comparable, SGX544 mp3 beats adreno in GL 2.5 yet adreno turns the tables on 2.7 by a similar margin.

 

[Nebuchadnezzar]

I actually bothered to screw around a bit with the Qualcomm sources; I don't think anybody ever mentioned that the L2 cache is only running at a maximum of 1188MHz.

Also seems to me the 320 is clocked at 450MHz in the S4 / 8064ab chip. It overrides the 400MHz value of non-ab chips on the max step.

void __init apq8064_init_gpu(void)
{
	unsigned int version = socinfo_get_version();

	if (cpu_is_apq8064ab())
		kgsl_3d0_pdata.pwrlevel[0].gpu_freq = 450000000;
.....

 

[Turbotab]

Well, Brian Klug has confirmed that on the actual devices, The One has a SoC ID = 109, and 'OTHER' device has an ID =153. From the socinfo.c and GPU board files from the kernel sources, we know the following.

ID 109 = 8064 = 400 MHz GPU

ID 153 = 8064ab = 450 MHz GPU

Apparently there is an interesting rationale behind this. Anyway, long story made short, if you're a spec whore the GS4's 1.9 GHz CPU & 450 MHz > 1.7 GHz & 400 MHz

I'd love to know how much Qualcomm is charging HTC & Samsung for each SoC.

 

[Ailuros]

Before the S4 launch I had heard a rumor that Samsung adjusted frequencies on the S600 to have as small as possible differences to its 5410. The info was clear that the 544 won't clock at 533 but at 480 and the Adreno in the 320 fairly close. Now make out of it what you want but considering how reliable the source has proven itself in the past to me I never doubted it one second.

 

[Nebuchadnezzar]

Before the S4 launch I had heard a rumor that Samsung adjusted frequencies on the S600 to have as small as possible differences to its 5410. The info was clear that the 544 won't clock at 533 but at 480 and the Adreno in the 320 fairly close. Now make out of it what you want but considering how reliable the source has proven itself in the past to me I never doubted it one second.

Not to rain on your source, but I see no evidence of this, except for a option to use or not the 532 step in the source:

https://github.com/AndreiLux/Perseu...vr/services4/system/exynos5410/sec_dvfs.c#L56

Which is however used by default:

https://github.com/AndreiLux/Perseus-UNIVERSAL5410/blob/samsung/drivers/gpu/pvr/Makefile#L82

I reckon you said that they had reduced it to 480 due to power issues, doesn't that contradict the performance claim? Either that or the released sources are not finalized and representative of shipping software.

 

[mboeller]

according to qualcomm the Adreno330 should be 59 faster than the ?? (I think they mean the Adreno 320) in GLBenchmark 2.5

http://www.engadget.com/gallery/qualcomm-media-workshop-beijing-2013/5841841/

If I could read the slides the presentation would be even more interesting

It seems they use a onchip GPU-Buffer for the Adreno330:
http://www.engadget.com/gallery/qualcomm-media-workshop-beijing-2013/5841849/

Also ~8400DMips @ ~1,35W is not bad either. It seems the 28nm process is much much better than the 32nm process. At least it seems so according to one slide.

 

[Rys]

Adreno has had dedicated memory since the beginning (usually referenced as GMEM), so that's not a new development for 330.

 

[Nebuchadnezzar]

Adreno has had dedicated memory since the beginning (usually referenced as GMEM), so that's not a new development for 330.

Are we talking about dedicated memory on the SoC PoP memory or actual on-die buffers?

 

[Exophase]

Are we talking about dedicated memory on the SoC PoP memory or actual on-die buffers?

GMEM is the tile memory, so on-die buffers. You can read about it in the user manuals for i.MX5 series chips that licensed the core back when it was still z430, before it was sold to Qualcomm. i.mx51 had 128KB while i.mx53 had 256KB.

The slide mboeller linked to definitely looks like tile memory, since it points what looks like a tile on a partially rendered tiled scene. A couple slides later you can see it calls out GMEM.

 

[Wishmaster]

GMEM is the tile memory, so on-die buffers. You can read about it in the user manuals for i.MX5 series chips that licensed the core back when it was still z430, before it was sold to Qualcomm. i.mx51 had 128KB while i.mx53 had 256KB.

The slide mboeller linked to definitely looks like tile memory, since it points what looks like a tile on a partially rendered tiled scene. A couple slides later you can see it calls out GMEM.

Maybe they significantly increased its size and that's why they showed it off as something unique. Although one could wonder what significantly could mean

 

[Helmore]

The z430 was rebranded as the Adreno 200 and the z460 the Adreno 220, right?