Between IMG's investments and consequent influence in start-up and/or smaller fabless semis, like Frontier and the like, and their own internal use within IP like VXD and end-product like Pure radios, IMG is already finding a decent base from which to try to expand META's reach.
PowerVR Series 6 now official
- Thread starter roninja
- Start date
Series 6 performance
The Nova A9600, built in 28nm, will deliver groundbreaking multimedia and graphics performance, featuring a dual-core ARM Cortex- A15-based processor running up to up to 2.5 GHz breaking the 20k DMIPS barrier, and a POWERVR Rogue GPU that delivers in excess of 210 GFLOPS. The graphics performance of the A9600 will exceed 350 million ‘real’ polygons per second and more than 5 gigapixels per second visible fill rate (which given POWERVR’s deferred rendering architecture results in more than 13 gigapixels per second effective fill rate). Thanks to Rogue Nova will support all existing APIs such as OpenGL, OpenGL ES 2.0/1.1, OpenVG and OpenCL as well as the next generation of OpenGL ES, codenamed ‘Halti’ and Microsoft DirectX. The Nova A9600 is sampling in 2011.
So what do folks think about the quoted performance figures?
The Nova A9600, built in 28nm, will deliver groundbreaking multimedia and graphics performance, featuring a dual-core ARM Cortex- A15-based processor running up to up to 2.5 GHz breaking the 20k DMIPS barrier, and a POWERVR Rogue GPU that delivers in excess of 210 GFLOPS. The graphics performance of the A9600 will exceed 350 million ‘real’ polygons per second and more than 5 gigapixels per second visible fill rate (which given POWERVR’s deferred rendering architecture results in more than 13 gigapixels per second effective fill rate). Thanks to Rogue Nova will support all existing APIs such as OpenGL, OpenGL ES 2.0/1.1, OpenVG and OpenCL as well as the next generation of OpenGL ES, codenamed ‘Halti’ and Microsoft DirectX. The Nova A9600 is sampling in 2011.
So what do folks think about the quoted performance figures?
impressive.
Last edited by a moderator:
Single or MP?
***edit: by the way if my math isn't broken you'd need a SGX554 MP16@200+MHz (or a SGX543/4 MP16@400+MHz) to reach that GFLOP rate.
Rogue with those performance metrics mentioned by ST Ericsson being single core is a too fat pill for me to swallow. Could very well be that you're right and I guess I'll manage to hammer that damn pill into a few dozen pieces in the meantime.
That gpu has more bogo-flops than my laptop's gpu.
And my laptop has a nv gt 330M. That CPU comes awfully close to my laptop's cpu, which is an i5 clocked @2.26GHz.I think my laptop just got pwnd by a freaking tablet.
Last edited by a moderator:
I agree with tangey - there was no mention of multi-core, so I would assume it is a single core.
If Series 6 can scale like SGX-MP (and I can't imagine it wouldn't) then a 16 core Series 6 would have performance figures of 5,600 million polygons per second and more than 208 gigapixels per second effective fill rate.
Which would be nice.
If Series 6 can scale like SGX-MP (and I can't imagine it wouldn't) then a 16 core Series 6 would have performance figures of 5,600 million polygons per second and more than 208 gigapixels per second effective fill rate.
Which would be nice.
Nice, it brings more than 20-fold GPU improvement over U8500. Sure makes those puny 5x improvements everyone announces look like crap. U8500 Mali-400 is single core, and of course that's:
SGX540 in S5PC110 is 90MTri/sec
Orion is 5x faster with a Mali-400MP4, or 450MTri/sec
But U8500 is only a Mali-400MP1 so only 112.5MTri/sec
And this is 20x faster so 2250MTri/sec.. hey I think they're short changing us on the numbers.
In all seriousness, this does look good. Refreshing to finally hear something technically related about Rogue, if only vaguely (will everyone hate me if I start calling the platform PowerVR Vague?) Let's see if ST-Ericsson can actually get U8500 out in something first, before it shifts from becoming vogue to announce a platform that's two ahead of what's reasonably deployed.
(I mean, PowerVR Vogue would at least be more flattering)
SGX540 in S5PC110 is 90MTri/sec
Orion is 5x faster with a Mali-400MP4, or 450MTri/sec
But U8500 is only a Mali-400MP1 so only 112.5MTri/sec
And this is 20x faster so 2250MTri/sec.. hey I think they're short changing us on the numbers.
In all seriousness, this does look good. Refreshing to finally hear something technically related about Rogue, if only vaguely (will everyone hate me if I start calling the platform PowerVR Vague?) Let's see if ST-Ericsson can actually get U8500 out in something first, before it shifts from becoming vogue to announce a platform that's two ahead of what's reasonably deployed.
(I mean, PowerVR Vogue would at least be more flattering)
Personally I think IMG are doing the right thing around teasing us with the specs, in their defense we are not really hearing too much so far about its competition. What have Nvidia, ARM, Vivante got up their sleaves in regards to their respective next generations. Are they even in the same ball-park? Guess we will have to wait and see.
From what has been revealed so far, we have 2 recently announced Cortex A15 SoCs both paired with different generations of PowerVR graphics. Nvidia will obviously have Tegra 4,5...etc in their plans, ARM Mali seems more doubtful we heard about the T614 but would anyone serious licence it ahead of Series 6 on performance alone, i doubt it. On cost, maybe?
From what has been revealed so far, we have 2 recently announced Cortex A15 SoCs both paired with different generations of PowerVR graphics. Nvidia will obviously have Tegra 4,5...etc in their plans, ARM Mali seems more doubtful we heard about the T614 but would anyone serious licence it ahead of Series 6 on performance alone, i doubt it. On cost, maybe?
Besides the obvious joke Samsung mentions in all of it's S5PC11x documentation 20M Tris/s for the SGX540@200MHz.
The higher clocked Mali400MP4 in Orion looks like that compared to a 540@200MHz:
http://www.glbenchmark.com/compare....Google Nexus S&D2=Samsung GT-i9100 Galaxy S 2
wherever there's fragment related processing the MP4 is in its own paradise; alas though if it hits on a healthy amount of vertices and/or geometry. Too bad Anand hasn't yet run all the low level results for the OMAP4430.
Strike a pose!
Hm, interesting..
If Orion runs its Mali-MP4 at 275MHz that means 1.375x clock over S5PC110, and 2x the TMUs, so an effective 2.75x TMU performance. On texture fetch we see 2.26x improvement, so I'd say that's about measuring up.
For ALU performance it looks like Mali can do 16x FP16 FMADD while SGX540 can do 8x FP16 FMADD (or under some circumstances 8x FP32 FMADD, but that's irrelevant if they're not being used in this comparison). Mali may also be able to do more work per-cycle.
I'm trying to find the best user of FMADDs and other simple ALU resources from the benchmark.. but the best indicators I can think of - "Common test: fragment weighted", "Geometric test: fragment weighted" and "Triangle test: textured, fragment lit", are performing worse on Mali than SGX. The results of more complex operations, on the other hand, exponential operations and especially trigonometric, scream on the Mali in comparison. For loops and branching are also a lot better, although that's pretty separate from core ALU performance.
It's possible that this is also from vertex shading being a bottleneck, but then you should see performance improve when moving from vertex lighting to fragment lighting, I would think. It looks to me like something else is bogging down fragment ALU operations, while all other fragment operations are working great. So I wouldn't count
If Orion runs its Mali-MP4 at 275MHz that means 1.375x clock over S5PC110, and 2x the TMUs, so an effective 2.75x TMU performance. On texture fetch we see 2.26x improvement, so I'd say that's about measuring up.
For ALU performance it looks like Mali can do 16x FP16 FMADD while SGX540 can do 8x FP16 FMADD (or under some circumstances 8x FP32 FMADD, but that's irrelevant if they're not being used in this comparison). Mali may also be able to do more work per-cycle.
I'm trying to find the best user of FMADDs and other simple ALU resources from the benchmark.. but the best indicators I can think of - "Common test: fragment weighted", "Geometric test: fragment weighted" and "Triangle test: textured, fragment lit", are performing worse on Mali than SGX. The results of more complex operations, on the other hand, exponential operations and especially trigonometric, scream on the Mali in comparison. For loops and branching are also a lot better, although that's pretty separate from core ALU performance.
It's possible that this is also from vertex shading being a bottleneck, but then you should see performance improve when moving from vertex lighting to fragment lighting, I would think. It looks to me like something else is bogging down fragment ALU operations, while all other fragment operations are working great. So I wouldn't count
(this thrread is veering off-topic, as its startin to talk about Orion instead of Rogue...but...)
http://www.samsung.com/global/business/semiconductor/support/brochures/downloads/systemlsi/Orion.pdf
The marketing stuff in the above states "3,200 pixels/s rendering performance."
Assuming thats fill rate, then Samsung must be allowing a x2 overdraw, giving 1600 pixel's raw fill rate, which would align with the mali400MP4 running at 400Mhz (assuming they get 1 pixel per clock per core).
Of course thats just the chip's max performance, and doesn't necessarily mean that's what it is running at in the phone.
275MHz on Galaxy S2. That figure comes from elsewhere.
Prompted by the above I had a hunt around and found linux orion mali-400 drivers, and indeed the options appear to be 134Mhz or 267Mhz, depending on whether its stepping 0 or 1.
Put yourself in Samsung's marketing department position. When you've marketed the S5PC11x with 1.0GPixels fill-rate and have claimed after that that the successor will deliver 5x times the graphics performance you can't easily claim that the Orion GPU is suddenly at 1.1GPixels/s. If they would have stated 400MPixels for the first things would be way easier.
I'm still not convinced yet that it's a single core, but it's way too early to know any other specifics unless IMG releases more info about it.
There are obviously architectural efficiency increases amongst many other factors like possible higher frequencies possible under 28nm that one should count in.
