PowerVR Series 6 now official

The integration of ray tracing with the PowerVR rasterization pipeline, both hardware and software, should be a focus of Imgtec at this year's GDC.

A mention at MWC 2014:

http://youtu.be/QTuf3ed0Jo4

 

The GX6650 has been getting attention as a competitor to Tegra's Kepler, but it's not likely to be its contemporary. The 6630 should be a closer match in terms of end product availability, and it should still bring with it quite a few of the 6650's competitive attributes like the six count of clusters and the "power gearing" and "triple compression" technologies (albeit in earlier, um, incarnations).

It won't have some of the significant architectural Series 6XT upgrades like the extra FP16 capacity and, perhaps of even greater impact, what I presume would be the rebalanced throughput characteristics of the XT refresh like potentially more triangle set-up, texture filtering, z units, etc. Hopefully, some of that was upgraded on the 6630 from earlier cores like the 64xx anyway, though.

 

Allwinner A80 results are on/off lately so in case they don't stick this time for whoever might be interested:


http://www.glbenchmark.com/device.jsp?benchmark=gfx30&os=Android&api=gl&D=Allwinner%20A80-EVB

http://www.glbenchmark.com/device.jsp?benchmark=gfx27&D=Allwinner+A80-EVB&testgroup=overall

Also first Mediatek MT6595 have appeared:

http://www.glbenchmark.com/device.j...id&api=gl&D=MediaTek+MT6595&testgroup=overall

I'd expect frequencies between the two to be comparable (MT6595 GPU is at 600MHz) and the difference in performance being mostly due to A80= G6230 vs. MT6595= G6200.

On a sidenote I'm still wondering WTF happened with the MT8135 tablet SoC; considering how much the MT6595 has advanced it might be an idea to use that one for tablets instead if that's even possible. As it stands if the 8135 doesn't get a higher GPU frequency it's going to pose pretty poor against the A80.

 

http://www.glbenchmark.com/device.jsp?benchmark=gfx30&os=Android&api=gl&D=Allwinner UltraOcta A80

No idea where that one is going to get used vs. the other. The UltraOcta has a claimed CPU frequency of 1800 and is either a dev or tablet platform, while the other has its GPU just clocked at 1600?

 

Comparing that UltraOcta A80 to Apple's A7 highlights just how limited these early Rogue cores become in the GfxBench 2.7 test.

Despite the Apple A7's GPU having double the clusters, the higher clock rate of the UltraOcta's G6230 pushes it to the higher frame rate for that particular GLES 2.0 workload. GLES 3.0 is obviously a different story.

 

The 6430 in the A7 is clocked around 450MHz, while the 6230 in the A80 at least at 600MHz if not more.

Rys confirmed that Manhattan is quite ALU bound. It would be a disaster if future games wouldn't be, otherwise all ULP GPU vendors would had placed their bets wrong.

 

That Allwinner UltraOcta A80 keeps on getting higher scores. It is now reporting a best score of 709 in Glbench 3.0, which puts in on a par with the best A330.

http://gfxbench.com/device.jsp?benchmark=gfx30&os=Android&api=gl&D=Allwinner UltraOcta A80

It actually beats ipad air in T-rex off-screen.

I wonder what frequency they are clocking the G6230 @ to get those results.

 

Those alpha blended fill rate tests in gfxbench3.0 don't help to estimate frequencies. The A80 gets roughly the same fillrate result; 4 vs 8 TMUs so assuming the A7 GPU is clocked around 450MHz it's either 900MHz for the A80 6230 (highly unlikely) or they're just cheating.

Fill-rate offscreen
iPadAir = 2929 MTexels/s (8 TMUs)
A80 = 2917 MTexels/s (4 TMUs)

Alpha blending offscreen
iPadAir = 6185 MB/s
A80 = 4354 MB/s

The alpha blending results could suggest a 600MHz frequency for the latter (vs. 450MHz on the A7), but the fillrate results sound like a dead end for now.

 

What about differences in efficiency and bottlenecks between the PowerVR 6230 and 6430?

 

I haven't the slightest idea how bandwidth looks like on A7 and even more so for A80 to have another important factor to think of.

The Manhattan test in GFXbench3.0 is quite ALU bound.

Apple A7 GPU G6430@450MHz (?)
4 * 64 * 0.45GHz = 115.2 GFLOPs FP32 ---> 13.2 fps (peak)

Rogue Han GPU G6230 = 11.4 fps

11.4 / 13.2 * 115.2 = ~100 GFLOPs FP32 / 128 ALU lanes = 0.78GHz = 780MHz

Now I believe that kind of frequency in an ultra thin tablet with an A80 when I see it. If any manufacturer can achieve it with a reasonable battery size and battery life in an ultra thin tablet all hail to Allwinner. In any other case "boo" as I'd react for any other player who tries to win impressions in benchmarks with unrealistic conditions.

 

The limiting factor in the T-Rex test has been covered here before, so the Allwinner A80 advantage over the Apple A7 makes sense there.

The improving Manhattan results for the A80 suggest some combination of perhaps, as mentioned, higher and higher frequency testing of the platform, possibly a different limiting factor for that test starting to take over at the higher frame rates, and/or, more optimistically, a good hardware and software implementation for the A80 platform in regard to drivers, available bandwidth, etc.

 

I'm assuming those recent A80 Glbenchmarks were achieved using this board:-

http://www.cnx-software.com/2014/04...-systems-products-to-become-available-in-may/

Skip to around 5:45.

"This is a board designed to show the maximum performance of A80"
Fan on top of the chip.

The video is a nice overview of Allwinner's products. Personally, I find that video blogger's accent/interview technique very irritating.

"is this secret".." is this special"

 

The best part is when he asks if they use the board to cheat, cause it has a fan. If they are using frequencies that cannot be sustained without active cooling then the GPU is not "similaah" to Apple's GPU

 

GDC talks are online.

Practical Techniques for Ray Tracing in Games
http://www.gdcvault.com/play/1020688/Practical-Techniques-for-Ray-Tracing

New Techniques Made Possible by PowerVR Ray Tracing Hardware
http://www.gdcvault.com/play/1020741/New-Techniques-Made-Possible-by

 

Somewhat OT but I albeit I'm all for changes and progress I can't say I'm all too crazy about the new IMG homesite layout.

A ducklin is sexy by itself; it doesn't need to accessorize:

 

The new website is terrible at high resolutions. What is it with so many website recently being designed like this? 50% of my page is empty with harsh white. At 4k resolutions it looks like around 75% of the page would be empty white.

It my mind it doesn’t look good for a graphics company that’s about tomorrows technology to have a new website that doesn’t factor in tomorrow or todays resolutions. Its like a website designed for 5 years ago.

 

The layout fits pretty well on my phone's display, recognition of the prevalence of mobile browsing usage these days.

 

Then they obviously need something that fits them all and no it's not rocket science either to either scale content on a webpage according to resolution or just have a mobile and desktop variant of each page.

 

Any reason to believe that with Nvidia and now AMD, IMG will fall behind in mobile GPU performance?

 

I don't think so. IMG is the only player in town with TBDR, which is ideal for mobile. That'll give them a clear edge methinks.