Discussion Qualcomm Snapdragon Thread

[jdubs03]

I don’t think I’ve seen this posted yet, but here are the presentation slides from earlier today: https://www.qualcomm.com/content/da...ents/Snapdragon-Summit_Oct2024_DAY1_CAmon.pdf

The genius is that they're MORE efficient at those clocks!! Nothing wrong w/ hitting 5Ghz as long as you're ultimately more efficient...

I give them props. Second gen Oryon is a big improvement clearly, and that is without version 9.2 like I mentioned. I think third gen Oryon will see some gains just from that upgrade (for example SME). Maybe they can do a bit of both, increase frequency but also needing to bump IPC.

Can’t deny the improvement:

Their next slide is a bit misleading though, particularly with geekbench single threaded and even really multi thread scores. Particularly when you consider the Geekerwan results.

Nonetheless I’ve always been a fan of competition. This will push everyone to improve.

Edit to show Geekerwans peak ST/MT results:

I will note the median A18P MT score is around 8500-8600, so that’s for sure a golden sample (probably freezer too like the 9400 level scores), that ST is right on though.

 

[DZero]

Qualcomm faces a serious uphill battle in laptops. They can claim single thread and battery leadership, but it is all by not much. Any AI leadership will also be by not much.

The gaps the M1 held was often jaw dropping. Oryon is impressive, but is no head turning feature. Both Strix Point and Lunar Lake are very snappy with long battery life. They both have better GPUs and Strix is the multithreaded leader.

If you see someone excited about a Strix laptop, what can you say to redirect them to Qualcomm? Barely perceptible extra snappyness? 15 hours of battery life instead of 12 1/2 hours? The ability to generate an article summary in 8.2 seconds instead of 8.8 seconds?

What would compel a Snapdragon purchase?

Also Strix would have a decent iGPU enough good for gaming.
And don't forget MTK that will come with nVIDIA

 

[Saylick]

I’ve been catching up on the news, but was it clarified what node the X Elite Mobile chip is fabbed on? I presume if it’s a 3nm class chip, it’s N3E?

 

[FlameTail]

I’ve been catching up on the news, but was it clarified what node the X Elite Mobile chip is fabbed on? I presume if it’s a 3nm class chip, it’s N3E?

The slide says it's "3nm", which mostly likely means N3E.

 

[FlameTail]

I don’t think I’ve seen this posted yet, but here are the presentation slides from earlier today: https://www.qualcomm.com/content/da...ents/Snapdragon-Summit_Oct2024_DAY1_CAmon.pdf


I give them props. Second gen Oryon is a big improvement clearly, and that is without version 9.2 like I mentioned. I think third gen Oryon will see some gains just from that upgrade (for example SME). Maybe they can do a bit of both, increase frequency but also needing to bump IPC.

Can’t deny the improvement:
View attachment 109939

If only they would bring this 2nd gen Oryon goodness to laptops...

Otherwise it's a long time away until 3rd gen Oryon lands in laptops (2026H1, if rumours are to be believed).

 

[poke01]

https://twitter.com/x/status/1848486107064734164

This dude mad that Amon called out Intel.

 

[FlameTail]

Die Area of the various components of the 8 Elite​

Component	Area (mm²)
SoC	124
Oryon-L core	2.2
Oryon-M core	0.88
12 MB L2	5.57
GPU	21.9
ISP	9.21
NPU	9.18
DPU	3.87
VPU	5.86
64 bit LPDDR5X	3.66
Modem	12.0

DPU = Display Processing Unit
VPU = Video Processing Unit (Encode/Decode)

Measurements are based on the dieshot published by Kurnal, and his labelling of the sub-components. Also, I cannot vouch that these numbers are 100% accurate, so treat then as estimates.

 

[coercitiv]

Looks much better than the stalled pipe cleaner that WoA received.

I'm reading through the slides/posts and laughing: Microsoft betrayed AMD/Intel only to be handed the trash SoC while QC were working on the good stuff. Damn, it's so fitting for MS.

 

[FlameTail]

Oryon-M is 40% of the size of Oryon-L

 

[FlameTail]

Now we know why there advertising onboard GMEM, despite previous Adreno GPUs also having it.

Because it's a gigantic 12 MB of GMEM in 8 Elite.

 

[FlameTail]

M3 : 3400 points @ 25W
LNL (Arc 140V) : 3300 points @ 35W
STX (Radeon 890M) : 3100 points @ 40W

8 Elite (Adreno 830) : 2600 points @ 11W

Now imagine if Qualcomm took the Adreno 830, doubled it's size, and slapped it into the X Elite Gen 2.

X Elite Gen 2 : 5200 points @ 22W (theoretical)

That will smash Intel/AMD's laptops iGPUs and be a bit better than Apple M5 even.

Qualcomm should go ahead and put a decently big iGPU into X Elite Gen 2, instead of skimping on the die area.

 

[Gideon]

Qualcomm should go ahead and put a decently big iGPU into X Elite Gen 2, instead of skimping on the die area.

While that would be great, I'd much rather wish they just fixed the drivers, even when they upsize it minimally.

I have more faith in Nvidia + Mediatek doing that though

 

[MS_AT]

Not so bad for a marketing presentation. They stated the compiler configuration they have used, only slightly hurting x64 competition, not enough to alter the scores in significant way. They picked consistently Geekbench 6.2 instead of 6.3 so they would look better when compared to Apple. And since 8 Elite is targeting mobile market, they can avoid detailed comparisons against X Elite Plus (was that the name of the middle SKUs for "cheaper" laptops?), which would be interesting. While I guess some people might consider it blasphemy, I would be curious if I could get a proper WoA install on my mobile

Now imagine if Qualcomm took the Adreno 830, doubled it's size, and slapped it into the X Elite Gen 2.

Fortunately we don't have to imagine the driver issues, these are already here with X Elite Intel example shows it's not so easy to catch up.

 

[FlameTail]

While that would be great, I'd much rather wish they just fixed the drivers, even when they upsize it minimally.

Yep, performance in benchmarks doesn't directly translate to games if the drivers are poor. That said though, drivers are software, and software is malleable. Whereas the GPU is hardware, and hardware is fixed. So it would be good to have a large iGPU in the first place.

 

[MS_AT]

Yep, performance in benchmarks doesn't directly translate to games if the drivers are poor. That said though, drivers are software, and software is malleable. Whereas the GPU is hardware, and hardware is fixed. So it would be good to have a large iGPU in the first place.

In theory yes, but if the confidence they will get it working is low, then you are paying for wasted sand

Oh, one thing I have not noticed before, for some of the comparisons they are comparing WoA X Elite vs Android 8 Elite or Android 8 Elite against Lunar Lake / HX370 on Windows.

 

[FlameTail]

Geekerwan's 1st video about X Elite;

Same video as the one posted on Bilibili yesterday, but now on YouTube with English subtitles.

 

[Meteor Late]

Oryon-M is 40% of the size of Oryon-L

Isn't Oryon-M the exact same architecture as Oryon-L? seems like being only 40% in size is a massive feat, when, for example, Zen 5c IIRC is 75% of Zen 5.
What are the main differences that allow it to be so small compared to L version? is it just less L1 cache (but not even half, 128 vs 192) and using dense libraries?

 

[soresu]

Zen 5c IIRC is 75% of Zen 5

Eh?

Regular Z5 Turin tops out at 128C.

Dense Z5 Turin tops out at 192C.

That's definitely greater area savings than 25%.

 

[Meteor Late]

Eh?

Regular Z5 Turin tops out at 128C.

Dense Z5 Turin tops out at 192C.

That's definitely greater area savings than 25%.

Maybe, AMD said 25% smaller core, that's what I was basing it on:

AMD dishes out more Zen 5 details — compact core is 25% smaller than the normal core, new SoC and chip architecture with dual CCXs

Tying it all together.

www.tomshardware.com

AMD revealed that its Zen 5c 'compact' cores are roughly 25% smaller than the standard full-fat Zen 5 cores

 

[soresu]

Maybe, AMD said 25% smaller core, that's what I was basing it on:

AMD dishes out more Zen 5 details — compact core is 25% smaller than the normal core, new SoC and chip architecture with dual CCXs

Tying it all together.

www.tomshardware.com

Hmmmm, might be that the CCX design changes allowed for further area savings on top of the core itself.

 

[FlameTail]

Isn't Oryon-M the exact same architecture as Oryon-L? seems like being only 40% in size is a massive feat, when, for example, Zen 5c IIRC is 75% of Zen 5.
What are the main differences that allow it to be so small compared to L version? is it just less L1 cache (but not even half, 128 vs 192) and using dense libraries?

It's not the same architecture as Oryon-L.

That's what the dieshots are suggesting.

We'll have to wait for SPEC2017 testing of the Oryon-M core to confirm.

 

[GTracing]

Hmmmm, might be that the CCX design changes allowed for further area savings on top of the core itself.

Zen5c Turin is using N3.

 

[mvprod123]

Core-to-Core Latency

 

[FlameTail]

X Elite for comparison;


The Oryon CPU in 8 Elite is much better than the one in X Elite.

• Vastly better ST and MT efficiency
• Lower c2c latencies
• Presence of 'E-cores
• Stable 4.3 GHz clock speed

I think calling it 2nd gen Oryon CPU is highly justified.

 

[Meteor Late]

Apparently another reviewer (S white database) got single core efficiency in Geekbench:

Don't know how trustable it is, but it wouldn't surprise me at all. While Qualcomm leads in MT both in performance and efficiency, there is a discrepancy in core counts and performance of "small" cores, so it will be more efficieny by having more cores and downclocking. Like to like 1 core comparison and Apple core is more efficient and powerful. There is a huge difference in mm2 area in the same node, however, so it's not a bad result at all.