Samsung Exynos 5250 - production starting in Q2 2012

[Wishmaster]

Surprising, I thought the t624 would be used...and mp6? Didnt see that one.

This has settled it then, a galaxy note 3 containing this beasty will be my next phone.
What about power consumption, hope they hav implemented big little better on this.

I'd suggest you wait for some benchmarks and real life reports. Remember how excited people were about 5410 and look what they think about it now, at least those that know a thing or two about SoC's.
Besides, 2x greater 3D performance is IMO purely marketing. It will probably trade blows with Adreno 330 but nothing more.

 

[Ailuros]

I'd suggest you wait for some benchmarks and real life reports. Remember how excited people were about 5410 and look what they think about it now, at least those that know a thing or two about SoC's.
Besides, 2x greater 3D performance is IMO purely marketing. It will probably trade blows with Adreno 330 but nothing more.

Depends on the frequency of the 628MP6; it probably has the same amount of SIMD ALU lanes as an Adreno330, but the latter is clocked at "only" 450MHz. If the 628MP6 clocks at 600MHz then that's a rough 33% increase in arithmetic throughput and part of it should make in difference even in real time.

***edit: http://www.arm.com/images/GPU_Compute_roadmap.jpg

Sounds like:

622 = 2 clusters, 2 TMUs
624 = 4 clusters, 4 TMUs
628 = 8 clusters, 8 TMUs
678 = 8 clusters, 8 TMUs (4 FLOPs/ALU lane)

Why does Samsung or ARM even call it 628MP6 in the first place and not Mali T626 from the get go?
In any case 6*16*2*0.6 = 115 GFLOPs/s, 3.6 GTexels/s.

 

[Nebuchadnezzar]

Why does Samsung or ARM even call it 628MP6 in the first place and not Mali T626 from the get go? .

The shared blocks are different between the T624 and T628, former scales 1-4 cores and the latter 1-8. Per ARM's diagramm:

Seems to be a substantial difference, especially if it means double the cache.

 

[Wishmaster]

Depends on the frequency of the 628MP6; it probably has the same amount of SIMD ALU lanes as an Adreno330, but the latter is clocked at "only" 450MHz. If the 628MP6 clocks at 600MHz then that's a rough 33% increase in arithmetic throughput and part of it should make in difference even in real time.

If it's really going to run at 600mhz it could be faster than 330, unless 600 will be available only in certain benchmarks

***edit: http://www.arm.com/images/GPU_Compute_roadmap.jpg

Sounds like:

622 = 2 clusters, 2 TMUs
624 = 4 clusters, 4 TMUs
628 = 8 clusters, 8 TMUs
678 = 8 clusters, 8 TMUs (4 FLOPs/ALU lane)

Why does Samsung or ARM even call it 628MP6 in the first place and not Mali T626 from the get go?
In any case 6*16*2*0.6 = 115 GFLOPs/s, 3.6 GTexels/s.

Probably cause it sounds and sells better

Can't wait to see some proper tests of adreno 330 vs. t628mp6 vs. tegra 4. When it comes to overall power efficiency I expect qualcomm to come on top but CPU and GPU will be the interesting part to watch

 

[Ailuros]

The shared blocks are different between the T624 and T628, former scales 1-4 cores and the latter 1-8. Per ARM's diagramm:



Seems to be a substantial difference, especially if it means double the cache.

Good point; I wonder if there's more than just what's visible in the graphs double in the 628

 

[Ailuros]

If it's really going to run at 600mhz it could be faster than 330, unless 600 will be available only in certain benchmarks

Even if it's 540MHz it's still most likely going to end up faster than the 330 unless QCOM manages to squeeze out more out of its drivers. The Mali T604 has 4 TMUs@533MHz, while the Adreno320 in the GalaxyS4 has 8 TMUs@450MHz; compare real time fillrate scores in GLB for instance and you'll see one of the differences.

Can't wait to see some proper tests of adreno 330 vs. t628mp6 vs. tegra 4. When it comes to overall power efficiency I expect qualcomm to come on top but CPU and GPU will be the interesting part to watch

Tegra4 with a heatsink and fan or without one? Without and in so far implementations it ends up below the iPad4 and the Adreno320@450MHz. Could it be that NV created Shield and the reference tablet device just to capture temporarily the benchmark top?

Anyway my prediction would be unless vendors clock the 330 higher than 450MHz it to end up somewhat below the Exynos5420 GPU and underneath that T4. And that also temporarily until this year runs out.

 

[lanek]

Huum , i dont know how credible those sbench are ( as far as antutu is credible ).. and like rumor point it to be in the Note3.

http://www.phonearena.com/news/Samsung-Galaxy-Note-3-benchmarked-again-at-AnTuTu_id45419

- Galaxy S4 ( Exynos ): score 27'000 pnts under Antutu
- Galaxy Note3 ( Exynos ): score 33'000 pnts

I hope Samsung will start too release a real refresh on the tablet side: high res ( 2560x1600 ), + high end SOC.

 

[Wishmaster]

Even if it's 540MHz it's still most likely going to end up faster than the 330 unless QCOM manages to squeeze out more out of its drivers. The Mali T604 has 4 TMUs@533MHz, while the Adreno320 in the GalaxyS4 has 8 TMUs@450MHz; compare real time fillrate scores in GLB for instance and you'll see one of the differences.

Fillrate was always qualcomms weak point, hopefully they'll finally do something about it.

Tegra4 with a heatsink and fan or without one? Without and in so far implementations it ends up below the iPad4 and the Adreno320@450MHz. Could it be that NV created Shield and the reference tablet device just to capture temporarily the benchmark top?

Of course I mean't tegra 4 without heatsink, preferably all tested devices should be smartphones or tablets if possible. IMO, shield was created just for press to ignite some talk in media about tegra

Anyway my prediction would be unless vendors clock the 330 higher than 450MHz it to end up somewhat below the Exynos5420 GPU and underneath that T4. And that also temporarily until this year runs out.

Those recent scores from xperia z ultra that got certified on gfxbench were roughly 20% higher than shields and were probably removed due to the very early stage of the firmware but I assume that 50% scores improvement are possible. Although hopefully fillrate efficiency will be better than on earlier adrenos (hope lies in lpddr3).

 

[Ailuros]

Fillrate was always qualcomms weak point, hopefully they'll finally do something about it.

Geometry throughput being a close second; that's the reason why I'm thinking whether the 628 is also doubling possibly raster or trisetup units besides caches f.e.

Those recent scores from xperia z ultra that got certified on gfxbench were roughly 20% higher than shields and were probably removed due to the very early stage of the firmware but I assume that 50% scores improvement are possible. Although hopefully fillrate efficiency will be better than on earlier adrenos (hope lies in lpddr3).

A 50% improvement over night at this stage sounds fairly impossible for any GPU out there. Anyone could do it but it would most likely also affect driver and in extension system stability. All the preliminary scores that appeared and disappeared for Adreno330 were about 20% higher than Shield.

By the way another reason why Samsung might have picked a MP6 is that a full 628 (MP8) might had been too much in terms of die area and anything less than MP6 might have put it into a fillrate disadvantage compared to the GalaxyS4.

 

[Ailuros]

Huum , i dont know how credible those sbench are ( as far as antutu is credible ).. and like rumor point it to be in the Note3.

http://www.phonearena.com/news/Samsung-Galaxy-Note-3-benchmarked-again-at-AnTuTu_id45419

- Galaxy S4 ( Exynos ): score 27'000 pnts under Antutu
- Galaxy Note3 ( Exynos ): score 33'000 pnts

I hope Samsung will start too release a real refresh on the tablet side: high res ( 2560x1600 ), + high end SOC.

I hardly ever pay attention to Antutu scores since its more than hard to define what they actually stand for, but that link above doesn't point at any Exynos SoCs. What am I missing?

 

[Wishmaster]

A 50% improvement over night at this stage sounds fairly impossible for any GPU out there. Anyone could do it but it would most likely also affect driver and in extension system stability. All the preliminary scores that appeared and disappeared for Adreno330 were about 20% higher than Shield.

50% was in comparison to adreno 320 from s4 pro (like in Xperia Z and nexus 4) that scores around 800-900 in glb2.7 so scores around 1200-1350 should be quite realistic for 330 IMO.

By the way another reason why Samsung might have picked a MP6 is that a full 628 (MP8) might had been too much in terms of die area and anything less than MP6 might have put it into a fillrate disadvantage compared to the GalaxyS4.

It's always interesting to see Samsung change their suppliers so often almost solely for their own devices.

 

[Ailuros]

50% was in comparison to adreno 320 from s4 pro (like in Xperia Z and nexus 4) that scores around 800-900 in glb2.7 so scores around 1200-1350 should be quite realistic for 330 IMO.

26-27 fps sounds like a reasonable start for Adreno330 that's true.

It's always interesting to see Samsung change their suppliers so often almost solely for their own devices.

Samsung has good reasons to stick with ARM for GPU IP.

 

[Wishmaster]

Samsung has good reasons to stick with ARM for GPU IP.

Probably, but year to year I wonder will we finally get to see some mass market device using Samsung SoC not manufactured by them. Unfortunately it's not like that and yet they still invest and innovate solely for the purpose of their own devices.
I'm quite interested what will be the future of their SoC business. But that's a subject for a different talk

 

[Ailuros]

Probably, but year to year I wonder will we finally get to see some mass market device using Samsung SoC not manufactured by them. Unfortunately it's not like that and yet they still invest and innovate solely for the purpose of their own devices.
I'm quite interested what will be the future of their SoC business. But that's a subject for a different talk

Open a new thread for it; it's definitely an interesting topic to discuss.

 

[Wishmaster]

Open a new thread for it; it's definitely an interesting topic to discuss.

Actually that's a very good idea
I'll start looking for some interesting materials and prepare first post for a new thread.
Thanks for the idea

 

[french toast]

Im thinking in terms of horsepower the t628 mp6 has more welly, but it might come down to driver maturity.

T628 has some IPC advantages over t604, thats without considering the extra 'cores' likely increased clocks, and increased bandwidth.

Im actually more interested in whether samsung and ARM can get big-LITTLE working better than on broken 5210.

 

[Mariner]

Bear in mind that the Mediatek Octa-A7 chip reputedly scores around 30,000 in Antutu.

It really is a pretty hopeless benchmark for anything other than messageboard gossip.

 

[Nebuchadnezzar]

Bear in mind that the Mediatek Octa-A7 chip reputedly scores around 30,000 in Antutu.

It really is a pretty hopeless benchmark for anything other than messageboard gossip.

I don't see the problem with it, all you need to do is to check the detailed scores instead of the final one.

 

[Mariner]

My point is that the final score is pretty useless in most respects.

On a chinaphone forum I visit, some people have taken the fact that the Mediatek Octa-A7 chip scores higher than the Snapdragon 800 as proof that it will be faster. Of course, we know this is more than a little bit unlikely.

 

[Ailuros]

My point is that the final score is pretty useless in most respects.

On a chinaphone forum I visit, some people have taken the fact that the Mediatek Octa-A7 chip scores higher than the Snapdragon 800 as proof that it will be faster. Of course, we know this is more than a little bit unlikely.

For Antutu I won't disagree; however just because one synthetic benchmark is of questionable value it doesn't mean that all are. Since the debate above was mostly about GLBenchmark scores, there are lightyears of differences between that one and something like Antutu. At least IMHO if someone knows how to interpret Kishonti's results there are far more represantive for mobile GPU performance than Antutu would be for....(wait what is that thing measuring again exactly?)....