Huawei Ascend D/ Hisilicon K3V2 - Quadcore with 16-core graphics

french toast said:
Well firstly there has to be some funky marketing going on as they are sure as hell refer to ALU's/shaders..not full GPU cores with TMU's/rasterizers etc..
Even if it is a proper quad GPU like A5X or even a 'compromised' 'quad' like T604..the results whilst being great aren't that great..unless they have used a low clocked variant to keep power consumption down?.:???:
Click to expand...

The K3V2 seems to use LP manufacturing process, where as the Tegra3 used Performance manufacturing process, BUT used LP for the 5th companion CPU. This is why Huawei is claiming that there K3V2 uses 30% less power.

The marketing funky claims are based on probable situations where the Tegra3 uses its main CPU's, compared to the K3V2's main CPU's. Claim 30% less power. Bingo.

Now, when the Tegra3 jumps to the 5th core, the K3V2 will still be running on one of its main CPU's.

So, in low power situations, we can assume that the K3V2 will use more power on one of its main CPU's.

And all the other CPU's are probably power gated when not needed. Question becomes who's strategy will be more power efficient.

For heavy loads or anything that needs to run on its main CPU's, i assume the K3V2 will be more power efficient.

For light loads... your guess is as good as mine.

french toast said:
How do we know for sure that A5X is 45nm? thought it was assumed to be on Sammy's 32nm process?
Click to expand...

http://www.anandtech.com/show/5686/apples-a5x-floorplan

There first gamble where based on a guess, and it looks like they where wrong.

french toast said:
Ditto with the K3V2 how do we know the manufacturing tech?
Click to expand...

Because the manufacture said so? Its 40nm ... And to be honest, its not a far stretch. If they said 28nm, there will have been more doubts about that, as it seem 28nm has plenty of problems.

The said that in the future, they are looking at A15 @ 28nm, and the current K3V2 @ 28nm.

But it seems clear that 28nm is not ready for prime time yet. That might also explain some of those numbers the Krait. And how we STILL have not gotten any power usage indicator for the Krait. If the claims are correct that 28nm is leaking too much power, that might explain some things?

Lazy8s said:
The workloads of some of these benchmarks and the variability of some of the comparative scores prevents me from buying some of the performance claims here.

HiSilicon says the texel fill is 1.3 billion per second and presumably IMR (especially with all of the SGI name dropping), so this SoC wouldn't have made for a better A5X to drive the new iPad display.

Some of the other theoretical performance figures sound nice, though.
Click to expand...

Immature drivers? We can assume that the drivers that Apple uses are already extremely optimized now.

That Vivante GC4000 seems to be a rather "unknown" asset up to now, so its very possible that there drivers still have more power underutilized?

Who knows, its just a gamble, but it can explain the difference between some of the results, and the claims.

If we look at for instants Intel, they have been the king of problematic performance for there drivers ( for integrated graphics ).
 
french toast said:
Everywhere is advertising that Adreno 320 is now 4x Adreno 225..instead of only 190% which was floating around before on the other thread...that seems more like it and that would destroy A5X with some good bandwidth (which as IMR instead of TBDR its gonna need)
Click to expand...

according to this presentation slide from Qualcomm it is 3-times the shader performance not 4-times:

7d493cb790.jpg


the complete presentation is here (scroll down):
http://www.notebookcheck.com/Qualcomm-Roadmap-fuer-die-Snapdragon-S4-Prozessoren.71190.0.html
 
Last edited by a moderator:
3dcgi said:
That's a strange thing to market considering multiple competitors could say the same thing.
Click to expand...
Hell, pretty much all of raster 3D graphics could be said to be descended from Silicon Graphics. I remember how their newfangled raster workstations were ridiculed due to their jagged lines (the screens were 1024x768). Of course, they soon updated their systems with anti-aliased line drawing and proceeded to take over the market from Evans&Sutherland and other vector graphics vendors, and the rest is history.
SGI->3Dfx->nVidia is just the gaming branch of SGI genealogy.
 
As mentioned, the GPU here clearly doesn't have sixteen cores in a classical sense since convention would then be to have sixteen TMUs (assuming the conventional one texel per clock here); I'm guessing this GPU actually has just four TMUs and a 325 MHz clock.

So, maybe a 4:1 ALU:TMU ratio, like an SGX554MP2.
 
Last edited by a moderator:
Lazy8s said:
As mentioned, the GPU here clearly doesn't have sixteen cores in a classical sense since convention would then be to have sixteen TMUs (assuming the conventional one texel per clock here); I'm guessing this GPU actually has just four TMUs and a 325 MHz clock.

So, maybe a 4:1 ALU:TMU ratio, like an SGX554MP2.
Click to expand...

^^Agreed.
 
In the 3rd post in this thread, I put a link to a video with the chip designer where he agreed with the interviewer that it was NOT mali, vivante or powervr.

In relation to rockchip moving from vivante to Mali, they also took a powervr license recently.
 
tangey said:
In the 3rd post in this thread, I put a link to a video with the chip designer where he agreed with the interviewer that it was NOT mali, vivante or powervr.

In relation to rockchip moving from vivante to Mali, they also took a powervr license recently.
Click to expand...

I missed that..well who knows then:oops:
 
The Ascend D quad smartphone with its K3V2 SoC scores 6470 (57.3 fps) on Egypt 720p Offscreen, putting it at a similar level of performance to Adreno 225 based phones in that benchmark.

The GL Environment reported to the benchmark doesn't necessarily verify the GPU IP partner as Vivante, but I'm suspecting Vivante is now pretty much a lock.

And technically, in that video interview posted earlier with the Hisilicon engineer, he only cut in to deny the suggestion of who the IP partner was after PowerVR and Mali were mentioned. When Vivante was suggested first, the engineer looked as if he might be preparing to qualify that suggestion instead of denying it before the interviewer moved on to the other suggestions a fraction of a second later. Hard to tell...

Before, I was figuring for a four TMU design and a 325 MHz clock, and that still seems like a good fit. I'm wondering, though, whether any of these designers are actually trying a design with few or even a single TMU while trying to run the GPU at the CPU's clock rate.
 
Last edited by a moderator:
Lazy8s said:
The Ascend D quad smartphone with its K3V2 SoC scores 6470 (57.3 fps) on Egypt 720p Offscreen, putting it at a similar level of performance to Adreno 225 based phones in that benchmark.

The GL Environment reported to the benchmark doesn't necessarily verify the GPU IP partner as Vivante, but I'm suspecting Vivante is now pretty much a lock.
Click to expand...

I note with interest in the list of gl calls supported, the following 2

GL_APPLE_texture_format_BGRA8888
GL_APPLE_texture_max_level

Wonder what they are about, would seem to be a clear reference to apple products. Anything to do with emulating PVRTC ?
 
tangey said:
I note with interest in the list of gl calls supported, the following 2

GL_APPLE_texture_format_BGRA8888
GL_APPLE_texture_max_level

Wonder what they are about, would seem to be a clear reference to apple products. Anything to do with emulating PVRTC ?
Click to expand...

Not likely. Both of the extensions are present in the OpenGL ES API Registry, and both of them have their IP status described as "No known IP issues/claims"; presumably, they just picked these extensions to be able to expose some of their functionality. The extension list also contains "GL_ARM_rgba8" and "GL_IMG_read_format", both of which are also present in the Registry.

All of these extensions seem rather mundane, though; presumably, if they were doing things like emulating PVRTC, they would have exposed "GL_IMG_texture_compression_pvrtc", like that Chainfire3D driver some time ago.


Notably, the extension list includes "GL_VIV_shader_binary", which strongly indicates the presence of a Vivante core.
 
Yep it's a Vivante GC4000: http://www.glbenchmark.com/phonedetails.jsp?D=Huawei+U9510&benchmark=glpro21

An octal core GPU as Vivante calls it or better 8*Vec4. If the GC4000 can't beat the SGX543MP2 in the iPad2 I wonder how freescale's iMX6 with a GC2000 is going to as they claimed themselves in some past marketing material.

http://www.vivantecorp.com/Product_Brief.pdf

In any case from the Huawei U9510 results I can't see any 200M Tris nor any 2.5 GPixels happening on that one.

Question remains where the heck the initial "16 cores" come from; counting alu lanes as cores would be 32 cores for 8*Vec4.

By the way the MSAA performance on Vivante cores hasn't improved by a lot. I doubt they're even tile based. The latest Mali400MP4@>400MHz http://www.glbenchmark.com/phonedetails.jsp?D=Samsung+GT-N8013&benchmark=glpro21 show that MSAA is for free.
 
Ailuros said:
The latest Mali400MP4@>400MHz http://www.glbenchmark.com/phonedetails.jsp?D=Samsung+GT-N8013&benchmark=glpro21 show that MSAA is for free.
Click to expand...

the (I assume 400MHz clocked) S2 test has about 8% less fill rate than the above test, which would tie in with the clock difference. The S2 tests beats it significantly on all the triangle tests, I assume we can therefore conclude from this that the triangle tests are screen resolution dependant and therefore are not terribly use for cross device comparision.

http://www.glbenchmark.com/compare....amsung GT-N8013&D2=Samsung GT-i9100 Galaxy S2
 
tangey said:
the (I assume 400MHz clocked) S2 test has about 8% less fill rate than the above test, which would tie in with the clock difference. The S2 tests beats it significantly on all the triangle tests, I assume we can therefore conclude from this that the triangle tests are screen resolution dependant and therefore are not terribly use for cross device comparision.

http://www.glbenchmark.com/compare....amsung GT-N8013&D2=Samsung GT-i9100 Galaxy S2
Click to expand...

The Samsung S3 results are gone again and while the Huawei results also disappeared they're now back again slightly below iPhone4S. In any case even on S2 4xMSAA is for free in the Egypt test.
 
Huawei have just been announced as a new IMG licensee including a Rogue core license. This applies to the going forward position so discount any PowerVR inside existing SoCs.
 
Which leaves one mystery partner for the list of so far 8 Rogue licensees to complete.
 
You must log in or register to reply here.
Back
Top