Samsung in talks with AMD & Nvidia for GPU ?

[tangey]

http://www.sammobile.com/2016/09/09...lks-with-nvidia-and-amd-for-gpu-technologies/

I don't know if this website has any history either for being accurate, or being a load of rubbish.

Does AMD even have IP suitable for smartphones ?

Also Samsung was the major target of the Nvidia patent suit a while back, which according to reports the various claims where either thrown out or withdrawn. So any mileage that this has anything to do with that ?

 

[Deleted member 13524]

AFAIK, both nvidia and AMD have stated they wouldn't say no to selling GPU IPs, for quite a while.

http://www.anandtech.com/show/7083/nvidia-to-license-kepler-and-future-gpu-ip-to-3rd-parties
http://www.kitguru.net/components/g...ics-ip-licensing-opportunities-going-forward/

 

[Ailuros]

http://www.sammobile.com/2016/09/09...lks-with-nvidia-and-amd-for-gpu-technologies/

I don't know if this website has any history either for being accurate, or being a load of rubbish.

Does AMD even have IP suitable for smartphones ?

Also Samsung was the major target of the Nvidia patent suit a while back, which according to reports the various claims where either thrown out or withdrawn. So any mileage that this has anything to do with that ?

Is the author affiliated with wccftech?

There might be some weird rumor circulating behind the curtains and Samsung may very well investigate for other options at Lord knows where. If true it means that whatever they've been internally experimenting with for an in house GPU was/is completely useless (remember there were repeated reports in the rumor mill over the years, that they're developing a ULP SoC GPU of their own....)

Besides that neither AMD's nor NVIDIA's business model bodes well with pure IP licensing, even if either/or would license IP it wouldn't be cheap. I'm left under the impression that they're stuck with ARM GPU IP because it comes significantly cheaper for them than with competing IP solutions. Which would mean what, that Samsung LSI would be willing overnight to spend times more for GPU IP than ever before?

I won't say impossible just yet, but the odds from a layman's perspective don't sound particularly encouraging.

 

[Wynix]

Since this is a pretty crazy rumour, let me talk out of my ass for a bit.

If we compare Samsung's latest SoC with Qualcomm's(who fabricate at Samsung foundries), the superiority of Qualcomm's GPU is obvious.
What if the motivation isn't purely about performance or even perf/watt but about perf/mm2 or perf/watt/mm2 and if the Adreno is significantly ahead, enough for Samsung to consider an expensive IP license deal worth pursuing, and then they more than make up the cost difference with reduced die size on every SoC.

 

[Ailuros]

Qcom could be likelier than AMD or others at least, but from public benchmark results (those that show throttling included) I don't see any supposed superiority for the Adreno530 vs. the 12 cluster T880 in Samsung's latest Exynos. In fact they're pretty damn close both in performance as well as when it comes to throttling. There's no public data for perf/mm2 since none of them mentions die area for the respected GPUs anywhere unless I've missed something and if QCOM should have a full tessellation pipeline for DX11 and doesn't channel stuff like tessellation through its ALUs then chances are high that QCOM's 530 consumes more die area after all.

 

[Nebuchadnezzar]

Qcom could be likelier than AMD or others at least, but from public benchmark results (those that show throttling included) I don't see any supposed superiority for the Adreno530 vs. the 12 cluster T880 in Samsung's latest Exynos. In fact they're pretty damn close both in performance as well as when it comes to throttling. There's no public data for perf/mm2 since none of them mentions die area for the respected GPUs anywhere unless I've missed something and if QCOM should have a full tessellation pipeline for DX11 and doesn't channel stuff like tessellation through its ALUs then chances are high that QCOM's 530 consumes more die area after all.

The A530 is half the size of the T880MP12 in the 8890. ARM is well aware of their PPA disadvantage and that's why G71 focused on the improved area efficiency.

 

[Deleted member 13524]

The A530 is half the size of the T880MP12 in the 8890. ARM is well aware of their PPA disadvantage and that's why G71 focused on the improved area efficiency.

Wow...
So that's why Mediatek won't implement any Mali GPU above MP4. They seem pretty huge, at least compared to Adreno.
Can you give any rough area comparison between the Adreno 530 and the A9's GT7600?

 

[Ailuros]

The A530 is half the size of the T880MP12 in the 8890. ARM is well aware of their PPA disadvantage and that's why G71 focused on the improved area efficiency.

Thanks Nebu; wasn't aware of that. Has the 530 a fixed function tessellation unit?

Wow...
So that's why Mediatek won't implement any Mali GPU above MP4. They seem pretty huge, at least compared to Adreno.
Can you give any rough area comparison between the Adreno 530 and the A9's GT7600?

MTK has always been extremely conservative with the GPUs in their SoCs. For smartphones for example they've been using for their higher end SoCs since MT6595 up to 6795 (HelioX10) a dual cluster G6200 with frequencies from 600 to 700MHz depending on case, that's 4 TMUs in all cases. Helio X20/25 right now use a quad Mali T880 at 780/860MHz (again 4 TMUs) and their GPU performance will only rise by a sizeable portion in the upcoming Helio X3x SoCs with a quad GT7400 (8 TMUs) at up to 800MHz.

 

[Nebuchadnezzar]

Wow...
So that's why Mediatek won't implement any Mali GPU above MP4. They seem pretty huge, at least compared to Adreno.
Can you give any rough area comparison between the Adreno 530 and the A9's GT7600?

Without saying mm², let's say 8890 is 1, A9 is 0.8-0.9 (depending on version) and A530 is 0.5. Qualcomm definitely has a large PPA lead however they're also extremely optimised/specialised in terms of implementation while the other two are IPs that can scale much higher in perf and implemented by third-parties.

Wow...
So that's why Mediatek won't implement any Mali GPU above MP4. They seem pretty huge, at least compared to Adreno.

MTK has always been extremely conservative with the GPUs in their SoCs. For smartphones for example they've been using for their higher end SoCs since MT6595 up to 6795 (HelioX10) a dual cluster G6200 with frequencies from 600 to 700MHz depending on case, that's 4 TMUs in all cases. Helio X20/25 right now use a quad Mali T880 at 780/860MHz (again 4 TMUs) and their GPU performance will only rise by a sizeable portion in the upcoming Helio X3x SoCs with a quad GT7400 (8 TMUs) at up to 800MHz.

PowerVR doesn't seem to be that far in advantage, I think it's just a case of MediaTek realizing that most people don't actually need larger GPUs and every cent/mm² counts in their target market. They have crazy small SoCs.

Thanks Nebu; wasn't aware of that. Has the 530 a fixed function tessellation unit?

Wasn't that something they advertised on the 430? I forgot.

 

[Ailuros]

No idea; I'm just sure that ARM Mali 7xx/8xx channel anything tessellation through the compute pipeline. PowerVR has yet to show a DX11 solution in hw (unless of course Adrenos also take a similar path as Malis for tessellation).

First sentence above in your post doesn't make QCOM's GPU IP a good candidate for what Samsung might want/need since it lacks flexibility (assuming QCOM is interested in licensing IP in the first place).

 

[tangey]

Nothing has been formally revealed about 8xt. 8xe talks about "up to 25%" area reduction compared to 7xe, and up to 100% per/mm2 improvement compared with "competing devices"

If some of the above rubs off into 8XT, it could suggests an increased advantage over Mali gpus in terms of size.

 

[Ailuros]

More die area isn't necessarily a bad thing if it lowers power consumption. Dedicated FP16 units might be relatively cheap, but when you go for a 2:1 ratio for FP32 SPs then it'll pile up eventually (the Apple A9 GPU has 192 FP32 & 384 FP16 SPs as a reminder). S8XT is a complete unknown yet and I'd assume DX11 is still obligatory; but if it shouldn't be, then it will add further in die area.

8XE as any former XE line of cores have as their highest priority perf/mm2, while anything XT on the other hand have as their highest priority perf/mW. Quite the same for any other vendor too in the ULP mobile market. http://images.anandtech.com/doci/10077/8XE-4.png?_ga=1.36146146.334006868.1458215934

 

[el etro]

If the volume selling is big enough to earn AMD/NV great profits is a huge deal. I don't think that any of two lacks behind current mobile GPU makers.

 

[Ailuros]

If the volume selling is big enough to earn AMD/NV great profits is a huge deal. I don't think that any of two lacks behind current mobile GPU makers.

That's not the real problem for an AMD or NV scenario. Neither company is laid out for pure IP licensing. It's hardly a coincidence that all of AMD's recent console deals are for semi custom SoCs and not just for IP.

 

[Dave Baumann]

https://www.extremetech.com/computi...nture-to-build-servers-for-the-chinese-market