PowerVR Series 6 now official

[tangey]

Eetimes article suggests G6200 in the MT8135 is ~80gflops

http://www.eetimes.com/document.asp?page_number=2&doc_id=1319064&image_number=1

80Gflops would sound more plausible to me, given its getting within 15% of the ipad4 on gl2.5. I understand Gflops doesn't necessary compute to graphics performance, but there must be a certain correlation.

Fill rate figures would be interesting.

 

[Ailuros]

80Gflops would sound more plausible to me, given its getting within 15% of the ipad4 on gl2.5. I understand Gflops doesn't necessary compute to graphics performance, but there must be a certain correlation.

Clock a G6200 at 625MHz and you have your 80 GFLOPs; I'd pray your expectations aren't for real otherwise it would mean quite crappy efficiency values. What's better to need 4x times more FLOPs for a ~4.0x increase or an 8x times increase for a 4x increase?

Fill rate figures would be interesting.

Sure if you go for 625MHz then 2.5 GTexels. I still prefer though the 1.3 GTexels scenario for obvious reasons

 

[Ailuros]

Nope sorry gentlemen:

http://www.anandtech.com/show/7160/asus-memo-pad-hd7-review/3

The 544MP1 is clocked at 286MHz in the HD7 implementation of the MT8125, yielding peak FP performance of just under 9 GFLOPS.

4*8*0.286 = 9.15 GFLOPs
4*9*0.286 = 10.30 GFLOPs

9.15-10.30 * 4x = =/<40 GFLOPs for the G6200 (or the graph in IMG's relevant PR blurb is off base); meaning that the G6200 frequency is somewhere in between 285 and 315MHz, with the first being likelier.

Now that MT8125 scores in GLB2.5 offscreen 9.9 fps according to Anand; compared to the supposed MT8135 a 4.6x times performance difference. Can't be sure but due SGX not necessarily digesting all that well alpha tests the difference in GLB2.7 could touch the 5x times difference.

 

[ltcommander.data]

Unless they've changed something last minute (what for exactly?) all so far variants are DX10.0. You can actually ask that PR gentleman in the blog link in the notes underneath if it's 10.0 or 10.1 after all.

So they've now updated the blog. G6200 is OpenGL 3.x and DX10.0 as traditionally known. They've also moved the target for future Rogues from DX11.1 to DX11.2. Maybe support for Tiled Resources/Sparse Textures will be included too.

 

[Ailuros]

So they've now updated the blog. G6200 is OpenGL 3.x and DX10.0 as traditionally known. They've also moved the target for future Rogues from DX11.1 to DX11.2. Maybe support for Tiled Resources/Sparse Textures will be included too.

Considering DX11.1 to DX11.2 is only an incremental change I doubt they "moved" much if anything; most of it is just checkbox stuff anyway. I wouldn't expect any DX11 game to be playable on such a GPU unless in the case of Rogue it's got a shitload of clusters.

By the time you can play a DX11 game on any such device I'd expect GPUs by then to be OGL_ES4.0 if not 5.0 entirely.

Other than that since you asked him it's natural that he saw the probable typo and corrected it for DX10.0. DX11.2 is new however not really a surprise either; by the way I like how users are cornering Alex over and over again with performance questions. Unless my instincts are seriously damaged I don't see the MT8135 G6200 having 80GFLOPs. I was for a period of time seriously worried about its efficiency but so far indications point at a very good perf/GFLOP ratio yes admittedly only in GLB2.5 so let's wait and see.

Qualcomm is lending devboards to Anand for testing purposes; can't Mediatek do the same?

 

[mboeller]

http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.ddi0438i/index.html

They are apparently up to r4 on the A15 too.

Thanks! That I didn't know.

 

[Plamensito]

PowerVR SGX554 MP1 притежава 8USEE2 = 19.2GFLOPS per 0.300MHz.

MP2 : 38.4
MP4 : 76.8

Ако приемем, че PowerVR 6 Rogue притежава 16USSE2, то GFLOPS ще се изчисли както следва : 16 х 4 MAD's x 2 ALU x 0.300 = 38.4 GFLOPS

Assuming that PowerVR 6 Rogue has 16USSE2, it GFLOPS will be calculated as follows:
16 x 4 MAD's x 2 ALU x 0.300 = 38.4 GFLOPS per 1 Core.

MP2 = 76.8
MP4 = 153.6

Now add the additional scalar of 1.125 (9 ÷ 8) :

PowerVR SGX554 MP1 = 21.6
PowerVR SGX554 MP2 = 40.2
PowerVR SGX554 MP4 = 86.4

PowerVR G6200 MP1 = 43.2
PowerVR G6200 MP2 = 86.4

Note: PowerVR SGX554 MP4 for iPad 4 operates 0.280MHz (80GFLOPS).

 

[Plamensito]

PowerVR SGX554 MP1 притежава 8USEE2 = 19.2GFLOPS per 0.300MHz.

MP2 : 38.4
MP4 : 76.8

Ако приемем, че PowerVR 6 Rogue притежава 16USSE2, то GFLOPS ще се изчисли както следва : 16 х 4 MAD's x 2 ALU x 0.300 = 38.4 GFLOPS

Assuming that PowerVR 6 Rogue has 16USSE2, it GFLOPS will be calculated as follows:
16 x 4 MAD's x 2 ALU x 0.300 = 38.4 GFLOPS per 1 Core.

MP2 = 76.8
MP4 = 153.6

Now add the additional scalar of 1.125 (9 ÷ 8) :

PowerVR SGX554 MP1 = 21.6
PowerVR SGX554 MP2 = 43.2
PowerVR SGX554 MP4 = 86.4

PowerVR G6200 MP1 = 43.2
PowerVR G6200 MP2 = 86.4

Note: PowerVR SGX554 MP4 for iPad 4 operates 0.280MHz (80GFLOPS).

In other words, MediaTek 8125 operates at 280 MHz :

16USSE2 x 4MADS x 2 ALU x 2 Core x 0.280 x 1.125 = 80GFLOPS!

 

[Ailuros]

In other words, MediaTek 8125 operates at 280 MHz :

16USSE2 x 4MADS x 2 ALU x 2 Core x 0.280 x 1.125 = 80GFLOPS!

So according to you a G6200 has both 2 clusters and 2 cores? I'd love a piece of documentation for that one before I further bent over laughing. Oh and you can forget what you folks call "scalar" on that one.

Just for sanity's sake: give me one good reason why someone would want to use a G6200"MP2" and not go straight for a G6400 instead.

 

[Plamensito]

So according to you a G6200 has both 2 clusters and 2 cores? I'd love a piece of documentation for that one before I further bent over laughing. Oh and you can forget what you folks call "scalar" on that one.

Just for sanity's sake: give me one good reason why someone would want to use a G6200"MP2" and not go straight for a G6400 instead.

Information is secure.
PowerVR GPU based on 2 vec4 + additional scalar - 4x2+1 = 9.
Now 9 ÷ 8 = 1.125 bonus!
Each core has 2 ALU.
PowerVR G6200 contains 16USSE2 (x2 SGX554).

Now 16USSE2 x vec4 x 2 ALU x 2 Core x 1.125 x 0.280MHz ( mhz for MediaTek) = 80GFLOPS.

By simply :

16USSE2 x 2 Core x 0.280 x 9GFLOPS = 80GFLOPS

See more about calculating PowerVR GPU here :

http://www.359gsm.com/forum/viewtopic.php?f=127&t=13396

 

[french toast]

I think there is some naming confusion. ..core = cluster? :/

 

[Plamensito]

I think there is some naming confusion. ..core = cluster? :/

Oooops, sorry

One cycle is a 9 GFLOPS (4x2+1=9).

Now : 16USSE2 x 2 Core x 0.280 x 9 = 80 GFLOPS

The PowerVR G6200 MP2 (280MHz MediaTek) is a 80 GFLOPS

The PowerVR SGX554 MP4 (280MHz Apple) is a 80 GFLOPS

 

[french toast]

Oooops, sorry

One cycle is a 9 GFLOPS (4x2+1=9).

Now : 16USSE2 x 2 Core x 0.280 x 9 = 80GFLOPS

The PowerVR G6200 MP2 (280MHz MediaTek) is a 80GFLOPS

The PowerVR SGX554 MP4 (280MHz Apple) is a 80GFLOPS

Seems logical, although that is one hell of alot of ALU resources for what is essentially a mid-low end chip.

Very interesting nonetheless .

 

[Plamensito]

Oooops, sorry

One cycle is a 9 GFLOPS (4x2+1=9).

Now : 16USSE2 x 2 Core x 0.280 x 9 = 80GFLOPS

The PowerVR G6200 MP2 (280MHz MediaTek) is a 80GFLOPS

The PowerVR SGX554 MP4 (280MHz Apple) is a 80GFLOPS

Officially

The PowerVR for one cycle is a 9GFLOPS (4x2+1)

PowerVR G6200 (MediaTek) :

16USSE2 x 2 Core x 0.280 x 9 = 80GFLOPS

PowerVR SGX554 (Apple iPad 4) :

8USSE2 x 4 Core x 0.280 x 9 = 80GFLOPS

 

[Plamensito]

Seems logical, although that is one hell of alot of ALU resources for what is essentially a mid-low end chip.

Very interesting nonetheless .

This is not true because:

PowerVR G6630 :

16USSE2 x 6 Core x 0.280 x 9 = 241GFLOPS

 

[french toast]

This is not true because:

PowerVR G6630 :

16USSE2 x 6 Core x 0.280 x 9 = 241GFLOPS

Yea and a giant 5000mah battery to go with it

 

[Plamensito]

Yea and a giant 5000mah battery to go with it

At first we see G6400/6430 in iPad 5:

16USSE2 x 4 Core x 0.280 x 9 = 161.3 GFLOPS

Adreno 330 = 129.6GFLOPS

 

[anexanhume]

This is not true because:

PowerVR G6630 :

16USSE2 x 6 Core x 0.280 x 9 = 241GFLOPS

Xenos = 240 GFlops. Now we can start the comparison to next gen!

 

[french toast]

At first we see G6400/6430 in iPad 5:

16USSE2 x 4 Core x 0.280 x 9 = 161.3GFLOPS

Adreno 330 = 129.6GFLOPS

If the figures you say stand up, you will have brought some valuable insight into the shader setup of mobile SOCs

How do you know what the execution resources are for the adreno uarch? Im not aware that they have been published?

Could you say how you made up those performance figures of adreno 330?
How many rops? Tmu? Alus? ...

 

[Plamensito]

If the figures you say stand up, you will have brought some valuable insight into the shader setup of mobile SOCs

How do you know what the execution resources are for the adreno uarch? Im not aware that they have been published?

Could you say how you made up those performance figures of adreno 330?
How many rops? Tmu? Alus? ...

4x2+1 = 9GFLOPS
32 ALU x 9 x 0.450 MHz = 129.6

or

4 x 2 + 1 = 9; 9÷8=1.125
32vec4 = 128 ALU
128 x 2 GFLOPS x 1.125 x 0.450 = 129.6