Samsung Orion SoC - dual-core A9 + "5 times the 3D graphics performance"

Around 1:49 it looked as though they had ported VillageMark.

 

Samsung is never following anyone's IP roadmap. I predict they'll take T658 and Rogue licenses in time.

If they thought they had even a 20% chance of being competitive, they'd try developing their own graphics IP again, again.

 

We know for know that T604 is capable of 68GFLOPs over MP4 and Rogue in the A9600 over 210GFLOPs over an unknown amount of cores. Even if the latter is a MP4 in the end the GFLOP rate/core is still very high. For 604 it's 17 GFLOPs/core and I wouldn't be surprised if 658 is twice as much considering that 10x times up to claim.

 

Presume they already have a T658 license by arrangement having access to ARM CPU cores. I think eventually they will also take up Rogue.

 

I need help figuring out this slide:



What do they mean with "relative performance? Relative to what?

 

Cortex A8/Mali 200. I wouldn't personally start such a graph at 0 but with a 1. That's a very interesting graph by the way as it projects T658 in 2014 and not earlier.

 

That can't be..
Otherwise 2014's dual Cortex A7 + Mali400MP2 would be as fast as 2009's Cortex A8 + Mali200.

Maybe the "relative performance" is only on a vertical level.. but then again, none of the "squares" are vertically-aligned.

Probably the weirdest graphic I've seen.

 

Of course is it vertically alligned. I don't even recall which "superphone" in 2009 contained a Mali200/A8

 

I dunno, but if a quad A9 + Mali-T604 in mid-2012 is only ~30% more powerful than a Exynos 4210 then ARM has some serious problems. If we took relative performance to mean peak and combining both GPU and CPU that'd seem to imply that Mali-T604 MP4 is much weaker than Mali-400MP4. I have to assume this is only comparing the GPUs despite listing the CPUs, but the relative performance is still extremely underwhelming given it spans a process that will presumably let it clock more than 30% higher. And then a Mali GPU in the middle of 2015 would not even be twice as powerful as Mali-400MP4... that's what, not even today's 40nm SGX543MP2 performance? On 20nm or smaller?? That'd be pathetic... And if dual core Cortex-A15s are still on the cutting edge of high end smartphones in the middle of 2015 then they'd better pretty damn highly clocked. Bizarre that ARM would list a quad core A9 next year but not a quad core A15 three years later.

I would hope this slide is just completely wrong.

 

It's wrong. The idea that a 272GFlops GPU (T658 MP8@500MHz) is less than twice as fast as a 12GFlops GPU (400 MP4@275MHz) is absolutely ridiculous (note: all GFlops figures are extremely easy to infer from public information). I'm under NDA and working on an extremely in-depth article on Mali-T658/T604/400, so stay tuned

 

May be the y axis is log scale? :/

 

272/8 =34Gflops for a single core @ 500Mhz.

That would be 13.6Gflops @ 200Mhz

SGX554 can't be all that far away from that, can it ?

If so, would seem that Mali's road map isn't terribly strong compared to IMG, given that Rogue is likely 50/70 Gflops per core , depending on whether the implementation in the A9600 is 3 or 4 cores. We don't know what the frequency is, but it can't be higher than 500Mhz surely ?

 

Per core GFLOPs should be a lot higher on Rogue.

Max theoretical at 500MHz 36GFLOPs/core. I still can't figure who'd want such a core with Rogue probably starting in 2013; mostly because of the latter's higher capabilities.

Why not? There aren't any specifics known about Rogue, but I wouldn't be surprised if the frequency is just a notch below 700MHz.

 

[SGX554 Can't be all that far away, can it ?]

Thanks for that, confirms that Mali next next generation (T604 has yet to see the light of day), is in performance terms roughly equivalent to an SGX core that has been licensable (and is likely to have *been* licensed) for some time.

Although of course the Mali core has much better compliance.

 

Actually in an ARM blog they state that Mali T658 can push up to 350 GFLOPS, see here.
Still I don't know how this compare to Power VR offering.

 

Without accounting for power, heat, area, architecture for clock speed targets, etc, noting similarity between two competing cores based just on GFLOPs doesn't say much.

 

It states also over 5 GPixels/s real fill-rate. 5200 / 8 = 650MHz.

 

What accuracy are those FLOPS?

 

I like the both T604/T658 in store next year part of that ARM blog.

 

They obviously aren't quoting FP64; since the T6xx family has USC ALUs it's most likely FP32 single precision.