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

[gdansk]

What about the 6300M? 12 CU with a single GDDR6 chip consuming 25W. Without the dedicated memory it should be about 22W. Cut in half and we're already at 11W. Still high for a phone but half the problem is that Samsung's 4nm isn't better than TSMC's 6nm.

If RDNA2 was on TSMC 4nm like MediaTek's latest it would be very competitive.

Edit: Apparently Samsung reiterates it is OK and there are no problems with performance. Weird... why skip the event then?

 

[Hitman928]

15W is still a lot more than 6W. Can they still have top class performance rivalling the established players in the mobile space?

The 15W chip also includes an SOC with a lot more IO that will burn a few watts as well. Take out all the extra IO the phone SoC doesn't need, drop down to 6 CU and lower clocks, I don't see why it can't fit into a 6 W envelope. Obviously the performance is going to be a lot less than what a laptop APU can do, but can it do well against the competition in the mobile phone space? I think so. I guess we'll find out. One thing that might be a detriment to RDNA2's success in that sphere is the types of games are different than what people play on PC and so mobile only GPU architectures might be more tuned to the low geometry, low physics type of games you find on mobile but I don't really follow the mobile GPU designs to really know. The other thing to remember is that AMD isn't designing the GPU, they are licensing the IP to Samsung's Exynos team and the Exynos team has kind of been a hot mess (pun intended) lately so who knows how good the execution is either. Whenever it does launch, the reviews should be interesting. My guess is that it will probably under perform in terms of perf/w, but that the issue will lie more in the process and execution than in the architecture/IP as the IP has already been shown to be pretty well balanced and scalable.

 

[moinmoin]

Apparently Samsung reiterates it is OK and there are no problems with performance. Weird... why skip the event then?

Article was updated with this info:
"After nailing the news of the event cancellation, Ice Universe added on Sina Weibo that the AMD GPU is having problems specifically heat problems. The target frequency is apparently 1.9 GHz, but the companies are currently only seeing reasonable thermals at 1.29 GHz. This is resulting in lots of infighting at Samsung about who dropped the ball."

And looks like Galaxy S22, which was supposed to come with Exynos 2200, may contain only Snapdragon.

Samsung claims it will now launch Exynos 2200 alongside Galaxy S22 instead.

 

[NTMBK]

They canned it coz it would have gotten more bad press than good. Looks like RDNA2 doesn't work for mobile. How could it? Their silicon is not power optimized for handhelds. Samsung should have gone with Tegra something.

Remind me how many phones shipped with a Tegra 4i SoC...

Tegra was a flop, post Tegra 3. The only big success they had was the X1 in the Switch, and Nintendo had to heavily underclock it to get acceptable thermals and battery life.

 

[igor_kavinski]

Remind me how many phones shipped with a Tegra 4i SoC...

Tegra was a flop, post Tegra 3. The only big success they had was the X1 in the Switch, and Nintendo had to heavily underclock it to get acceptable thermals and battery life.

A pared down X1 (probably halved) could work. Haven't really compared its processing power with Adreno etc.

 

[igor_kavinski]

The other thing to remember is that AMD isn't designing the GPU, they are licensing the IP to Samsung's Exynos team and the Exynos team has kind of been a hot mess (pun intended) lately so who knows how good the execution is either.

Ah. No chance a team can just take an IP and just execute all the way through to a flawless product, unless it's Apple.

 

[Hitman928]

Ah. No chance a team can just take an IP and just execute all the way through to a flawless product, unless it's Apple.

They'll have support docs and FAEs to help guide and answer questions, but ultimately it will be up to the design team to execute and bring together the whole SoC.

 

[TESKATLIPOKA]

Article was updated with this info:
"After nailing the news of the event cancellation, Ice Universe added on Sina Weibo that the AMD GPU is having problems specifically heat problems. The target frequency is apparently 1.9 GHz, but the companies are currently only seeing reasonable thermals at 1.29 GHz. This is resulting in lots of infighting at Samsung about who dropped the ball."

And looks like Galaxy S22, which was supposed to come with Exynos 2200, may contain only Snapdragon.

Samsung claims it will now launch Exynos 2200 alongside Galaxy S22 instead.

1.9GHz as a target frequency? For a smartphone GPU? I am a bit sceptical about this info, because RDNA2 is not the most power efficient at 1.9GHz, but at a lower frequency(1.5GHz is my guess).
I checked and I couldn't find any smartphone GPU clocked higher than 1GHz.
BTW 6CU RDNA2 at 1.29GHz has similar TFLOPs as Adreno 650.

 

[Stuka87]

1.9GHz as a target frequency? For a smartphone GPU? I am a bit sceptical about this info, because RDNA2 is not the most power efficient at 1.9GHz, but at a lower frequency(1.5GHz is my guess).
I checked and I couldn't find any smartphone GPU clocked higher than 1GHz.
BTW 6CU RDNA2 at 1.29GHz has similar TFLOPs as Adreno 650.

RDNA2 is actually very efficient at high clocks. The RX 6500 XT for instance runs at 2,610 MHz with a TDP of 107W. Yes, this is obviously way above what a mobile GPU would want. But that is a crazy high clock speed for that TDP considering an RX 580 ran at almost half the clock speed and used double the power. And with the way GPU power scaling typically works, dropping the clock to 1,500 MHz would probably get the power down to 30W or so (educated guess). And that's with a much more powerful GPU than what would be required for a mobile device.

 

[gdansk]

I think the 25W 6300M does suggest 1500 MHz is efficient for RDNA2 on 6nm.
But based on other tests of Samsung's 4nm it doesn't look like a good match.

 

[TESKATLIPOKA]

We are talking about a GPU for a smartphone and in that case power consumption and thermals should have the highest priority, so you need to keep voltage as low as possible.

uzzi38 made very interesting graphs and spreadsheet using his RX 6700XT.

Frequency (Average)	1236 MHz	1912 MHz	2200 MHz	2565 MHz
Increase in frequency	0 %	+55 %	+78 %	+108 %
Voltage	0.701 V	0.861 V	0.964 V	1.12
Increase in Voltage	0 %	+23 %	+38 %	+60 %
GPU only power	30 W	69 W	101 W	175 W
Increase in GPU power draw	0 %	+130 %	+237%	+483%

RDNA2 using TSMC's 7nm node is the most power efficient at ~0.7 V and ~1250 MHz, go over this and the efficiency will drop because voltage has a detrimental effect on power consumption.
For example:
If you locked the frequency of the GPU to 1236MHz, but increased the voltage from 0.701V to 0.861V, just this would increase power consumption by ~51%(0.861^2/0.701^2).
Now your Perf/W ratio dropped by 1/3, because the performance didn't change, but power consumption increased by 1/2.

I am not really surprised Samsung can't achieve reasonable thermals at 1.9GHz compared to 1.29GHz, because this frequency increase means >2X higher power consumption for the GPU alone.

 

[NTMBK]

Samsung announces Exynos 2200 mobile processor with Xclipse GPU based on AMD RDNA2 architecture - VideoCardz.com

Samsung Introduces Game Changing Exynos 2200 Processor With Xclipse GPU Powered By AMD RDNA 2 Architecture The new premium mobile processor comes with hardware-accelerated ray tracing and state of the art Arm-based processing technology Samsung Electronics, a world leader in advanced...

videocardz.com

There we go, they actually announced it now. (Xclipse, that sure is... a name...) Lots of talk about raytracing, no benchmark numbers or clock speeds. Let's see how the final devices perform.

It would be interesting if this chip ends up in Windows for ARM laptops- though I expect that the terms of the deal with AMD prohibits that.

 

[moinmoin]

I'm confused, why now but not earlier?

 

[igor_kavinski]

I'm confused, why now but not earlier?

They were waiting for the right time. They got the latest predictions from the best astrologer in the world and he/she told them, "Today is not the day to bring out a product starting with the letter X".

 

[DisEnchantment]

I am not really surprised Samsung can't achieve reasonable thermals at 1.9GHz compared to 1.29GHz, because this frequency increase means >2X higher power consumption for the GPU alone.

RDNA2 on desktop is on N7, if so called second generation 4nm process cannot cut power in half then something is wrong with the process and Samsung is abysmally behind by at least 2 process generations vs TSMC.

"Samsung’s Xclipse GPU is the first result of multiple planned generations of AMD RDNA graphics in Exynos SoCs,” said David Wang, Senior Vice President of Radeon Technologies Group at AMD
But seems XClipse is the first in the series, so Samsung have confidence in the architecture which is a good sign.
On AMD side, feedback from XClipse and Aerith/VGH could be useful which could bode well for the future of Laptops and PC handhelds.
Or maybe not, they will just clock RDNA3 to hell, which seems to be the case.

AMD is also developing ARM cores in parallel with x86, I wonder if this will go into some semi custom device like a future Exynos.

 

[moinmoin]

RDNA2 on desktop is on N7, if so called second generation 4nm process cannot cut power in half then something is wrong with the process and Samsung is abysmally behind by at least 2 process generations vs TSMC.

There previously were suggestions that with 4nm Samsung is finally closing in on TSMC's N7 efficiency wise, so...

And I'm still not sure how much AMD is actually involved beyond licensing. The GPU name doesn't sound very Radeon-like.

 

[TESKATLIPOKA]

RDNA2 on desktop is on N7, if so called second generation 4nm process cannot cut power in half then something is wrong with the process and Samsung is abysmally behind by at least 2 process generations vs TSMC.

I don't know how good or bad 4nm from Samsung is. I just wrote the difference in power consumption between those two clocks based on N22 and 7nm.

BTW there is a leak saying 109 FPS in GFXBench Aztec Ruins.
Snapdragon 888 manages 116(+6%) and way Xiaomi manages 143(+30%) FPS.
It doesn't look good for Exynos 2200, but let's not forget It's an engineering unit and AMD has FSR. They could be the GPU king, just need to enable It.

 

[DisEnchantment]

I don't know how good or bad 4nm from Samsung is. I just wrote the difference in power consumption between those two clocks based on N22 and 7nm.

BTW there is a leak saying 109 FPS in GFXBench Aztec Ruins.
Snapdragon 888 manages 116(+6%) and way Xiaomi manages 143(+30%) FPS.
It doesn't look good for Exynos 2200, but let's not forget It's an engineering unit and AMD has FSR. They could be the GPU king, just need to enable It.

They need some actual game though, because with gfxbench even desktop GPUs are slower than mobile GPUs.
Same test on 6600XT shows it slower than Dimensity 9000
But probably most mobile games are architected around these mobile GPUs

 

[TESKATLIPOKA]

Samsung Xclipse 920 (AMD RDNA2) GPU is up to 25% faster than Adreno 730 in Vulkan test

Samsung Exynos SoC up to 50% faster in OpenCL and 25% in Vulkan Geekbench test than Snapdragon 8 Gen 1

Although only synthetic benches, It's much better than the leak before.
We will have to wait for reviews to see the end result.

 

[NTMBK]

Samsung Xclipse 920 (AMD RDNA2) GPU is up to 25% faster than Adreno 730 in Vulkan test

Although only synthetic benches, It's much better than the leak before.
We will have to wait for reviews to see the end result.

I hope Anandtech get a review unit. This is one of the most interesting SoCs in a while, even if it turns out to suck.

 

[moinmoin]

I hope Anandtech get a review unit. This is one of the most interesting SoCs in a while, even if it turns out to suck.

I'm afraid Andrei left AT last November. Not sure who is filling his shoes trying to get a good look into mobile SoCs, if anyone.

 

[GodisanAtheist]

I'm afraid Andrei left AT last November. Not sure who is filling his shoes trying to get a good look into mobile SoCs, if anyone.

- Probably nobody at this point. We're watching AT slowly decompose before our very eyes.

Won't be too long before the front page goes the way of Xbit labs and such, a one respected and terrifying dinosaur on its last gasps on its way to extinction in an age of visual medium Youtube spastic garbage.

Also, get off my lawn!

 

[Mopetar]

Maybe Anand will emerge from the secluded labs in Cupertino and return to the site he created to lead it into a new and prosperous golden age.

Honestly though as long as Ian is around we're going to get a lot of high quality articles for whatever he is able to cover. Even if the site doesn't cover everything, the coverage it does provide will still be some of the best in the business.

 

[Saylick]

Chips and Cheese has kind of been the spiritual successor to AT with their highly technical articles. A bunch of tech websites use the words "deep dive" so loosely these days (/rant: just because you can summarize bullet points shown in technical slides given to you by the manufacturer does not make your article a "deep dive" /endrant) that it's refreshing to see someone else actually trying to carry the mantle.

 

[moinmoin]

Do you want to see the Samsung's "official introduction" video for Exynos 2200? Hint: you don't if you are interested in the tech. Aside nice CG graphics this contains pretty much zero info. The GPU is being praised for ray tracing and VRS, a screen can be run at 144hz...