Samsung Exynos 5250 - production starting in Q2 2012

[Nebuchadnezzar]

My bad. T628 then

The T624 is supposed to be 50% faster than the T604. The T628 is again twice as fast as that as it's basically just double the T624 in every aspect. We would have heard something if the performance jump would be that huge, so I doubt it.

 

[Jubei]

The T624 is supposed to be 50% faster than the T604. The T628 is again twice as fast as that as it's basically just double the T624 in every aspect. We would have heard something if the performance jump would be that huge, so I doubt it.

Looking at ARMs site, the only difference between T604/T628 seems to be improved API support. I would have thought ARM would mention a 50% performance improvement.

Whereas they do mention that T604 is 5X previous Mali and T628 is 2X T624

 

[loekf]

Looking at ARMs site, the only difference between T604/T628 seems to be improved API support. I would have thought ARM would mention a 50% performance improvement.

Whereas they do mention that T604 is 5X previous Mali and T628 is 2X T624

In other news, saw that another competitor of Samsung, namely Broadcom took a 32-bits/64-bits ARM architecture license (so not a CPU license, but architecture license), allowing them to develop their own cores compatible with the ARM v7/v8 ISA (so we now have Qualcomm, Apple, Microsoft (?), AMD, Nvidia, AppliedMicro and Broadcom).

Makes you wonder why Samsung doesn't do this ?

Just throwing more and more cores into a SoC, is that a guaranteed winner ?

 

[Nebuchadnezzar]

Looking at ARMs site, the only difference between T604/T628 seems to be improved API support. I would have thought ARM would mention a 50% performance improvement.

Whereas they do mention that T604 is 5X previous Mali and T628 is 2X T624

I don't remember where it was, but they're supposedly getting performance gains due to architectural efficiency gains in the 6x4/8 variants while retaining the unit counts.

It's placed higher in the product lineup:
http://images.anandtech.com/doci/6136/Screen Shot 2012-08-06 at 11.46.02 AM.png

ARM announced the second generation of the Mali-T600 Series GPUs targeting tablets, smartphones and smart TVs. Those new GPUs provide up to 50% performance increase over the first generation Midgard GPUs (T-604 & T-658) and include support for Adaptive Scalable Texture Compression (ASTC), a texture compression technique. The company explains that ASTC significantly optimizes GPU performance and increases battery life in devices.

Read more: http://www.cnx-software.com/2012/08...i-t624-mali-t628-and-mali-t678/#ixzz2HVrBQ2xm

In other news, saw that another competitor of Samsung, namely Broadcom took a 32-bits/64-bits ARM architecture license (so not a CPU license, but architecture license), allowing them to develop their own cores compatible with the ARM v7/v8 ISA (so we now have Qualcomm, Apple, Microsoft (?), AMD, Nvidia, AppliedMicro and Broadcom).

Makes you wonder why Samsung doesn't do this ?

Just throwing more and more cores into a SoC, is that a guaranteed winner ?

They already have an architectural license AFAIK.

 

[Laurent06]

In other news, saw that another competitor of Samsung, namely Broadcom took a 32-bits/64-bits ARM architecture license (so not a CPU license, but architecture license), allowing them to develop their own cores compatible with the ARM v7/v8 ISA (so we now have Qualcomm, Apple, Microsoft (?), AMD, Nvidia, AppliedMicro and Broadcom).

Makes you wonder why Samsung doesn't do this ?

Samsung has an ARMv8 license. Broadcom already has a subscription license that gives them access to all ARM cores.

Ref:
http://www.realworldtech.com/arm64/
http://www.eetimes.com/electronics-news/4213168/ARM--Broadcom-expand-ties

 

[Helmore]

And let's not forget Marvell and their Armada cores. Those are ARMv7, but it wouldn't surprise me if they have an ARMv8 license.

 

[Nebuchadnezzar]

http://images.anandtech.com/galleries/2554/CES2013-111056.jpg

From: http://www.anandtech.com/show/6602/samsung-announces-exynos-5-octa-soc-4-cortex-a7s-4-cortex-a15s

They have some power consumption comparison using LabView live measuring just as they did last year with the 4210 vs the 4212. If those graphs are to scale which I believe they are, the lowest idle power is extremely low with peaks reaching 2W.

I wonder what type of typical load they are synthesizing here.

 

[Tabris]

rumor from China:
SGX544MP3@533MHz

 

[tangey]

rumor from China:
SGX544MP3@533MHz

any links ?

Samsung does have SOMETHING coming out with SGX5XT in it, I'd be amazed if it was the flagship Exynos octa core, would be massive score for IMG to get it. I expect samsung will be using IMG in some variant to make it easier to get microsoft to qualify it for windows RT.

Howeve it is a bit weird that in a 15 min video conversation with engadget that was dominated by discussion on the Exynos octa, I did not hear a mention made of arm graphics, nor did samsung mention arm graphics in their presentation. Would appear that graphics is becoming the most secretive aspect of socs.

 

[Jubei]

I watched the keynote and Dr Woo says the 3D performance is 2X any previous generation Exynos including Exynos 4 Quad. So im assuming he considers Exynos 5 Dual to be the same generation

 

[anexanhume]

I still can't wrap my head around having four A7's too. Surely it makes more sense to spin up one or two A15 cores as opposed to all 4 A7 cores?

http://ces.cnet.com/8301-34435_1-57563033/samsung-announces-new-octa-8-core-processor-at-ces-2013/
Samsung usually never pulled figues out of their ass in this regard and the 4412 marketing figures from last year were pretty accurate to reality, if so, 70% savings is quite remarkable.

I believe it. Node and a half jump with HKMG added. Then you have an A7 core to turn on instead of an A15 for any standby tasks.

 

[juicytuna]

I still can't wrap my head around having four A7's too. Surely it makes more sense to spin up one or two A15 cores as opposed to all 4 A7 cores?

From the big.LITTLE whitepaper:

A final point to consider is that a big.LITTLE system incorporating Cortex-A15, Cortex-A7, CCI-400 and
the GIC-400 is optimal for all big.LITTLE use-models. There is no configurable feature or optimization
that favors a particular use-model. However, to reduce software complexity in the big.LITTLE task
migration use model it is recommended that the same number of cores be implemented in the Cortex-A15
cluster and Cortex-A7 cluster.

 

[anexanhume]

From the big.LITTLE whitepaper:

Thanks. I wonder if people with control over their hardware and software (apple) will abide by that, if they even use the A7.

 

[ltcommander.data]

I believe it. Node and a half jump with HKMG added. Then you have an A7 core to turn on instead of an A15 for any standby tasks.

Are you comparing to Exynos 4412? That was on 32nm so it's only a half node to 28nm for the Octa and the 32nm process already introduced HKMG.

 

[liolio]

From the big.LITTLE whitepaper:

Well A7 are tiny but that looks like a pretty wasteful approach.
They hint that the software doesn't really know how to deal the hardware. It looks like the code either on the quad A15 or on the quad A7.

 

[anexanhume]

Are you comparing to Exynos 4412? That was on 32nm so it's only a half node to 28nm for the Octa and the 32nm process already introduced HKMG.

Samsung is comparing to their own 45nm process for the 70% reduction. IIRC the claimed power savings from 32nm to 28nm alone was around 15-20%.

http://images.anandtech.com/galleries/2554/CES2013-111056.jpg

From: http://www.anandtech.com/show/6602/samsung-announces-exynos-5-octa-soc-4-cortex-a7s-4-cortex-a15s

They have some power consumption comparison using LabView live measuring just as they did last year with the 4210 vs the 4212. If those graphs are to scale which I believe they are, the lowest idle power is extremely low with peaks reaching 2W.

I wonder what type of typical load they are synthesizing here.

 

[Nebuchadnezzar]

Samsung is comparing to their own 45nm process for the 70% reduction. IIRC the claimed power savings from 32nm to 28nm alone was around 15-20%.

Personally I interpreted the 70% figure coming from the difference in average power between 32nm to 28nm compared from the 5250 to the new big.LITTLE architecture.

 

[anexanhume]

Personally I interpreted the 70% figure coming from the difference in average power between 32nm to 28nm compared from the 5250 to the new big.LITTLE architecture.

All their hard numbers and charts compared to 45nm, so I'm assuming that's where the 70% comes from since they didn't specify. If they're willing to go two nodes back for what they fully disclose, why not do the same when they don't even tell you what you're comparing to?

 

[Exophase]

Yeah I'm really having a hard time deciphering just what Samsung is talking about too..

All I know is the process bump will give some improvement. AFAIK Samsung says the density goes up by around 33%. Anandtech threw out an estimate of 15% improvement on power consumption, presumably at peaks. That and Cortex-A7s are a good step, and welcome even so soon after Exynos 5250 (assuming it doesn't take a long time to enter products.. is it just me or has the announce to ship cycle for mobile parts been shortening dramatically?)

What I'm most curious about is if they've been able to drive down the voltage for peak clocks appreciably. Since that has such a strong bearing on the overall press the CPU technology receives. Sites have been absolutely lighting up about how Cortex-A15 gives you a little better performance than Atom for a huge power premium.

 

[anexanhume]

Yeah I'm really having a hard time deciphering just what Samsung is talking about too..

All I know is the process bump will give some improvement. AFAIK Samsung says the density goes up by around 33%. Anandtech threw out an estimate of 15% improvement on power consumption, presumably at peaks. That and Cortex-A7s are a good step, and welcome even so soon after Exynos 5250 (assuming it doesn't take a long time to enter products.. is it just me or has the announce to ship cycle for mobile parts been shortening dramatically?)

What I'm most curious about is if they've been able to drive down the voltage for peak clocks appreciably. Since that has such a strong bearing on the overall press the CPU technology receives. Sites have been absolutely lighting up about how Cortex-A15 gives you a little better performance than Atom for a huge power premium.

Not just you. Cortex A8 was announced in October 2005 and didn't see a consumer product until June 2009 (palm pre, iphone 3gs etc.). A7 was announced October 2011 and looks to be in a consumer product somewhere between 1.5 and 2 years. The SoC platform to integration has shortened too, but I think the architecture to device timeline really illustrates the quickened pace.