Discussion Qualcomm Snapdragon Thread

[soresu]

Strix Halo wasn't announced at Computex because it is likely going to be announced at CES 2025.

There certainly won't be any Zen6 SKU announced until Q4 '25 at minimum.

 

[soresu]

That's why Microsoft is claiming that their new Prism emulation layer is as good as Rosetta2. I have heard that the emulation penalty is only like -15% performance.

As with new CPU and GPU µArch benches you will find that Prism's supposed -15% is just an average of a dataset (or worse, the best result by a significant margin).

Given there is currently still a lot of software to be ported natively to WoA they have to be relying on a rather limited data set to work with.

 

[Hitman928]

ARM?

They achieved 17% IPC increase in one year (Cortex X4 -> Cortex X925). I expect Oryon V2 will also bring substantial IPC improvements.

I think we should wait on 3rd party reviews before making any conclusive IPC claims. For ARM, even according to their own numbers, the IPC increase with X925 is < 10% in SPECint, which is what we typically use to measure IPC.

 

[FlameTail]

So from what we know far, this is how X Elite's IP relates to mobile Snapdragon;

GPU based on 8G2's.

NPU based on 8G3's.

CPU based on 8G4's.

Pretty funny

 

[soresu]

It could be the most power sipping high performance CPU at the moment

Once the X core has the op fusion in A725 it should definitely be significantly more efficient, though as yet the jury is out on on X925's efficiency until we get first silicon.

 

[soresu]

So from what we know far, this is how X Elite's IP relates to mobile Snapdragon;

GPU based on 8G2's.

NPU based on 8G3's.

CPU based on 8G4's.

Pretty funny

Probably just because it is the first gen product with a less synchronised SoC design team vs the SD8.

Later gens I would expect to align IP across different SoC products.

 

[StinkyPinky]

Looks like there's a boost frequency bug with early production Samsung Snapdragon laptops. I'm sure we're all shocked to learn that it's a total dog when locked at 2.5 ghz!

Snapdragon X Elite in the wild is allegedly slower than iPhone 12 — first benchmarks of Samsung Book4 Edge disappoint

Wait for firmware updates before condemnation

www.tomshardware.com

This seems pointless since we know nothing about this except that it was locked at 2.5Ghz. Obviously it would be slow at that speed.

Probably a bug or perhaps he set the laptop to some sort of energy saver mode. I wouldn't read much into it.

 

[Hitman928]

This seems pointless since we know nothing about this except that it was locked at 2.5Ghz. Obviously it would be slow at that speed.

Probably a bug or perhaps he set the laptop to some sort of energy saver mode. I wouldn't read much into it.

Seems to be a bug with that laptop. There are a lot of entries in the GB database with scores similar to his, most of them actually.

 

[podspi]

I am pretty bummed about recall being delayed, tbh. Why not just have it off, or even hidden by default?

Oh, well. At least it seems that Phi Silica will be ready on day one. Depending on how good that model actually is, it could be a real game changer.

Now that I think about it, AI will probably be a point of differentiation among operating systems in the future. Not just features, but how well the built-in OS APIs are able to do things like text completion and summarization.

 

[poke01]

We got some early benchmaks at 65 watts and 28 watts. Its seems ST decreases a lot when its set to 28 watts. This is lower than base M2 in CB 2024!

 

[The Hardcard]

View attachment 101208
We got some early benchmaks at 65 watts and 28 watts. Its seems ST decreases a lot when its set to 28 watts. This is lower than base M2 in CB 2024!

Another sign of firmware/BIOS issues? Another perplexing single thread hit.

 

[FlameTail]

First look at Galaxy Book4 Edge

The long vents on the sides are not for the fans, but for the speakers!

 

[FlameTail]

I hope Oryon V2 also brings a substantial power reduction;

View attachment 100592
If you take a measuring ruler to this graph of Oryon V1, you can see that the power consumption increased by roughly 50% (9W->14W) for that last 10% of frequency (3.8 GHz -> 4.2 GHz).

So the Oryon architects have an easy opportunity to cut the power consumption by 1/3rd, if they reduce the frequency by 10%.

4.2 GHz = 100% power
3.8 GHz = 66.66% power

Now, if we say Oryon V1 was ported to N3P (the node Oryon V2 will presumably use), then we get a 20% power reduction.

View attachment 100594
View attachment 100595
4.2 GHz - Oryon V1 - N4P = 100% power
3.8 GHz - Oryon V1 - N4P = 66% power
3.8 GHz - Oryon V1 - N3P = 53% power

So by reducing the frequency by 10%, and going from N4P -> N3P, the power consumption is almost halved!

Now, this gives them a solid foundation to significantly increase IPC in Oryon V2, without blowing up the power consumption.

@SpudLobby The good thing about X Elite being on standard N4 (not N4P), is that it's going to set up the X2 Elite for a huge node jump (N4->N3P).

 

[FlameTail]

Here are the tables, for reference:

 

[SpudLobby]

@SpudLobby The good thing about X Elite being on standard N4 (not N4P), is that it's going to set up the X2 Elite for a huge node jump (N4->N3P).

Yep, there should be a substantial power savings or rather also a substantial clock gain without hurting power much just from electric characteristics of the transistors alone.

And in practice what with how long it took, I suspect they have some phydes changes to iterate on, and things like more L2 cache and extra SLC might be on the menu — at the minimum the extra L2 will be.

This is also why gains in practice from one product to the next on a new node can be so much larger than simply stated by the performance/power of the process — if designers throw on extra L2, L3, SLC cache and make serious commitments to power & IPC upgrades, you can get much more than just + 15% and - 30% performance (as measured by frequency, because you’d have IPC too) iso power and power reductions iso-frequency.

 

[SpudLobby]

One interesting thing is: Qualcomm’s floor frequency for their yields is about 3.4GHz On N4P, with 4GHz also existing for some products at the top end and a more rare 4.2GHz.

Apple’s floor on N4 is 3.42GHz functionally with the A16, and 3.48GHz with M2 on N5P.

Not far off all things considered, Apple is probably ahead still, but not by much on the same nodes.

N3B seems to allow Apple to hit 4GHz on laptop SKUs and 3.8GHz on phone SKUs, with a slight hit to power from where they were on the laptops and definitely a bit more with phones.

But importantly they could actually hit the clocks (or near them) on mass yields, and Apple doesn’t frequency bin.

Then with N3E and the M4, we see Apple can pull off another 10%, and gets to 4.4GHz. Now again, this pushes power up, so we see they’re taking more than the process gives them and whatever cache/phydes changes they made also weren’t enough to help here — platform power still went up. Core power probably doubled and platform power might’ve gone up 2-4 watts.


They’re also binning P cores with that now, in a horizontal sense, but Qualcomm does that already anyways in addition to the vertical frequency binning. Either way, be it due to Apple’s tools and layout in combination (they’ve recently had notably higher clocks than Cortex Arm competitors and this has even increased in the last few years) yields, if 4.4GHz on mass parts with N3E is possible for Apple and at a power point that Qualcomm already has no issue taking with N4P, then I think we could see substantial clocks gains throughout the range with X2 on N3P. You wouldn’t get as much power savings at these peaks, but it’s still lower than AMD/Intel do at the platform level.


I think that base SKUs for X2 will be running 4GHz, with 4.2-4.6GHz models becoming much more common.

 

[FlameTail]

One interesting thing is: Qualcomm’s floor frequency for their yields is about 3.4GHz On N4P, with 4GHz also existing for some products at the top end and a more rare 4.2GHz.

?
X Elite is N4.

 

[FlameTail]

We already have an ominous Oryon V2 rumour.

Snapdragon 8 Gen 5 reportedly hits 5 GHz.

微博

随时随地发现新鲜事！微博带你欣赏世界上每一个精彩瞬间，了解每一个幕后故事。分享你想表达的，让全世界都能听到你的心声！

m.weibo.cn

 

[SpudLobby]

?
X Elite is N4.

nah pretty sure it’s N4P.

 

[FlameTail]

nah pretty sure it’s N4P.

 

[SpudLobby]

We already have an ominous Oryon V2 rumour.

Snapdragon 8 Gen 5 reportedly hits 5 GHz.

微博

随时随地发现新鲜事！微博带你欣赏世界上每一个精彩瞬间，了解每一个幕后故事。分享你想表达的，让全世界都能听到你的心声！

m.weibo.cn

I wouldn't put much stock in that. Like, 8 Gen 5 is a while off for one, but for two, that's a ridiculous gain for a phone part, and even on N3P - or N2, I don't expect to see that kind of frequency.

Like, look at the A17 Pro's frequency already on N3P and the power having gone up, they pushed it a bit too much. With N3E and maybe some tweaks, they'll keep it similar, but I doubt A18 is going to be 4.2GHz.

It's really naive to think 8 Gen 5 will run at 5GHz, at best this just says the core and process allow that kind of frequency. And we don't even know what kind of process 8 Gen 5 is going to be on.

The way I see it: if Oryon V2 gets + 25% IPC on GB6 and similar on Spec, and a 10% frequency gain at the top end + a 15+% gain for the bottom part, along with some dedicated E Cores for MT/MT^2 or MT/W and background stuff (not that they need it), it's going to be even more competitive vs AMD/Intel than current parts for numerous reasons.

 

[SpudLobby]

View attachment 101212

Why would you assume Anandtech is actually the ground source of truth on this?


Qualcomm doesn’t explicitly say N4P for the 8 Gen 3 either, but it’s widely reported it is N4P, so if the source is “well Qualcomm said N4” my answer is lol.

 

[SpudLobby]

View attachment 101208
We got some early benchmaks at 65 watts and 28 watts. Its seems ST decreases a lot when its set to 28 watts. This is lower than base M2 in CB 2024!

That doesn’t actually mean ST is consuming that much power for one thing. It just means for that tDP (which includes MT and mixed workloads) they decided to lower clocks for the dual-boost to fit within that constraint.

And second of all, who cares if it’s lower than the base M2 in ST — by like 3%? It’s the same thing at that point. No one is going to make or break a decision based off that. No one cares. AMD and Intel aren’t doing any better in CB24 at the same power lol.

Honestly the Apple fan complaints about the X Elite are extremely annoying. It’s just a bunch of “M4 better” no sh&@. We know that. QC doesn’t have to catch them right now, nor were they ever going to right off the bat if we’re being honest.

 

[FlameTail]

The power consumption of X Elite in ST is too high. 15W may be less than Intel/AMD, but it's not a huge enough difference. Apple M3 is not in this graph, but I assume it'll be like 10W.

X2 Elite should prioritise reducing this power consumption.

I annotated the above graph by mapping the curve to frequency. It's not perfect though, because the curve is for platform power, not core power.

Still, we can see that the power has roughly doubled when going from 3.4 GHz to 4.2 GHz.

 

[FlameTail]

Qualcomm doesn’t explicitly say N4P for the 8 Gen 3 either, but it’s widely reported it is N4P, so if the source is “well Qualcomm said N4” my answer is lol.

Yeah, it would be pretty suspect if the X Elite is on standard N4, when N4P is available.

The explanation for why X Elite would be using N4, is that it's development had started at a time when N4P was not available. It's a reasonable explanation. But Qualcomm could have easily ported it from N4 -> N4P. According to Semiaccurate, we know that they did multiple re-spins of the Hamoa die...