Discussion AMD's Soundwave ARM APU: The Beginning of Transformation !!!

[LightningZ71]

Apple demonstrated a way, actually a few ways, to do both instruction and OS emulation/conversion on the fly with their switch to ARM and the Rosetta 2 implementation. It isn't as costly as it used to be, and, to be frank, CPUs have a lot of performance left on the table in most situations these days.

 

[Tigerick]

RDNA4.5 iGPU = Subset of RDNA5 GPU



After done with XDNA3, let's move on to RDNA4.5 iGPU. Due to N3P node, RDNA 4.5/5 will get biggest upgrade in term of numbers of SP per CU. Each CU will get two times of SP, from 64 to 128. For example, STX with 16CU will have same amount of SP as upcoming MDS1 with 8CU. MDS1 might have same TF if the clock speeds have remained the same.

It is not in the table above but RDNA4 still remain 64 SP per CU. That's why RDNA4.5 is actually considered subset of upcoming RDNA5. By doubling SP per CU, the amount of CU will be cut by half, therefore each Radiance core will need to have more than 3-4 times powerful. I am putting question mark not because I have doubt, it is because I don't know final clock speed of APU. Clock speed will determine the final TF, that's why Soundwave only need 12CU(24CU) to get double performance of E2600. Even with 24CU, SWV still won't be faster than STX/MDS1 also due to clock speed because SWV is designed to be power efficient APU; see my thread in Custom APU.

In conclusion, Soundwave APU will get first mover of latest technology from AMD in 2026, namely RDNA4.5 and XDNA3, period.

 

[DZero]

Wait... isn't Exynos 2600 and 1680 will get that GPU uArch too?

 

[Tigerick]

Wait... isn't Exynos 2600 and 1680 will get that GPU uArch too?

Exynos 2600, yes. My calculation is 5.7TF



Exynos 1680 should feature Xclipse 550 GPU, RDNA3.5+ with 2CU, same family as E2500. Considered this as entry SoC.

 

[DZero]

Exynos 2600, yes. My calculation is 5.7TF

View attachment 133883

Exynos 1680 should feature Xclipse 550 GPU, RDNA3.5+ with 2CU, same family as E2500. Considered this as entry SoC.

Interesting, let's see how develops. I expect
E2600 going 10% less than SD 8 Gen 5 Elite, but better than SD 8 Gen 5
E1680 going on par with olde Exynos 2200. And that could be interesting to watch.

 

[Nothingness]

Cyberpunk running on DGX10 with FEX and Proton: https://wccftech.com/nvidia-dgx-spa...5-fps-cyberpunk-2077-path-tracing-dlss-4-mfg/

Doesn't look that bad.

 

[Tigerick]

https://x.com/events_msft?s=20

About times; Let's see tomorrow is what I expect Microsoft going to announce...

Tomorrow #MSIgnite kicks off the next wave of AI transformation.

Join us live at 9AM PT as Microsoft leaders unveil what's next in AI, cloud, and security.

And AMD is joining as well:

https://twitter.com/x/status/1990072367964365230

 

[Tigerick]

Microsoft Ignite - November 18–21, 2025

Microsoft Ignite - November 18–21, 2025 in San Francisco, CA. Get the edge you need to drive impact in the era of AI. Join us to bolster your knowledge, build connections, and explore emerging technologies.

ignite.microsoft.com

Geez, NV is sponsoring the official Microsoft Ignite backpack meanwhile AMD is sponsoring water bottle. Wonder what are NV and AMD cooking?

 

[DrMrLordX]

Wonder what are NV and AMD cooking?

If you push TDPs high enough, you catch on fire? Maybe? Ignite!

 

[Tigerick]

Wait... isn't Exynos 2600 and 1680 will get that GPU uArch too?

After some diggings, E1680's GPU might be using some variant of E2600's RDNA4.5 GPU cause the older model, E1580 comes with RDNA3 4CU which perform similar with E1680's 2CU. The most interesting question should be why is AMD licensing their latest Radeon IP in entry SoC from Samsung?

	Exynos 1580	Exynos 1680	Exynos 2600		SD 7+ G3	SD 8s G4	SD 8 G5	SD 8 G6		SD 8 Elite G4	SD 8 Elite G5	SD 8 Elite G6
End Product	Tab S10 FE				Xiaomi Pad 7	Xiaomi Pad 8	Xiaomi Pad 9?			Xiaomi Pad 8 Pro	Xiaomi Pad 9 Pro?
Launch Date	Q4 2024	?	?		Q1 2024	Q2 2025	Q4 2025	Q4 2026		Q4 2024	Q3 2025	Q3 2026
Node	4LPP+	SF2 ?	SF2		N4P	N4P	N3P	N2		N3E	N3P	N2
Die Size (mm2)					89.54	105.56	103.55			124.1	126.2
Modem		Integrated 5348	External 5410
CPU	1+3+4 8-core	1+4+3 8-core	1+3+6 10-core		1+4+3 8-core	1+3+2+2 8-core	2+6 8-core			2+6 8-core	2+6 8-core	2+3+3 8-core
Prime CPU	2.9GHz A720	2.9GHz ?	3.8GHz C1 Ultra		2.8GHz Cortex-X4	3.21GHz Cortex-X4	3.8GHz Oryon			4.32GHz Oryon	4.61GHz Oryon
GPU	RDNA3 4CU	RDNA4.5 2CU	RDNA4.5 8CU		A732 3S	A825 2S	A829 2S			A830 3S	A840 3S	4S ?
FP32	1.33 TF	~1.38 TF	~4.1 TF		1.46 TF	2.36 TF	2.5 TF			3.38 TF	3.69 TF
NPU	14.7 TOPS	?	~100 TOPS		~10 TOPS	~16 TOPS				~50 TOPS	~80 TOPS

PS: Sammobile thought E1680 is a downgrade from E1580. Well, if you read my table, you should think otherwise.

 

[DZero]

After some diggings, E1680's GPU might be using some variant of E2600's RDNA4.5 GPU cause the older model, E1580 comes with RDNA3 4CU which perform similar with E1680's 2CU. The most interesting question should be why is AMD licensing their latest Radeon IP in entry SoC from Samsung?

	Exynos 1580	Exynos 1680	Exynos 2600
Launch Date	Q4 2024	?	?
Node	4LPP+	SF2 ?	SF2
CPU	1 + 3 + 4 8-core	1 + 4 + 3 8-core	1 + 3 + 6 10-core
Prime CPU	2.9GHz A720	2GHz ?	3.8GHz C1 Ultra
GPU	RDNA3 4CU	RDNA4.5 2CU	RDNA4.5 8CU
NPU	14.7 TOPS	?	~100 TOPS

PS: Sammobile thought E1680 is a downgrade from E1580. Well, if you read my table, you should think otherwise.

1680 uses 2 nm??? I tought that uses 4nm or even 3nm

Meanwhile, I don't be surprised if we see an Exynos 1630, to replace the 1330 with the RDNA3 with 2CU.

 

[Doug S]

1680 uses 2 nm??? I tought that uses 4nm or even 3nm

Meanwhile, I don't be surprised if we see an Exynos 1630, to replace the 1330 with the RDNA3 with 2CU.

Does Samsung have a working process between 4LPP+ and SF2? AFAIK they never got their 3nm node to work at all.

 

[DZero]

Does Samsung have a working process between 4LPP+ and SF2? AFAIK they never got their 3nm node to work at all.

And the Exynos 2500?

 

[regen1]

Does Samsung have a working process between 4LPP+ and SF2? AFAIK they never got their 3nm node to work at all.

Samsung's renaming/rebranding of some nodes has made it blurry.
They rebranded their upcoming SF3P to SF2.

Samsung's new roadmap unveils its 2nm process nodes and outlines backside power delivery plans

Samsung renames SF3P to SF2 due to better performance, power, and density.

www.tomshardware.com

Samsung Foundry Unveils Updated Roadmap: BSPDN and 2nm Evolution Through 2027

web.archive.org

As of now they had somewhat limited success on their 3nm node family with a cryptomining ASIC, W1000, Exynos 2500. Upcoming Exynos 2600 is supposedly on SF2(which is rebranded SF3P).

 

[Doug S]

And the Exynos 2500?
View attachment 134240

I remember reading somewhere that Exynos 2500 had been canceled and Samsung went all-Qualcomm for the S25 but I guess only the second part was true since it seems like they did find a few things to put the 2500 in.

Based on @regen1's post it sounds like SF2 is part of the same node family as SF3, so choosing it hoping to save money would be like choosing TSMC N5 to save money versus N4P - whatever higher per wafer cost you pay is canceled out by the improved density. Even more so if they've got SF2's yielding higher than SF3.

 

[DZero]

I remember reading somewhere that Exynos 2500 had been canceled and Samsung went all-Qualcomm for the S25 but I guess only the second part was true since it seems like they did find a few things to put the 2500 in.

Based on @regen1's post it sounds like SF2 is part of the same node family as SF3, so choosing it hoping to save money would be like choosing TSMC N5 to save money versus N4P - whatever higher per wafer cost you pay is canceled out by the improved density. Even more so if they've got SF2's yielding higher than SF3.

Indeed, but it might be weird, I mean, I suspect that SF2 is around 50% yield, while SF3 might reached 60%. Let's see how ends that. I don't be surprised if even Samsung launches the Exynos 1630 or 1730 at 3nm with them.

 

[Tigerick]

I have created two OS threads: WoA11 and Android PC and pinned them in the frontpage in case anyone wants to know more about upcoming OS.

TLDR: Both Android OS and WoA will be starting new race with ARM PC platforms. All OEMs are going to use whatever APU being made by Qualcomm, NV, AMD, Mediatek and Samsung?. What we are seeing right now are just the first wave of ARM APUs. Do you all still think Soundwave is just one-off project from AMD?

Each platform has their own Pros and Cons. Who will win the war, nobody knows? But I do know who the loser would be: Intel. As Gemini said about Aluminium OS:

Google's new OS will treat ARM as the future and x86 as the legacy platform.

 

[Tigerick]

View attachment 133103
Hoho, what is RDNA4-derived micro-architecture ???

Doesn't it refer to RDNA4+ or RDNA4.5 ???

Per agreement in 2019: Why does AMD license their latest RDNA IP to Samsung with Android support? Think guys....

I have calculated FP32 with above info. There are few ways to hit 6TF with 1.4GHz. One of them is through doubling of SP per CU. Let's see is it the RDNA4.5 improvement which we will see in Soundwave and E2600. As I said E2600 5G is little brother of Soundwave with half memory bandwidth. Do you still think Soundwave which is fabbed by N3P and 128-bit memory bus comes with tiny CPU and GPU cores? I have decoded the XDNA3 AIE, now it is RDNA4.5 turn...

PS: E2600 with 8CU's SNL scores hit 3100+, higher than A19 Pro. Go figures.

Xclipse 960 which is RDNA4-derived micro-architecture is being codenamed as AMD Juno with clockspeed reduced to 985MHz. Samsung has boosted the clockspeed of C1 Ultra to 3.8GHz to show better ST performance; still below D9500's 4.21GHz: 10-core vs 8-core. The reduction of GPU's clockspeed has decreased the FP32 from 5.7TF to 4TF. Still higher than 8 Elite Gen5's 3.7TF though.

Hoho, what is Juno's graphics IP fabbed by SF2? I bet the codename is not in the official roadmap of AMD. Let's see whether Juno is having same IP as upcoming Soundwave APU which is fabbed by N3P.

Leaker Believes Samsung Exynos 2600 Mobile Chip Will Feature AMD "JUNO" iGPU

Earlier in the month, Samsung's marketing team initiated their Exynos 2600 promotional campaign. At the time, official teaser material did not provide any substantial hints about the next-gen mobile chipset's technical underpinnings. Weeks later, a fresh leak has laid out the cutting-edge mobile...

www.techpowerup.com

 

[Gideon]

Hoho, what is Juno's graphics IP fabbed by SF2? I bet the codename is not in the official roadmap of AMD. Let's see whether Juno is having same IP as upcoming Soundwave APU which is fabbed by N3P.

Leaker Believes Samsung Exynos 2600 Mobile Chip Will Feature AMD "JUNO" iGPU

Earlier in the month, Samsung's marketing team initiated their Exynos 2600 promotional campaign. At the time, official teaser material did not provide any substantial hints about the next-gen mobile chipset's technical underpinnings. Weeks later, a fresh leak has laid out the cutting-edge mobile...

www.techpowerup.com

Kinda ironic if AMDs own Medusa APUs end up having older version of RDNA than Samsung phones (which seems to be the case if the rumor is true)

 

[poke01]

Exynos 2600 | Mobile Processor | Samsung Semiconductor Global

Exynos 2600 is a next-generation 2nm GAA mobile SoC, combining a deca-core Arm® C1-Ultra CPU, a powerful NPU, and the Samsung Xclipse 960 GPU for standout performance. Learn more.

semiconductor.samsung.com

Yeah it’s most likely rdna4

 

[Tigerick]

Exynos 2600/2700 vs Soundwave APU: CPU Analysis

	Dimensity 9500 5G	Exynos 2600 5G		Dimensity 9600 5G	Exynos 2700 5G		Soundwave	M5	X2 Elite X2E-80-100
Integrated 5G Modem	Yes	External, 5410		Yes	?		?	No	Yes
Node	N3P	SF2		N2	SF2P		N3P	N3P	N3P
Memory	64-bit LPDDR5x-10667	64-bit LPDDR5x-10667		96-bit LPDDR6-10667	96-bit LPDDR6-10667		128-bit LPDDR5x-9600	128-bit LPDDR5x-9600	128-bit LPDDR5x-9523
Memory BW	85.33 GB/s	85.33 GB/s		114.1 GB/s	114.1 GB/s		153.6 GB/s	153.6 GB/s	152 GB/s
Prime Cores	1 x C1-Ultra 4.21GHz	1 x C1-Ultra 3.8GHz		C2 Ultra	C2 Ultra			4 x 4.6GHz	6 x Oryon cores 4GHz Single-Core 4.7GHz Dual-Core 4.4GHz
Performance Cores	3 x C1 Premium 3.5GHz							6 x 3.04GHz	6 x 3.4GHz
Pro1 Cores		3 x C1 Pro 3.25GHz
Pro2 Cores	4 x C1 Pro 2.7GHz	6 x C1 Pro 2.75GHz

 

[Tigerick]

Exynos 2600/2700 vs Soundwave APU: GPU Analysis

	Exynos 2500 5G	Exynos 2600 5G	Exynos 2700 5G		Soundwave	Medusa Point MDS1	Gorgon Point
Node	SF3	SF2	SF2P		N3P	N3P	N4P
Memory	64-bit LPDDR5x-9600	64-bit LPDDR5x-10667	96-bit LPDDR6-10667		128-bit LPDDR5x-9600	128-bit LPDDR5x-9600?	128-bit LPDDR5x-8533
Memory BW	76.8 GB/s	85.33 GB/s	114.4 GB/s		153.6 GB/s	153.6 GB/s	136 GB/s
GPU	RDNA3+	Juno RDNA 4.5 ?	Juno RDNA 4.5 ?		RDNA 4.5 ?	RDNA 4.5 ?	RDNA 3.5
WGP	8	4	4		6	4	8
CU	16	8	8		12	8	16
SP Per CU	64	128	128		128	128	64
Total SP	1024	1024	1024		1536	1024	1024
GPU Frequency	999 MHz	999 MHz ?	?		?	3100 MHz ?	3100 MHz
FP32 (Dual Issue)	4.1 TF	4.1 TF	~ 5 TF ?			12.7 TF	12.7 TF

 

[Tigerick]

Thanks to a new architecture, the computing performance of the Exynos Xclipse 960 GPU is twice as high as that of its predecessor.⁴ This drives a ray tracing performance improvement of up to 50%.⁵

4,5: Based on internal testing in comparison to the Exynos 2500. Results may vary depending on operating conditions and other factors.

Hoho, let's decode the above statement regarding Exynos 2600's iGPU: Xclipse 960 aka AMD's Juno RDNA IP. It is based on new architecture; definitely refer to new generation of RDNA. Before I proceed, let's assume the story between Samsung and AMD: Samsung wants to showoff more selling features of X960 but AMD does not want to reveal too many details about new RDNA IP because it is the same IP would be used for all future APUs (N3P). AMD definitely have a final word on the statement; thus, the claims should be correct.

Check my table regarding the relationship between STX (N4P) and E2500 (SF3): the performance delta is caused by the clock speed; the features should be similar...Exynos 2500 is using RDNA3+/3.5 IP; thus it is safe to assume Xclipse 960 is using RDNA4+/4.5 IP. However RDNA4s (N4P) are designed for desktop dGPU; some leakers keep saying RDNA4 are not for mobile SoC, then what IP Samsung/AMD is using? Let's check the features below:

• The computing performance of the Exynos Xclipse 960 GPU is twice as high as that of its predecessor. This is bold claim: Samsung claims computing performance is double than Xclipse 950. Hoho, even RDNA4's improvement is only 20% faster in rasterization's IPC. Then where is 100% performance claim from? Hoho, if you read my speculation about doubling SP per CU aka Compute Unit; then with same clock speed, I could claim double performance per CU as well. It is not full picture cause the number of CUs has been cut by half, that's why FP32 still remains the same.
• If above feature is considered overselling, then the second statement is underselling: This drives a ray tracing performance improvement of up to 50%. With 8CU, X960 still manage to offer 50% better RT performance than X950's 16CU/RT. Hoho, anyone wants to estimate how powerful is each Radiance core in X960?

As I speculated before, Soundwave (N3P) is like bigger brother of E2600. E2600 is having 8CU with 4.1TF, then how many CUs would Soundwave having? Don't believe what AMD marketing people told you; I rather believe in MLID's story regarding Senior Management about SWV. And if anyone still believes in Medusa Point is having RDNA3.5+ GPU, hoho...

 

[marees]

And if anyone still believes in Medusa Point is having RDNA3.5+ GPU, hoho...

AMD has a history of bundling old GPU tech in their APus. Witness vega gpu tech in AMD APus — long after AMD stopped supporting vega

Samsung could have pushed AMD for premium ip as their hardware could go into flagship devices but soundwave seems the complete opposite of flagship to me. Same is the case for medusa point APUs

 

[Tigerick]

Exynos 2600/2700 vs SWV/SWV2? APU: NPU Analysis





Samsung Exynos 2600 has presented new generation of NPU with 32K (32,768) MAC units. It is the largest units atm. That's why Samsung is claiming "improves generative AI performance by 113% compared to its predecessor, E2500." That's mean E2600 has about 126 TOPS, biggest by far. By calculation, Samsung only needs to set the NPU speed @ 1.92GHz. That's the difference between phone SoC and mobile SoC. Due to low power, phone SoC will need to set the clock speed as low as possible. However, compared to A19Pro at 8192@2.17GHz, Samsung goes a bit extreme by employing 32,768@1.92GHz at the expense of die area. That's why E2600's NPU is about 3.5 times faster than A19 Pro.

@Fjodor2001 If you interested on AMD XDNA AIE, here is the updated version. Based on roadmap from FAD, XDNA4 should offer better scale and power efficiency. I am updating XDNA4 to employ 32,768 MAC units, that's mean next APUs (N2P?) with XDNA4 should offer 160 TOPS; double than SWV's 80 TOPS. FYI, MDS1 and Medusa Premium are both employing XDNA3 which could offer NPU TOPS in range of 100-110 depending on the clock speed. The formula is presented in the Excel sheet above if you are interested on the calculation of TOPS.