News [CNBC] AMD, Samsung partner on mobile graphics tech

[igor_kavinski]

Samsung can still use these dies to make the budget gamer's chromebook. That will also afford them more battery life and a higher power budget, allowing the RDNA2 chip to clock higher and deliver decent fps.

 

[TESKATLIPOKA]

There are still some unanswered questions to simply put the blame for the weaker performance on RDNA2 architecture.
How much space does this GPU need compared to the competition?
What is the actual frequency and voltage?

 

[Hitman928]

To be honest, I'm not sure why Samsung did this. Latest Mali would have been at least as good.

Now they can advertise it as having powerful desktop GPU features (ray tracing, VRS, etc.). Even if they are just a waste of space in mobile (ray tracing) or vastly unsupported in mobile games (VRS), it looks good for marketing. I suspect that they will also want to use their RDNA2 IP licensed features if they want to push more into laptops where these features might actually be of use.

Based upon the review @moinmoin posted, it also looks like they are adjusting how much power can be consumed by the SOC to extend battery life which obviously hurts performance but greatly increased battery life over the prior generation during gaming.

 

[Bigos]

I have not seen any mention of this Samsung SoC RDNA2 graphics in open source drivers (both kernel and mesa). As such, it looks like Samsung is developing all of the software in-house. Please correct me if I'm wrong, I will gladly revise the following paragraph if that is the case.

I generally have as much faith in Samsung software as I have in the ads I receive in spam emails. I would thus not easily rule out Samsung being incompetent enough to screw out the drivers. I can't fathom how they could not have just used the open source solution that is readily available on Linux. Sure, their SoC is specific so it would require some kernel changes, but mesa should be usable as-is. I hate companies that develop "open-source" software behind the scenes and do not collaborate with the other players when it makes sense.

Having said that, we don't really know how much poor software affects the performance and power efficiency. It would take dedicated individuals years of development to port the open source Linux drivers to see how it works (it wouldn't be the first time third-party developers fix Samsung software).

 

[moinmoin]

I have not seen any mention of this Samsung SoC RDNA2 graphics in open source drivers (both kernel and mesa). As such, it looks like Samsung is developing all of the software in-house. Please correct me if I'm wrong, I will gladly revise the following paragraph if that is the case.

I generally have as much faith in Samsung software as I have in the ads I receive in spam emails. I would thus not easily rule out Samsung being incompetent enough to screw out the drivers. I can't fathom how they could not have just used the open source solution that is readily available on Linux. Sure, their SoC is specific so it would require some kernel changes, but mesa should be usable as-is. I hate companies that develop "open-source" software behind the scenes and do not collaborate with the other players when it makes sense.

Having said that, we don't really know how much poor software affects the performance and power efficiency. It would take dedicated individuals years of development to port the open source Linux drivers to see how it works (it wouldn't be the first time third-party developers fix Samsung software).

While Android is (still) running with a Linux kernel it's unfortunately customary in the mobile phone chipset manufacturer space to just offer binary blobs as driver modules. So Samsung's driver code is likely based on AMD's closed source drivers, if even that. It would be great to have more open ARM based ecosystems in mobile phones.

 

[DisEnchantment]

Android uses mainly OpenGL ES 2.0 via the EGL library. User space drivers I have seen are different from upstream ones.
AMD's RDNA2 architecture is not OpenGL ES friendly, it is lacking many features and is emulating them where possible, like ASTC for example.
Performance will tank when running popular apps which have been in the Store for a while.
RDNA2 is more compatible with OpenGL 4.5+.

Android R is using 5.4 kernel, and minimum version needed for proper functioning of all RDNA2 HW blocks(like SDMA) features is 5.9 if I am not mistaken, according to Alex/Christian.
So no direct reuse of upstream code possible.

When used with Vulkan it should be better, but most Android apps are not using Vulkan yet. And only native apps can use Vulkan.
Also Vulkan driver is accessed via the Android Vulkan runtime along with Gralloc/BufferQueue and not directly.
And RT on Linux is only accessible via Vulkan anyway which was added recently to the spec.
Maybe with Android 14 /T (experimental) and above (Kernel 5.10+), RDNA2+ will work better but for now it is too forward looking.
But it is OK, the Exynos 2200 chip will be much better for other use cases like Automotive or Industrial when using Linux directly.
When the HW architecture evolves, SW is going to better as well.

 

[Frenetic Pony]

I think this thread is missing the perf/per watt. Which is quite good. We can look past benchmarks here I think and appreciate that the ported arch is really quite efficient, in comparison with another companies Snapdragon 8 Gen 1 implementation the Exynos 2200 "only" manages 30fps in Genshin Impact vs 48 for Sanpdragon, but it draws an entire two watts less. With 6.4 vs 4.4 watts, given that the GPU isn't the only thing drawing power, we can see the 34% drop in performance vs 32% less power should mean the 920 is running straight up more efficiently here.

Why Samsung chose to run the chip so low I don't know, though I can guess it involves competing with Apple in at least one category. And judging by that battery life while actually playing a game, matching the Ultra the big Iphone 13 pretty much 1 to 1, we can see there they succeeded. It certainly doesn't gel with their "Teh Gameeerz RGB lightup phone!" marketing they put out. And it isn't the world's most performant chip in any way, but it doesn't mean it was a failure.

 

[Tup3x]

Now they can advertise it as having powerful desktop GPU features (ray tracing, VRS, etc.). Even if they are just a waste of space in mobile (ray tracing) or vastly unsupported in mobile games (VRS), it looks good for marketing. I suspect that they will also want to use their RDNA2 IP licensed features if they want to push more into laptops where these features might actually be of use.

Based upon the review @moinmoin posted, it also looks like they are adjusting how much power can be consumed by the SOC to extend battery life which obviously hurts performance but greatly increased battery life over the prior generation during gaming.

Latest Mali supports ray tracing. Not sure about VRS. VRS benefits are small and usually if it offers any performance boost, you will also notice the visual downgrade.

If they plan to make some kind of new SoC for their laptops, then I understand.

 

[soresu]

Latest Mali supports ray tracing. Not sure about VRS. VRS benefits are small and usually if it offers any performance boost, you will also notice the visual downgrade.

With foveated rendering in VR it would not be noticeable if lower shading rate is concentrated in the visual periphery.

100% likelihood that PSVR2 optimised titles will make use of this and other FR related techniques.

I'm pretty sure that the latest QC Adreno PR materials mentioned VRS too, so I would be surprised to find the feature absent in Mali next gen, though I've yet to hear of a single halfway decent VR HMD using a Mali based SoC.

 

[uzzi38]

I think this thread is missing the perf/per watt. Which is quite good. We can look past benchmarks here I think and appreciate that the ported arch is really quite efficient, in comparison with another companies Snapdragon 8 Gen 1 implementation the Exynos 2200 "only" manages 30fps in Genshin Impact vs 48 for Sanpdragon, but it draws an entire two watts less. With 6.4 vs 4.4 watts, given that the GPU isn't the only thing drawing power, we can see the 34% drop in performance vs 32% less power should mean the 920 is running straight up more efficiently here.

Why Samsung chose to run the chip so low I don't know, though I can guess it involves competing with Apple in at least one category. And judging by that battery life while actually playing a game, matching the Ultra the big Iphone 13 pretty much 1 to 1, we can see there they succeeded. It certainly doesn't gel with their "Teh Gameeerz RGB lightup phone!" marketing they put out. And it isn't the world's most performant chip in any way, but it doesn't mean it was a failure.

Android Authority found that sustained performance in the same chassis (S22 Ultra) is about the same in GPU tests: https://www.androidauthority.com/snapdragon-8-gen-1-vs-exynos-2200-3122407/

On the CPU side it's a clear loss for Samsung, so it's clear they've done a worse job at their physical design in that aspect also.

 

[Tup3x]

Received my S22 few days ago. It looks like OpenGL ES is handled by ANGLE and it uses Vulkan backend.

 

[DisEnchantment]

Received my S22 few days ago. It looks like OpenGL ES is handled by ANGLE and it uses Vulkan backend.

Nice info. I am not sure how much perf degradation when running through this translation layer, but I believe it won't be negligible.
But Google is moving ahead with Vulkan for Android and Khronos already have a profile for Android.
I think Android T(current preview) or U should have broad support for Vulkan throughout like in WebView etc.

Vulkan-Profiles/profiles/VP_ANDROID_baseline_2021.json at main · KhronosGroup/Vulkan-Profiles

Vulkan Profiles Tools. Contribute to KhronosGroup/Vulkan-Profiles development by creating an account on GitHub.

github.com

I guess for the ASTC texture decompression AMD would need to support in HW eventually instead of doing in SW. But I hope this does not lead to bloated HW for desktop.
Radeon Vulkan drivers are excellent (RADV+AMDVLK)

 

[Frenetic Pony]

Android Authority found that sustained performance in the same chassis (S22 Ultra) is about the same in GPU tests: https://www.androidauthority.com/snapdragon-8-gen-1-vs-exynos-2200-3122407/

On the CPU side it's a clear loss for Samsung, so it's clear they've done a worse job at their physical design in that aspect also.

Small win for the new 920 over Snapdragon, and that's paired with a worse CPU config, so the win might be slightly above that even. Still, definitely not a huge win.

And it's... familiar to see a mixup like this between engineering and marketing. All those, obviously false, headlines blasting that the new GPU is terrible have already gone out. Engineering and support want a good customer experience on Exynos and so throttle to save battery life, while whoever's in charge of the Snadragon body knows not to throttle anything because the tech press is largely awful. C'est la vie.

 

[Thala]

All those, obviously false, headlines blasting that the new GPU is terrible have already gone out. Engineering and support want a good customer experience on Exynos and so throttle to save battery life, while whoever's in charge of the Snadragon body knows not to throttle anything because the tech press is largely awful. C'est la vie.

Snapdragon still has the much more powerful GPU, despite it is performing about the same when thermally throttled in the Samsung phones. If you look at other phones with the Snapdragon with a better cooling solution - it is performing much better as well. I'd say the Exynos GPU is still terrible in the bigger picture.

 

[NTMBK]

Samsung are planning to keep going with RDNA: https://www.tomshardware.com/news/samsung-to-keep-using-amd-rdna-gpus-for-socs

I'm glad they're keeping going. Hopefully the second version will be a big improvement.

 

[darkswordsman17]

I have a strong hunch that software support is a big reason why. Which right now the AMD GPU is likely at a significant disadvantage as its newer to this market and earlier posts indicate its definitely not being fully utilized with the current state of Android and the like. But that will change.

I have a hunch that AMD GPUs will make porting games much much easier. Might not matter much right now, but it could matter a lot in the future. Especially as VR/AR headsets develop.

I think the bigger issue is that Samsung didn't properly develop a device to play to the strengths of this SoC and instead just stuffed it into their phone (whilst marketing it like a premium gaming option when that is the current state of things for it on Android). Now perhaps if they'd done something to address that that would've been ok. Or if they'd released a Switch like gaming tablet or a Quest like standalone VR headset, I think it would be much more well received. Maybe something like the Steamdeck that could run both Android and SteamOS.

 

[Tup3x]

Samsung are planning to keep going with RDNA: https://www.tomshardware.com/news/samsung-to-keep-using-amd-rdna-gpus-for-socs

I'm glad they're keeping going. Hopefully the second version will be a big improvement.

They need to improve a lot... and I hope games switch to using Vulkan since this thing has same issue with OpenGL ES as Intel with DX9. Though, even in Vulkan the performance is is behind Qualcomm and perf/W is not where it should be. To be honest, I excepted a lot more when I upgraded from Kirin 970 (Huawei P30) to S22.

 

[Frenetic Pony]

They need to improve a lot... and I hope games switch to using Vulkan since this thing has same issue with OpenGL ES as Intel with DX9. Though, even in Vulkan the performance is is behind Qualcomm and perf/W is not where it should be. To be honest, I excepted a lot more when I upgraded from Kirin 970 (Huawei P30) to S22.

No it's not, it's ahead. Most long term test, which throttle to the same watts and are vastly preferred by good reviewers, show a slight advantage for the Exynos. Links: 1l 2

BTW as these two SOCS/Phones have the same cooling solutions and chips built on the same process this is the most fair comparison you could possibly make, so yeah those two tests are super valid.

Probably it'll only get better if the next version is RDNA3 based, challenging Apple in perf per watt certainly isn't out of the question if this gets a similar bump as the desktop versions in efficiency.

The big problem is the rest of the SOC. Samsung doesn't, or at least didn't, have the management to get a good SOC made, instead budgeting no time to get miracles done. Recently their head of silicon was fired and a new one moved in, so maybe things will change for the better.

 

[Tup3x]

No it's not, it's ahead. Most long term test, which throttle to the same watts and are vastly preferred by good reviewers, show a slight advantage for the Exynos. Links: 1l 2

BTW as these two SOCS/Phones have the same cooling solutions and chips built on the same process this is the most fair comparison you could possibly make, so yeah those two tests are super valid.

Probably it'll only get better if the next version is RDNA3 based, challenging Apple in perf per watt certainly isn't out of the question if this gets a similar bump as the desktop versions in efficiency.

The big problem is the rest of the SOC. Samsung doesn't, or at least didn't, have the management to get a good SOC made, instead budgeting no time to get miracles done. Recently their head of silicon was fired and a new one moved in, so maybe things will change for the better.

No offense, but I've had the "pleasure" to se this since launch and it does not compare with Snapdragon. When it comes to battery longevity and real world performance it just is not on par. If this would be as fast as Snapdragon the heat and power usage would be horrendous (or throttle the Snapdragon and it would look even better). There are plenty of guys here in Finland who have bought Snapdragon version and they all say that it is much better than the Exynos version.

Samsung's process is one part of the problem but not only is the rest of the SoC mediocre, RDNA2 isn't quite ready for phones yet (or Samsung did a poor job implementing it).

 

[TESKATLIPOKA]

Even If Xclipse 920 doesn't perform as Adreno 730, It's still faster than the previous generation Snapdragon and Exynos. That's not bad for a first try knowing Samsung N4 process is not the best.
Maybe the frequency(1.3GHz ?) is too high and need higher voltage resulting in worse power efficiency on this process. Maybe If they added 2 more CU and set clocks at 1GHz to get comparable performance, It would have been better.

Specs from Wikipedia:

SoC	GPU	Frequency	GFlops in FP32
Exynos 2100	Mali G78 MP14	854 MHz	1,530
Snapdragon 888	Adreno 660	840 MHz	1720.3
Exynos 2200	Xclipse 920 6CU ?	1300 MHz ?	998.4 ?
Snapdragon 8 Gen 1	Adreno 730	818 MHz	2236.1

I don't think frequency and GFlops are correct for these SoCs.

I am glad that Samsung didn't stop Its cooperation with AMD. RDNA3 in mobile should look a lot better than RDNA2. We will see with Phoenix Point how good RDNA3 is at limited TDP.

edit: correct naming is 4N -> 4nm

 

[moinmoin]

That's not bad for a first try knowing Samsung N4 process is not the best.

Just a side node since this is easily getting confusing (especially with Intel joining and renaming), the N4 style moniker are names by TSMC. Samsung still uses sizes verbatim, in this case 4nm.

 

[Joe NYC]

Samsung still pursuing its independent designs, which still include RDNA2

Samsung Galaxy S23 will receive a 3-nm Exynos Quadra SoC. Samsung will continue to use AMD's RDNA2 architecture architecture - Digit News

 

[TESKATLIPOKA]

Samsung still pursuing its independent designs, which still include RDNA2

Samsung Galaxy S23 will receive a 3-nm Exynos Quadra SoC. Samsung will continue to use AMD's RDNA2 architecture architecture - Digit News

In the article, there is no mention of actually using RDNA2, It could be also RDNA3. I hope It's based on RDNA3.

 

[DisEnchantment]

Samsung Electronics and AMD Extend Strategic IP Licensing Agreement To Bring AMD Radeon™ Graphics to Future Mobile Platforms

Companies broaden scope of mobile graphics collaboration to bring leadership AMD Radeon graphics technology to expanded portfolio of Samsung Exynos SoCs

news.samsung.com

Samsung Electronics, a world leader in advanced semiconductor technology, and AMD (NASDAQ: AMD) today announced they have signed a multi-year agreement extension to bring multiple generations of high-performance, ultra-low-power AMD Radeon graphics solutions to an expanded portfolio of Samsung Exynos SoCs. Through the licensing extension, Samsung will bring console-level graphics quality and optimized power consumption to more mobile devices, offering an incredibly immersive and long-lasting gaming experience.

Looks like Samsung is getting RDNA3 and RDNA4 generations.
I hope there is a version of RDNA3 or RDNA4 somewhere in David Wang's drawer that cut back some of the ML capabilities in RDNA. And hopefully without the supposed bugs of the RB+ etc.
Lots of markets don't need bloated multi purpose CU/WGPs. Mobiles already have dedicated ML inferencing block and Steam Decks don't need it as much.

AMD's open source Linux drivers and user space mesa libraries are great, might be worth it to go with Radeon IP just because of this. Having worked with some of our SoC suppliers their drivers don't even pass Vulkan Compliance test fully and nobody knows what is happening when things wont work (no sources and nobody in the open knows what problems were there to even suggest some solution).

 

[Tup3x]

I sure hope that this doesn't mean that they'll use Exynos in their S series phones next year. Exynos 2200 was rather bad and the GPU part was especially lacking.