Discussion Qualcomm Snapdragon Thread

[Raqia]

Preview of the Die Shots

A18 Pro: 110 mm^2
8 Elite: 125-130 mm^2
9400: 125-130 mm^2

GPU Cluster in Green. CPU Cluster in Red. P-Core is the Solid Red Block.

Comments say:

A18 Pro P-Core: 2.9 mm^2 (2MB of L2)
8 Elite P-Core: 2.1 mm^2 (No L2)
9400 P-Core: 3.2 mm^2

A18 Pro CPU Cluster: 21 mm^2 (18 mm^2 without SME)
8 Elite CPU Cluster: 22 mm^2
9400 CPU Cluster: 22 mm^2

A18 Pro GPU Cluster: 21 mm^2
8 Elite GPU Cluster: 23 mm^2
9400 GPU Cluster: 22 mm^2

View attachment 109108

The Mali GPU in the 9400 wasn't much changed in cluster count and has the same number of compute units as in the 9300 with more cache. The Adreno in the 8Gen3 (~137 mm^2) seems to occupy relatively less of the SoC die than in the 9300 (~150 mm^2). Lower left in each die (not to scale):
This means in terms of die area, the Adreno's beefed much more relatively speaking. The 830 promises to be a major architectural overhaul, maybe due to the added complexity and cache requirements of supporting more FP32 oriented GPU compute work loads...

 

[FlameTail]

Preview of the Die Shots

A18 Pro: 110 mm^2
8 Elite: 125-130 mm^2
9400: 125-130 mm^2

Interestingly, all those chips are on N3E.

A17 Pro : 103 mm² N3B
8 Gen 3 : 137 mm² N4P
D9300 : 141 mm² N4P

So we looking at about a 10% reduction in die size compared to previous gen of Snapdragon/Dimensity.

Not bad for 4nm -> 3nm shrink.

A18 Pro P-Core: 2.9 mm^2
8 Elite P-Core: 2.1 mm^2 (No L2)
9400 P-Core: 3.2 mm^2 (2MB of L2)

So 8 Elite's Oryon Phoenix-L core is 18% smaller than the Phoenix core in X Elite (2.55 mm²).

A18 Pro CPU Cluster: 21 mm^2 (18 mm^2 without SME)
8 Elite CPU Cluster: 22 mm^2
9400 CPU Cluster: 22 mm^2

22 mm² for 8 Elite's CPU cluster. Hmm. What could we infer from that?

X Elite's CPU cluster of 4 Oryon cores (with L2) is 16 mm²

A18 Pro GPU Cluster: 21 mm^2
8 Elite GPU Cluster: 23 mm^2
9400 GPU Cluster: 22 mm^2

Qualcomm is spending the most die area on GPU. Will that translate to the best performance?

 

[Raqia]

22 mm² for 8 Elite's CPU cluster. Hmm. What could we infer from that?

That is consistent w/ a 2x + 6x CPU config: added area over Apple A18 2x + 4x CPU w/o SME (mobile Oryon likely doesn't include it) could be accounted for by the extra 2 'E' cores (but could also be added cache).

 

[FlameTail]

Looking at middle die, which is 8 Elite;
I think I can see 3 slices in the GPU.
That would be 3 slices of four compute units each
For a total of 12 compute units (Geekbench OpenCL benchmark revealed that Adreno 830 has 12 compute units).

Does X Elite's Oryon core use HP library?

My guess as to what the 8 Elite's CPU cluster looks like;

2 × Phoenix-L
12 MB L2 cache

6 × Phoenix-M
6 MB L2 cache

 

[adroc_thurston]

Does X Elite's Oryon core use HP library?

God no, it'll be a lot bigger if it did.

 

[FlameTail]

>22 mm² for 8 Elite's CPU cluster. Hmm. What could we infer from that?

That is consistent w/ a 2x + 6x CPU config: added area over Apple A18 2x + 4x CPU w/o SME (mobile Oryon likely doesn't include it) could be accounted for by the extra 2 'E' cores (but could also be added cache).

	CPU cluster	P-core
A18	18 mm²	2.9 mm²
8 Elite	22 mm²	2.1 mm²

This suggests that the E-cores in 8 Elite are much larger than those in A18.

In fact, the E-core is what I am most curious about. The secret of Phoenix-M.

 

[Doug S]

Does X Elite's Oryon core use HP library?

With N3E you don't have to choose HP or LP for the entire die. They can use different transistor types in different blocks. I'm not sure of the granularity of FinFlex, the ideal would be to use lower power/denser cells everywhere and sprinkle in higher power ones where necessary to close timing loops. Theoretically you could achieve nearly the frequency that would be possible with an all HP design, with power nearly as low as an all LP design.

I got the impression it is not so granular yet (or maybe ever?) though whether the limitation is in the process or the tools I don't know.

 

[Hitman928]

With N3E you don't have to choose HP or LP for the entire die. They can use different transistor types in different blocks. I'm not sure of the granularity of FinFlex, the ideal would be to use lower power/denser cells everywhere and sprinkle in higher power ones where necessary to close timing loops. Theoretically you could achieve nearly the frequency that would be possible with an all HP design, with power nearly as low as an all LP design.

I got the impression it is not so granular yet (or maybe ever?) though whether the limitation is in the process or the tools I don't know.

You can mix 2 fin and 1 fin or 3 fin and 2 fin libraries within a block, but you are limited to one or the other per row.

 

[Raqia]

Also, consider that the Dimensity 9400 CPU cluster draws ~18W in multi-threaded SPEC:

The knee of the curve is around 10W where you get 90% of the performance at peak TDP. It's unclear what performance the 8 Elite delivers at 20W, so this doesn't rule out that it is more efficient than the 9400...

 

[Doug S]

You can mix 2 fin and 1 fin or 3 fin and 2 fin libraries within a block, but you are limited to one or the other per row.

What about 1 fin and 3 fin? Can you mix those?

 

[Hitman928]

What about 1 fin and 3 fin? Can you mix those?

That isn't an option that I've seen, at least not for N3.

 

[FlameTail]

With N3E you don't have to choose HP or LP for the entire die. They can use different transistor types in different blocks. I'm not sure of the granularity of FinFlex, the ideal would be to use lower power/denser cells everywhere and sprinkle in higher power ones where necessary to close timing loops. Theoretically you could achieve nearly the frequency that would be possible with an all HP design, with power nearly as low as an all LP design.

I got the impression it is not so granular yet (or maybe ever?) though whether the limitation is in the process or the tools I don't know.

I think this video by High Yield explains FinFlex fairly well;

 

[FlameTail]

TechInsights Forecast for ARM Notebook shipments by processor architecture;

They predict rapid marketshare gain for ARM notebooks. No doubt, almost all of this growth will be fuelled by Windows-on-ARM, and Qualcomm will be a key player making gains... if they play their cards right.

 

[DavidC1]

TechInsights Forecast for ARM Notebook shipments by processor architecture;
They predict rapid marketshare gain for ARM notebooks. No doubt, almost all of this growth will be fuelled by Windows-on-ARM, and Qualcomm will be a key player making gains... if they play their cards right.

I highly doubt it, unless Intel/AMD falls flat on it's face again, and Pantherlake-U loses 35% of battery life over Lunarlake, meaning back to Meteorlake levels.

Forecasters are notorious for assuming whatever trend is in motion will continue especially when it fits their bias.

I know Microsoft was pushing WoA sometimes to the detriment of the market recently. Intel and Microsoft have a very dysfunctional love/hate relationship where they blame each other for their problems not realizing they hate each other because they see their own flaws as if looking into a mirror.

 

[okoroezenwa]

Forecasters are notorious for assuming whatever trend is in motion will continue especially when it fits their bias.

Pointedly remembering the silly forecasts of 15% Windows Phone market share.

 

[FlameTail]

Rumour: Samsung has changed it's plans to put the Dimensity 9400 chip in the base model of the Galaxy S25. Instead, all three main models (S25/S25+/S25U) will leverage the Snapdragon 8 Elite.

[Exclusive] The negotiations between Samsung and MediaTek, which initially aimed to include the Dimensity chip in the Galaxy S25, have shifted to placing the Dimensity chip in the S25 FE instead. The S25 will exclusively use Snapdragon chips.

Source

 

[FlameTail]

Snapdragon 8 Elite visits Geekbench 5

QUALCOMM Sun for arm64 - Geekbench

Benchmark results for a QUALCOMM Sun for arm64 with a Qualcomm ARMv8 processor.

browser.geekbench.com

2236 Single | 9015 Multi

For comparison;

~10% behind A18 Pro in Single Core, but ~30% better in Multi Core.

 

[jdubs03]

Snapdragon 8 Elite visits Geekbench 5

QUALCOMM Sun for arm64 - Geekbench

Benchmark results for a QUALCOMM Sun for arm64 with a Qualcomm ARMv8 processor.

browser.geekbench.com

2236 Single | 9015 Multi

For comparison;
View attachment 109200
~10% behind A18 Pro in Single Core, but ~30% better in Multi Core.

Seems about right. Those 2 extra “e” cores definitely help. Overall 10% behind, but IPC ~16% behind (same deficit as in GB6.3).
For the A18 Pro, at least on iOS 18.0.1, the MT score seems to hover between 6400-6600, with some stragglers on each side. Mid 2400s on the ST is most of what I’m seeing. A18 is of similar reductions as in the Geekerwan graph. The GB version for these is 5.5.1. Not sure how that differs from 5.4.4; I’d assume not much.

 

[poke01]

Seems about right. Those 2 extra “e” cores definitely help. Overall 10% behind, but IPC ~16% behind (same deficit as in GB6.3).
For the A18 Pro, at least on iOS 18.0.1, the MT score seems to hover between 6400-6600, with some stragglers on each side. Mid 2400s on the ST is most of what I’m seeing. A18 is of similar reductions as in the Geekerwan graph. The GB version for these is 5.5.1. Not sure how that differs from 5.4.4; I’d assume not much.

There is a difference between 5.5.1 and 5.4.4 for MT

5.4.4

iPhone17,2 - Geekbench

Benchmark results for an iPhone17,2 with an ARM processor.

browser.geekbench.com

iPhone17,2 - Geekbench

Benchmark results for an iPhone17,2 with an ARM processor.

browser.geekbench.com

5.5.1

iPhone17,2 - Geekbench

Benchmark results for an iPhone17,2 with an ARM processor.

browser.geekbench.com

iPhone17,2 - Geekbench

Benchmark results for an iPhone17,2 with an ARM processor.

browser.geekbench.com

 

[Raqia]

Snapdragon 8 Elite visits Geekbench 5

QUALCOMM Sun for arm64 - Geekbench

Benchmark results for a QUALCOMM Sun for arm64 with a Qualcomm ARMv8 processor.

browser.geekbench.com

2236 Single | 9015 Multi

For comparison;
View attachment 109200
~10% behind A18 Pro in Single Core, but ~30% better in Multi Core.

Looks on target for performance, the 8-Elite is more likely to hit projected power numbers than the X-Elite...

Performance Delivered a New Way Part 2âââGeekbench versus SPEC

Author: Ram Srinivasan, Performance Architect, NUVIA

medium.com

It's unclear how the X-Elite would fare if they are using the methodology they describe here, but it's likely the 8-Elite (big core) would do better...

 

[DZero]

Rumour: Samsung has changed it's plans to put the Dimensity 9400 chip in the base model of the Galaxy S25. Instead, all three main models (S25/S25+/S25U) will leverage the Snapdragon 8 Elite.

Source

Don't be surprised if the price hike is aroung USD 200. Samsung might face troubles on selling?
Meanwhile, Dimensity on FE? Exynos is deader than dead then.

 

[Raqia]

I highly doubt it, unless Intel/AMD falls flat on it's face again, and Pantherlake-U loses 35% of battery life over Lunarlake, meaning back to Meteorlake levels.

Forecasters are notorious for assuming whatever trend is in motion will continue especially when it fits their bias.

I know Microsoft was pushing WoA sometimes to the detriment of the market recently. Intel and Microsoft have a very dysfunctional love/hate relationship where they blame each other for their problems not realizing they hate each other because they see their own flaws as if looking into a mirror.

Don't forget that Intel is using TSMC N3E for Lunar Lake, which cuts into their margins substantially; 18A yields are rumored to be abysmal. FOVEROS packaging, while the future of the industry in general, is also much more expensive than monolithic SoCs and Intel is forced to do this now instead of later. Even if Intel holds on, they won't be able to reinvest as much into their own R&D.

Windows on ARM's key component are now in place with native builds of popular browsers, productivity and media editing suites. Once the remaining VPN companies port their code, I see no problem with deploying ARM devices in many large corporate environments which deploy pre-defined, pre-tested suites of software. If you need x64 performance, just RDP into such an instance...

 

[mvprod123]

Don't forget that Intel is using TSMC N3E for Lunar Lake, which cuts into their margins substantially; 18A yields are rumored to be abysmal. FOVEROS packaging, while the future of the industry in general, is also much more expensive than monolithic SoCs and Intel is forced to do this now instead of later. Even if Intel holds on, they won't be able to reinvest as much into their own R&D.

Windows on ARM's key component are now in place with native builds of popular browsers, productivity and media editing suites. Once the remaining VPN companies port their code, I see no problem with deploying ARM devices in many large corporate environments which deploy pre-defined, pre-tested suites of software. If you need x64 performance, just RDP into such an instance...

*N3B

 

[Doug S]

Pointedly remembering the silly forecasts of 15% Windows Phone market share.

And there wasn't even a "trend" there, they just assumed "its Microsoft so they'll win".

 

[hemedans]

And there wasn't even a "trend" there, they just assumed "its Microsoft so they'll win".

There was a trend,

from 2012 Q4 (announcement of wp8) to 2013 Q3 (Sale of Nokia device to Microsoft) Wp sold lot of device and managed to capture 4% marketshare in just one year,

Things falls apart when Microsoft took Nokia division and discontinue everything, most people who Bought Wp because of Nokia move to other platforms.