PowerVR Furian Architecture

It depends (TM). There's a lot of MULs but the ALU performance "per flop" is probably not higher for 8XT than for 7XTP (especially as 7XTP had serial FMAs, unlike 7XT and 8XT where these ALUs are parallel) - the benefit is that it's higher perf/mm2, not necessarily higher perf/flop. But for the USC overall rather than just the ALU, we also gain efficiency from more pipelines processing in parallel with the "primary" pipeline for "free", e.g. complex ops (reciprocal etc.), conditional branches, texturing, memory load/store, etc...

And a lot of low-level optimisations to avoid various real-world bottlenecks that can't be described just by looking at peak rates. In terms of overall performance, for a lot of mobile workloads (including the most complex benchmarks), the ALU:TEX ratio was arguably slightly too high on 7XTP, so in terms of overall perf/mm2 we benefit from having slightly fewer flops per pixel/clk - and for higher-end cores/content, we have the ability to scale up to 3 USCs per 8-wide TPU (vs 2 USCs currently).

Also we tend to talk a lot about ALU efficiency, but I think it's worth highlighting that the new 8-wide TPU is *really* efficient - it's a lot smaller than 2x the 7XTP 4-wide TPU at pretty close to 2x the performance :) Geometry performance is also great, framebuffer compression is slightly improved, etc... All of that is less "sexy" than ALU changes so it doesn't get as much attention, but it all adds up to a very efficient and balanced architecture.
 
It's notable in the IMG PR yesterday about the first announced Furian Core 8525, that in comparing it to 7200, the comparison were big on performance benchmarks and overall Gflops improvements, but light on size/power improvements. Even though the list of "more than" comparative statements was precursored with
The Furian architecture is designed for enhanced performance, with power efficiency, setting it apart from competing solutions in performance per mW. Compared to the Series7XT GT7200 GPU, the GT8525 achieves:
None of the comparative statements mentioned power or size advantages. It seems logical that the list was on a per-clock comparative basis, but without any inclusion of comparative power/size, the list doesn't actually give an indication of expectations for either area efficency or performance per mW improvements.
 
Last edited:
Sorry, I don't think I can comment more precisely on area/power despite how many area spreadsheets and synthesis reports I've looked at over the last few months :) Perf/mm2 and perf/watt are both significantly better than 7XT(P).
 
is it possible(in soc smartphone) GT8525 clock at 1GHz on TSMC 16nm++? i ask it because MT6795 (Helio X10) have PowerVR 6200@700MHz and MT8176 have PowerVR 6250@650MHz(All on 28HPm process). if yes i think that at Helio X30 same performace or better because of new architecture(192GFlops and 204,8GFlops). by the way TSMC's 16FF+ (FinFET Plus) technology can provide above 65 percent higher speed.
 
is it possible(in soc smartphone) GT8525 clock at 1GHz on TSMC 16nm++? i ask it because MT6795 (Helio X10) have PowerVR 6200@700MHz and MT8176 have PowerVR 6250@650MHz(All on 28HPm process). if yes i think that at Helio X30 same performace or better because of new architecture(192GFlops and 204,8GFlops). by the way TSMC's 16FF+ (FinFET Plus) technology can provide above 65 percent higher speed.

Of course is it possible to clock a 8525 at 1GHz but it won't come without any penalties either since very high frequencies always are subject to intense throttling in a small passively cooled device. The GT7400P in the Helio X30 clocks at 800MHz and the SoC is manufactured under 10FF TSMC.
TSMC might claim for its own reasons how much more frequency each new process or added die area allows (or a combination of both) but manufacturers usually avoid to push things to the boundaries of each process to keep things on the safe side. In order to add N persentage for one design compared to another die area would have to be the same which I don't think is the case for a 8525 vs. a 6250.

They probably managed to cut back some area ifor 8XT, but there are stil quite a few new functionalities added in Furian, which weren't present in the initial Rogue variants.
 
imagination just announced four cluster powervr furian GT8540 and make comparison against series 7xt GT7450, i think it just rebrand name of GT7400 plus on Helio X30 GPU.
The result

So what are the results of all these changes? As we can see from the graph below, in the popular industry standard Kishonti GFXBench Manhattan 3.0 benchmark, the Series8XT GT8525 offers significantly more frames per second (fps) per watt, or fps/W, than our PowerVR GT7450, our previous generation equivalent GPU hitting around 35fps compared to 15fps previously. (‘Frames per second’ refers to how smoothly a game might run on a device, so the higher number the better for the end-user experience). That it does so in a smaller silicon area (as indicated by the smaller circle) is an additional benefit.
Buckle-1.png

PowerVR GT8525 vs GT7450 – Manhattan 3.0 FPS vs FPS/W and Area

@Ryan Smith @Nebuchadnezzar
to verify that figure, any plan anandtech to make review with gpu power analysis for meizu pro 7 with helio x30?
 
10wpod5.png

source
data from anandtech for Manhattan 3.0 Offscreen Power Efficiency (System Load Power) highest only 5.71 fps/W vs. 15 fps/W, maybe because different how to measure?
 
imagination just announced four cluster powervr furian GT8540 and make comparison against series 7xt GT7450, i think it just rebrand name of GT7400 plus on Helio X30 GPU.

Buckle-1.png

PowerVR GT8525 vs GT7450 – Manhattan 3.0 FPS vs FPS/W and Area

@Ryan Smith @Nebuchadnezzar
to verify that figure, any plan anandtech to make review with gpu power analysis for meizu pro 7 with helio x30?


Probably just an innocent typo but the graph is comparing the dual cluster 8525 against the (4 cluster) 7450 and not the 8540.

source
data from anandtech for Manhattan 3.0 Offscreen Power Efficiency (System Load Power) highest only 5.71 fps/W vs. 15 fps/W, maybe because different how to measure?

I think Anandtech is comparing fps/SoC W vs. IMG fps/GPU W.
 
source
data from anandtech for Manhattan 3.0 Offscreen Power Efficiency (System Load Power) highest only 5.71 fps/W vs. 15 fps/W, maybe because different how to measure?
Yes one is measuring system power one is measuring solely GPU power.

I don't have a Pro 7 - I can try to reach out to Meizu in regards to it but can't promise anything.
 
Probably just an innocent typo but the graph is comparing the dual cluster 8525 against the (4 cluster) 7450 and not the 8540.

sorry for my typo lol, but the good news is the possibility of powervr furian to raise the efficiency more higher with wide cluster and lower clock is very large.

I think Anandtech is comparing fps/SoC W vs. IMG fps/GPU W.
Yes one is measuring system power one is measuring solely GPU power.

I don't have a Pro 7 - I can try to reach out to Meizu in regards to it but can't promise anything.

you're right and confirmed by Nebuchadnezzar. On that blog said, "the performance/W goes up to an incredible 75%! For the end user, this means that they will be able to enjoy a device that will last for longer, especially when under load, such as when gaming." if i make rough estimate based on that 75% system power+display Meizu Pro 7 Helio X30 is around 4W so that if mediatek only change gpu from rogue 7XT+ to furian on Meizu pro 7 Helio X30 the power consumption is around 2.28W @ 40FPS on GFXBenchmark Manhattan 3.0.

I hope high end SOC like Samsung and Exynos realize how good power furian is and use it.
 
the possibility of powervr furian to raise the efficiency more higher with wide cluster and lower clock is very large.

i said that because powervr furian GT8525 need 1066.66 mhz to equal GFXBenchmark manhattan 3.0 score of powervr series 7xt+ @800 mhz
 
if i make rough estimate based on that 75% system power+display Meizu Pro 7 Helio X30 is around 4W so that if mediatek only change gpu from rogue 7XT+ to furian on Meizu pro 7 Helio X30 the power consumption is around 2.28W @ 40FPS on GFXBenchmark Manhattan 3.0.
Display is always subtracted from these figures, and 4W is too low for active system. GPUs are generally around 30-50% of active system power, that image shows 14fps/W at 40fps meaning around 2.85W so system power should be at least 5.7Wish which is only 7fps/W system.

We don't publish MH3.0 figures anymore but I still have them; 8890 is 7.5fps/W, 8895 8.5fps/W, 820 8.7fps/W and the 835 is at 15fps/W.
I hope high end SOC like Samsung and Exynos realize how good power furian is and use it.
Doubtful they'll ever use it given that they're doing their own GPU.
 
Last edited:
Oh they're still wasting millions in that infamous GPU project? :p Monkey see monkey do considering Apple, but the latter at least had decent material to steal from ;)

On a sidenote: IMG needs volume right now and I'm less concerned about high end IP like Furian. The best thing that could happen to them right now is Series9 low end and mid range IP reaching more than a few good deals through Spreadtrum, Mediatek and the likes.
 

If the future landscape in that market should consist mostly of Samsung, Apple and QCOM then obviously yes (presupposition Samsung has viable GPU IP this time...). I still don't think that China f.e. would want a scheme like that to materialize, meaning that there still will be a very viable market for GPU IP irrelevant if the majority for that will come from ARM in the future or not.
 
There's really no large value semiconductor market in play today where a single source of an IP product is viable and desirable. There exist plenty of niche markets where just one design is enough, but in something high value and high risk, like complex highly-integrated consumer application processors, designers need choice in case of missteps.
 
Back
Top