[Ailuros]

Originally Posted by Jubei

Ailuros is correct. Now that i read through the pdf properly, Samsung was meauring it as 2x better than 4412 at wxga resolution. Im assuming its 4412 given that they mentioned cpu frequency of 1,4 GHz

Well to put it on another level I don't believe that a T604 would reach almost 10k frames in GLBenchmark2.5 (1080p offscreen) before I see it.

[Nebuchadnezzar]

GLBenchmark from a leaked device: http://briefmobile.com/nexus-10-rece...ider-treatment

Resized picture for the sake of sanity:




Pretty equal to the A6, but gets beaten in low-level benches. Also tied up with the A5X. I guess the it won't be a match for the A6X. On the other hand beats the Adreno 320 handily.

I wonder how drivers will improve it, it's still a new architecture. Remember that the Mali 400 got roughly 40% boost from early release to several months later through drivers.

It does keep its 2x performance over a ~440-533MHz Mali400. 5x 266MHz Mali400 seems to be roughly validated.

[tangey]

I guess it's all semantics, but when I read in the original PR for T-604:-
"The Mali-T604 delivers up to 5x performance improvement over previous Mali graphics processors...."

My reasonable expectation is that this is on a clock to clock comparison, i.e. an archeiture improvement

[Wishmaster]

Originally Posted by Nebuchadnezzar

Pretty equal to the A6, but gets beaten in low-level benches. Also tied up with the A5X. I guess the it won't be a match for the A6X. On the other hand beats the Adreno 320 handily.

Where does it beat adreno 320 handily? Even at briefmobile they said that adreno 320 was faster at GLbenchmark, not always but it wasn't beaten as you say it was. And when you compare scores from the fastest adreno 320 at glbenchmark it wins with nexus 10 at almost all offscreen tests.

[Jubei]

Shouldnt the fillrate be quite a bit higher?

[glw]

I see that BriefMobile also ran SunSpider, that score is much worse than I expected. The Chromebook with the same chip is twice as fast. I can accept that Chrome on Android is not the same version as Chrome on ChromeOS but that is a huge difference for software from the same company.

[Ailuros]

Originally Posted by Jubei

Shouldnt the fillrate be quite a bit higher?

Not really; in fact the fillrate efficiency on T604 is higher than on Mali400MP4@533MHz:

http://www.glbenchmark.com/phonedeta...group=lowlevel

[Ailuros]

Originally Posted by Nebuchadnezzar

GLBenchmark from a leaked device: http://briefmobile.com/nexus-10-rece...ider-treatment

Resized picture for the sake of sanity:


Pretty equal to the A6, but gets beaten in low-level benches. Also tied up with the A5X. I guess the it won't be a match for the A6X. On the other hand beats the Adreno 320 handily.

I wonder how drivers will improve it, it's still a new architecture. Remember that the Mali 400 got roughly 40% boost from early release to several months later through drivers.

It does keep its 2x performance over a ~440-533MHz Mali400. 5x 266MHz Mali400 seems to be roughly validated.

Mali400MP4@533MHz = 1977 frames
Mail400MP4@440MHz = 1827 frames (latest Galaxy SIII score)
Mali400MP4@266MHz = 1475 frames (latest Galaxy SII score)

3313 / 1977 = 1.68x
3313 / 1827 = 1.81x
3313 / 1475 = 2.24x

With an up to 40% driver improvement in the future they'll slightly exceed the 2x times mark over the 533MHz Mali400MP4 under the presupposition that the latter won't see a single inch of addtional performance against the T604.

Thost 2x or 5x times metrics don't have to be for GLBenchmark-anything by the way. I'm sure it won't be too hard to find a case with high enough geometry load where a T604 could literally slaughter a 400MP4@533MHz.

[Jubei]

Originally Posted by Ailuros

Not really; in fact the fillrate efficiency on T604 is higher than on Mali400MP4@533MHz:

http://www.glbenchmark.com/phonedeta...group=lowlevel

Aha thanks. I just thought it would be closer to the theoretical peak fillrate of 2 gtexels

[Ailuros]

Originally Posted by Jubei

Aha thanks. I just thought it would be closer to the theoretical peak fillrate of 2 gtexels

For that it would have to be a true TBDR.

[Nebuchadnezzar]

Originally Posted by Wishmaster

Where does it beat adreno 320 handily? Even at briefmobile they said that adreno 320 was faster at GLbenchmark, not always but it wasn't beaten as you say it was. And when you compare scores from the fastest adreno 320 at glbenchmark it wins with nexus 10 at almost all offscreen tests.

All the low level scores are higher than the 320 by quite some bit, it generally points out do driver immaturity on the "normal" scores. We'll have to see.

It's a pity that GLBenchmark is the only benchmark out there offering fixed resolution benchmarking, I would like to see some Taiji offscreen scores, as that's more representative of today's gaming loads.

[JohnH]

Originally Posted by Nebuchadnezzar

...

It's a pity that GLBenchmark is the only benchmark out there offering fixed resolution benchmarking, I would like to see some Taiji offscreen scores, as that's more representative of today's gaming loads.

What's you're basis for saying it's more representative?

[Jubei]

Originally Posted by Ailuros

For that it would have to be a true TBDR.

Understood. Makes me wonder if ARM underestimated where the competition would be in 2013.

When you go to their T604 product website:

http://www.arm.com/products/multimed.../mali-t604.php

Its almost like they are saying: "Skip this product" in favour of the T624

Same if you look at the T658 (wich was announced with T604) they write that you should look at T678 instead

[Xmas]

Originally Posted by tangey

I guess it's all semantics, but when I read in the original PR for T-604:-
"The Mali-T604 delivers up to 5x performance improvement over previous Mali graphics processors...."

My reasonable expectation is that this is on a clock to clock comparison, i.e. an archeiture improvement

Target clock rates are part of the architecture, so I'm not sure you can split it like that.

Originally Posted by Ailuros

For that it would have to be a true TBDR.

To reach its peak fillrate in a fillrate test? Why?

[Ailuros]

Originally Posted by Xmas

To reach its peak fillrate in a fillrate test? Why?

I might be wrong but I only have seen TBDRs so far coming damn close to their peak theoretical fillrate in synthetic fillrate tests.

Originally Posted by Jubei

Understood. Makes me wonder if ARM underestimated where the competition would be in 2013.

I'm taking a different perspective: the cost to support native FP64 isn't exactly cheap rather the contrary. Inevitably it will lead to tradeoffs for performance on other aspects of the GPU like FP32 throughput for example. Whether ARM made the correct design decision to include FP64 that early or not is something that remains to be seen.

When you go to their T604 product website:

http://www.arm.com/products/multimed.../mali-t604.php

Its almost like they are saying: "Skip this product" in favour of the T624

I'm not even in the clear yet what T624 exactly stands for compared to T604.

Same if you look at the T658 (wich was announced with T604) they write that you should look at T678 instead

I wouldn't pay too much attention to that one since it's not exactly absurd when you have a page for product A to also try to attract a potential customer for a higher end variant than A. T658/678 seem to have twice the amount of clusters of 604/624, which could mean that 658 doubles floating point throughput (and texel fillrates amongst others) compared to 604 at the same frequency. The much more burning question would be how big T604 exactly is and how much bigger or the other higher end variants.

[davetherave]

New to the Forum, although I've been following it for years.

I am curious about this chipset especially seeing as its likely to be included in the soon to be announced Nexus 10.

Currently the Exynos 5250 is used in the Samsung Chromebook which has been pegged as having up to 6.5 hours battery life. Surely if the 5250 is being used in the Nexus 10 and has to power a larger screen (2560×1600) compared to the Chromebook (1366x768) as well as having to be in a smaller form (tablet rather than laptop), will this not eat up battery like no tomorrow?

Should be expect a smaller cut chip that reduces the power drain or will the chipset be clocked slower to slow the battery drain?

What's your ideas?

[Ailuros]

An idea would be that since a high end tablet isn't going to be cheap and especially with a 2560*1600 display, the manufacturer can include a huge enough battery to ensure a reasonably high battery life. Ever checked how huge the battery on an iPad3 is?

[Wellington]

Exynos 5250 topping GlBenchmark : http://glbenchmark.com/phonedetails....chmark=glpro25.

[Pressure]

Well, it's also clocked at 1.70 Ghz ...

[Ailuros]

Originally Posted by Pressure

Well, it's also clocked at 1.70 Ghz ...

It's a devboard running at 1024 x 552 for the onscreen results. I initially thought those were in 2560 and got excited until I saw the actual display resolution.

[Wellington]

Well, it's got HDMI so guess it can run at a much higher res.

The offscreen test is what counts.

[Ailuros]

Originally Posted by Wellington

The offscreen test is what counts.

Definitely; it still seems to need a whole damn lot of driver love IMHO.

[Nebuchadnezzar]

First actual 5450 release rumours: http://translate.google.co.jp/transl...%3Fuid%3D96826

Move to 28nm in the 5450 and tentatively in the next Samsung flagship (S IV) by next spring/summer. That's earlier than I expected it. If it still comes with the planned T658, that would be quite a big SoC.

[Ailuros]

The 32nm to 28nm transistion will swallow at least some of the additional die area.

[Jubei]

Nexus 10 now officially announced. 399 dollars for 16 GB version and 499 for the 32 GB version