Discussion AMD's Future APU Gone ARM?

[Nothingness]

Evidently you need to learn what the word other means.

I definitely missed the context of the part I was quoting, sorry

But still: Ampere don't belong to Oracle. That said, I agree with what you wrote about Ampere; they don't seem to be doing well, and I'm not sure there's room for them given that Amazon and now Google are using their in-house SoC. I guess they might still find some place at smaller cloud providers that don't have enough money to make their own internal development (but then I'm not convinced by their in-house CPU that's still MIA).

 

[uzzi38]

I definitely missed the context of the part I was quoting, sorry

But still: Ampere don't belong to Oracle. That said, I agree with what you wrote about Ampere; they don't seem to be doing well, and I'm not sure there's room for them given that Amazon and now Google are using their in-house SoC. I guess they might still find some place at smaller cloud providers that don't have enough money to make their own internal development (but then I'm not convinced by their in-house CPU that's still MIA).

I'm convinced they're going to end up going the way of Marvell frankly. I'm not convinced an AmpereOne that launches now can create a compelling enough case against Sierra Forest or Turin Dense to get one of those smaller cloud vendors to even look at them.

 

[FlameTail]

The ARM server is dying? Oh no.

 

[Ghostsonplanets]

On the QCOM Server thing, it need to be remembered that Server isn't a small one-off venture you can enter and leave at whim. Customers will want much more than a "Best in Class" Core. You need serious software, ISV and engineering support available, you need a roadmap, you need to invest into Server specific details (Uncore, Scalability, Cores per Socket, perf/W, etc) and you need a serious and committed roadmap and execution that shows your partners that you won't walk away at anytime.

AMD literally struggled a lot with the latter part, once their Epyc based systems started to make rounds, because people didn't believed they would execute and that Zen was a one-off thing. It took AMD many years for them to create a reliable brand image that eased their way into Server.

QCOM, just like AMD at the time, only has a "Best in Class" core and that's all. If they want to go server, they'll need to execute on a lot of things and it will be a slow and long-term process. They can't enter and leave this space like they did before. So, it remains to be seen if we'll truly see QCOM in server space. Hardware is only part of the solution.

No way, IMO that Exynos will use Mali.

Mali? The recent designs are actually really good but the driver situation is concerning. But I would welcome a big player to try to force their way into WoA and make it so that Arm has to start to give more support towards Mali on WoA.

Seeing possible more entrants in the GPU space for Windows is exciting. Soon we'll have AMD, Nvidia, QCOM and Arm with GPX IPs for Windows. Maybe even Huawei with Maleoon?

 

[Ghostsonplanets]

The ARM server is dying? Oh no.

Less dying and more that it's slowing down. Arm server was born due to Arm being able to exploit the downsides of Server x86 and offer significantly better perf/W and many cores per socket. But, to support Arm, you need to revamp and give serious engineering resources for it to be stand toe to toe with x86 solutions. And you need assurance that roadmaps will be executed.

As AMD and Intel start to plug-in the gap and offer significantly better perf/W and many cores solutions, Arm attractiveness is smaller because the path of least resistance is to stay on the reliable, tried and tested solutions.

But Arm is still attractive. Specially due to the fact you can create a custom, tailored for your needs, solution. But, then again, there are advantages and disadvantages.

 

[NTMBK]

I'm convinced they're going to end up going the way of Marvell frankly. I'm not convinced an AmpereOne that launches now can create a compelling enough case against Sierra Forest or Turin Dense to get one of those smaller cloud vendors to even look at them.

My personal theory is that Nvidia will acquire them. They're even based in Santa Clara.

 

[jpiniero]

My personal theory is that Nvidia will acquire them. They're even based in Santa Clara.

Don't see the point from nVidia's perspective. The stock ARM cores are probally adequate for their needs.

 

[NTMBK]

Don't see the point from nVidia's perspective. The stock ARM cores are probally adequate for their needs.

Extra experienced hardware engineers to work on their server platform, access to Ampere's customers, a custom core to use in both server and consumer products without ARM per-core licensing fees.

 

[Mahboi]

AMD literally struggled a lot with the latter part, once their Epyc based systems started to make rounds, because people didn't believed they would execute and that Zen was a one-off thing. It took AMD many years for them to create a reliable brand image that eased their way into Server.

Hold on a second, AMD got Naples out in 2017. And they weren't amazing. It really took off with Zen 2. How many years are we talking here?

 

[FlameTail]

Extra experienced hardware engineers to work on their server platform, access to Ampere's customers, a custom core to use in both server and consumer products without ARM per-core licensing fees.

That would make Nvidia a formidable force to be reckoned with

 

[Tigerick]

No way, IMO that Exynos will use Mali.

You should at least check the past generation of Exynos as listed below:

EXYNOS	Node	Prime CPU	Total Cores	RAM		GPU	GPU Arch	Clock	FP32
2200	4LPE	Cortex-X2 2.8GHz	1 + 3 + 4 = 8	LPDDR5		Xclipse 920	RDNA2 6CU	1306 MHz	1 TF
2400	4LPP+	Cortex-X4 3.2GHz	1 + 2 + 3 + 4 = 10	LPDDR5X		Xclipse 940	RDNA3 12CU	1109 MHz	3.4 TF
2500	SF3	Cortex-X5 3.2GHz ?	1 + 5 + 4 = 10	LPDDR6 ?		Xclipse 950	RDNA3+ ?

Hmm, what surprise me is the amount of CU in E2400. Could E2400 also gearing for WoA laptop in the future???

 

[Tigerick]

This person:

https://twitter.com/x/status/1775724470256419218

Exynos 2500 A: 8C in a 1+5+2*
Exynos 2500 B: 10C in 2+6+2*

*If I'm not mistaken

They do seem to be more or less reputable with Samsung leaks. I think they were the first to claim the Exynos 2500 core configuration.

Ho, I actually discussed the leaks @ ARM Thread . And Yes based on the table I created above, it seems Samsung is preparing to combine with AMD to cater WoA laptop market. Based on Wiki:

On 3 June 2019, AMD and Samsung announced a multi-year strategic partnership in mobile graphics IP based on AMD Radeon GPU IP.

AMD should provide RDNA IP to Samsung so that Samsung can integrate with Exynos SoC. AMD also should provide drivers support for Android and upcoming WoA. My question is what does AMD get in return beside tiny licensing fee?

 

[Mahboi]

AMD should provide RDNA IP to Samsung so that Samsung can integrate with Exynos SoC. AMD also should provide drivers support for Android and upcoming WoA. My question is what does AMD get in return beside tiny licensing fee?

Ownership of an IP that is so dominant that it is in PC GPUs, every console outside of Nintendo (including all Steam Deck & imitations), all high end phones and fundamentally everywhere.
Seems like a strong thing to have under your foot...

 

[NTMBK]

Ownership of an IP that is so dominant that it is in PC GPUs, every console outside of Nintendo (including all Steam Deck & imitations), all high end phones and fundamentally everywhere.
Seems like a strong thing to have under your foot...

ARM has IP in almost every smartphone, most of the embedded market, etc. Their revenue is 10x smaller than AMD's revenue. Why would AMD want to turn themselves into a poor man's ARM?

 

[Mahboi]

@NTMBK remember when one day AMD announced that "FSR was going to be co-developed with Samsung/Qualcomm/whoever else"? For that.
Getting everyone on your boat is also giving some interesting possibilities.

 

[FlameTail]

Hmm, what surprise me is the amount of CU in E2400. E2400 also gearing for WoA laptop in the future???

The E2400 needs that many CUs to stay competitive with the iGPUs of rival Dimendity 9400 and Snapdragon 8G3, in terms of both performance and efficiency.

 

[uzzi38]

Extra experienced hardware engineers to work on their server platform, access to Ampere's customers, a custom core to use in both server and consumer products without ARM per-core licensing fees.

Nvidia aren't interested in the cloud computing market, the entire reason for Grace is to just basically use it as an IO hub for their GPUs.

If Nvidia wanted a custom core btw, they'd just... Make their own. Like they've done multiple times in the past.

They don't care for the performance of the cores they use all that much, so stock ARM cores are all they need

 

[FlameTail]

If Nvidia wanted a custom core btw, they'd just... Make their own.

And probably fail.

Designing a good CPU core requires significant resources and talent.

 

[uzzi38]

You should at least check the past generation of Exynos as listed below:

EXYNOS	Node	Prime CPU	Total Cores	RAM		GPU	GPU Arch	Clock	FP32
2200	4LPE	Cortex-X2 2.8GHz	1 + 3 + 4 = 8	LPDDR5		Xclipse 920	RDNA2 6CU	1306 MHz	1 TF
2400	4LPP+	Cortex-X4 3.2GHz	1 + 2 + 3 + 4 = 10	LPDDR5X		Xclipse 940	RDNA3 12CU	1109 MHz	3.4 TF
2500	SF3	Cortex-X5 3.2GHz ?	1 + 5 + 4 = 10	LPDDR6 ?		Xclipse 950	RDNA3+ ?

Hmm, what surprise me is the amount of CU in E2400. Could E2400 also gearing for WoA laptop in the future???

The original 3WGP implementation was genuinely just too small for E2200. The original mRDNA was rather competitive in perf/Area, but the missed clock targets combined with the relatively small iGPU meant Samsung fell behind a bit. Even then, at the time 1.3GHz was a surprisingly high clock for an iGPU.

With E2400, they dedicated more area to the iGPU and got a much more competitive iGPU as a result. 6WGP is probably more in line with the competition frankly, I'm not sure it's worth reading into any more than that.

 

[uzzi38]

And probably fail.

Designing a good CPU core requires significant resources and talent.

Ampere isn't the best option if what you want is a high performance core anyway. The custom cores designed for AmpereOne were targeting only a very modest improvement over N1 with the main goal being making the core more area efficient more than anything else.

 

[Nothingness]

Ampere isn't the best option if what you want is a high performance core anyway. The custom cores designed for AmpereOne were targeting only a very modest improvement over N1 with the main goal being making the core more area efficient more than anything else.

With their 16 KB icache, I somehow doubt they can beat N1 for larger workloads. Per socket performance might be better due to a larger number of cores. But we all know single thread (and MT with a small number of threads) performance matters.

 

[Ghostsonplanets]

The original 3WGP implementation was genuinely just too small for E2200. The original mRDNA was rather competitive in perf/Area, but the missed clock targets combined with the relatively small iGPU meant Samsung fell behind a bit. Even then, at the time 1.3GHz was a surprisingly high clock for an iGPU.

With E2400, they dedicated more area to the iGPU and got a much more competitive iGPU as a result. 6WGP is probably more in line with the competition frankly, I'm not sure it's worth reading into any more than that.

I wonder if they'll keep increasing WGP count on future Exynos designs or if they'll stabilize on 6 WGPs. With Exynos 2500 using RDNA 4 and SF3P, there should be some area savings to include 2 WGPs. Wider design at low clocks is a good way to match competition with low power, at expense of area.

 

[igor_kavinski]

Designing a good CPU core requires significant resources and talent.

They have those in spades. But they have become so fat and lazy from the "free" AI hype income pouring in that they can't be bothered to do anything else other than GPUs as AI accelerators. They are the hardware company equivalent of Steam's Valve. Remember Half Life 3? Best guess why the world is still deprived from that?

Spoiler

Fat and lazy from all the "free" income pouring in from the Steam store

 

[FlameTail]

I wonder if they'll keep increasing WGP count on future Exynos designs or if they'll stabilize on 6 WGPs. With Exynos 2500 using RDNA 4 and SF3P, there should be some area savings to include 2 WGPs. Wider design at low clocks is a good way to match competition with low power, at expense of area.

Iirc Exynos 2500 will use SF3, not SF3P.

Snapdragon 8 Gen 4 will bring a ~35% uplift over the Snapdragon 8 Gen 3's GPU performance in 3Dmark Wildlife Extreme.

So Exynos 2500 will also have to bring a similar increase over Exynos 2400, to keep up with the 8G4.

 

[Ghostsonplanets]

Iirc Exynos 2500 will use SF3, not SF3P.

Snapdragon 8 Gen 4 will bring a ~35% uplift over the Snapdragon 8 Gen 3's GPU performance in 3Dmark Wildlife Extreme.

So Exynos 2500 will also have to bring a similar increase over Exynos 2400, to keep up with the 8G4.

I see where I was wrong now. What they now call SF3 is what was known previously as 3GAP. Thought they would keep it when they did the transition to the new naming scheme, but, apparently, they did an Intel 10nm hahaha.