ARM Bifrost Architecture

So G71 MP8 has about 2/3rds the performance of Adreno 530 in Snapdragon 820 and the same performance as Tegra K1.
 
There is an update coming to the Mate 9 to fix some GPU driver issues. Hopefully it will improve the GPU performance to match the results on the Kirin 960 presentation where it outperformed the Adreno 530.
 
There is an update coming to the Mate 9 to fix some GPU driver issues. Hopefully it will improve the GPU performance to match the results on the Kirin 960 presentation where it outperformed the Adreno 530.
Yeah those figures are on the much lower side with regards what to expect from G71.
Cheers
 
So G71 MP8 has about 2/3rds the performance of Adreno 530 in Snapdragon 820 and the same performance as Tegra K1.

As others pointed out above those first results are even lower than Huawei mentioned when they introduced the 960. I'd also expect that with some drive fine tuning they'll at least reach the first predicted values. As it stands now the G71M8@900MHz comes out lower than the T880MP12@650MHz, which sounds way too low for an architecture like Bifrost.

***edit: a lot better now: https://gfxbench.com/device.jsp?ben...li-G71&did=41625847&D=Huawei Mate 9 (MHA-xxx)

The driver overhead score is still way too low, which may be a sign of headroom for further optimisations and it still seems to throttle quite a bit both in TRex as well as in Manhattan3.1. 3-4 tests though for either/or is still a too weak sample.
 
Last edited:
As others pointed out above those first results are even lower than Huawei mentioned when they introduced the 960. I'd also expect that with some drive fine tuning they'll at least reach the first predicted values. As it stands now the G71M8@900MHz comes out lower than the T880MP12@650MHz, which sounds way too low for an architecture like Bifrost.

***edit: a lot better now: https://gfxbench.com/device.jsp?ben...li-G71&did=41625847&D=Huawei Mate 9 (MHA-xxx)

The driver overhead score is still way too low, which may be a sign of headroom for further optimisations and it still seems to throttle quite a bit both in TRex as well as in Manhattan3.1. 3-4 tests though for either/or is still a too weak sample.
Now it is much better.
 
Well scores improved quite a bit, however the driver overhead (2) score is still low, which seems to be a general trend with Malis for unknown reasons.
 
In a mobile game most likely yes. From what I recall it's measuring CPU overhead in Manhattan3.1.
I have compared fps of games running on a Mali/Adreno GPU (Note3/Note4) and Mali's would get better results compared to the superior raw performance of their counterparts ( the Note 3 Mali GPU was like half the raw power but just run games at about the same fps' as the Adreno 330 ). I don't know about current versions because I don't have them to compare and almost nobody seems to do this kind of tests ( just some chinese youtube accounts that are hard to find and difficult to understand for me xD ).
 
Here's how different mobile GPUs fare in the specific synthetic test: https://gfxbench.com/result.jsp?benchmark=gfx40&test=639&order=score&base=gpu&ff-check-desktop=0
As with all synthetic tests that target a very specific aspect, results are an indication for the specific aspect. It was created with the lower ÇPU overhead in graphics through low level APIs in mind.

I would not take it very seriously when it puts lower end Mali GPUs configurations over the Mali T880-MP12 inside the Exynos 8890. And there is a lot of variations between devices with the same SOC ( maybe som fake results are mixed in or tweaked kernels in action ). This tables are made with the uploaded results from users... not very trustworthy.
 
It's a low level synthetic test and Gfxbench is the most reliable GPU bound synthetic you can get for mobile SoC GPUs for years now. It's an indication and does not interpret into average gaming performance. And no the results aren't from garden variety users only, but members of the press, manufacturers and Kishonti personell amongst others.

All that it would mean at this stage, is that ARM should be working feverishly on its Vulkan drivers if it wants to have a lower CPU overhead in low level APIs in future Vulkan based games.
 
It's a low level synthetic test and Gfxbench is the most reliable GPU bound synthetic you can get for mobile SoC GPUs for years now. It's an indication and does not interpret into average gaming performance. And no the results aren't from garden variety users only, but members of the press, manufacturers and Kishonti personell amongst others.

All that it would mean at this stage, is that ARM should be working feverishly on its Vulkan drivers if it wants to have a lower CPU overhead in low level APIs in future Vulkan based games.

I was not questioning their tests, just the results that were shown. Yeah, there are results from reliable sources too, but the majority are from the garden users. When you see a low end GPU over a high end one, or the same soc getting twice the performance in different devices... you can see something is wrong.

And as synthetic benchmarks they are, they don't always represent real world gaming performance as I declared before: https://www.gamebench.net/blog/samsung-galaxy-note-4-beats-every-other-phone-asphalt-8
 
I said already that it doesn't represent real gaming performance and it won't be any issue either with an API like Vulkan either. If you're really interested to see what the specific low level test does, you can download it and watch it first hand.
 
Last edited:
Mali G71 MP20 and LPDDR4X for 30 to 34GB/s bandwidth (assuming 2*32bit) in 10nm Exynos 8895:

http://www.samsung.com/semiconducto..._Worlds_first_10nm_FinFET_Process_Technology/
http://www.anandtech.com/show/11149/samsung-announces-exynos-8895-soc-10nm

If we look at how well the G71 MP8 in the Kirin 960 fares right now, we might be looking at something that matches the apple A9X in GPU performance.
Depending on the clocks, it might even bite at the heels of an Iris Pro 580.
The samsung web page says 60% more graphics performance from 8890 so I guess it is clocked aroud ~450MHz.
 
Hum..then it's not going very high.
Maybe they're targeting thermally sustainable performance this time, for GearVR.
 
Hum..then it's not going very high.
Maybe they're targeting thermally sustainable performance this time, for GearVR.

Well HiSilicon went up to 900MHz in the Kirin 960 because they used "only" 8 clusters. Samsung is obviously going the opposite route where they're going wider and most likely with a way more modest frequency. 450MHz sounds like a very reasonable frequency if you want to almost entirely avoid throttling (remember the Mate9 can go from 35.3 in Manhattan 3.1 onscreen in the long term performance test down to 19.2 fps), however it sounds at least to me a wee bit low. Eventually if needed they could also implement a boost frequency like say 533MHz for quick spikes like for compute. It's still a very powerful config since we should remember that we are actually talking about 20 TMUs in a tiny smartphone SoC ;)
 
I like the pace at which ARM is interating on its GPU, G71 was the first steps of their new architecture, the G72 should be interesting to follow.
Contrary to Intel which seems to have eased its effort on the GPU front, ARM is improving its product and closing the gap with its competitors with every iterations of their products.
 
Back
Top