Discussion Qualcomm Snapdragon Thread

[Raqia]

This is interesting;

How Chrome doubled its Speedometer scores on Android

Today’s The Fast and the Curious post covers how Chrome achieved best-in-class Speedometer scores on mobile devices, resulting in faster...

blog.chromium.org

The higher level efforts at in OS scheduler level are a bit unsung for general performance. Xiaomi apparently cuts CPU frequency and boosts memory frequency when it detects CPU stalls due to memory pressure:

Qualcomm seems to have finally achieved parity with Apple for web performance and one of the last differentiating features at the silicon level between iOS and Android has fallen, and I think for the other disciplines (GPU, ISP, NPU, Modem etc.) it has been consistently better for several years. It's just up to partners to build and polish features on top of this foundation now...

That said, Apple enjoys a much lower cost per manufactured SoC due to its volume and close partnership with TSMC in funding and early adopting new nodes.

 

[jdubs03]

The higher level efforts at in OS scheduler level are a bit unsung for general performance. Xiaomi apparently cuts CPU frequency and boosts memory frequency when it detects CPU stalls due to memory pressure:

Qualcomm seems to have finally achieved parity with Apple for web performance and one of the last differentiating features at the silicon level between iOS and Android has fallen, and I think for the other disciplines (GPU, ISP, NPU, Modem etc.) it has been consistently better for several years. It's just up to partners to build and polish features on top of this foundation now...

That said, Apple enjoys a much lower cost per manufactured SoC due to its volume and close partnership with TSMC in funding and early adopting new nodes.

I don’t have Chrome to test on my I16PM but that is definitely close.

-Two run average on Safari was 33.5.

I’ll test again when iOS 18.2 comes out.

 

[Tup3x]

I don’t have Chrome to test on my I16PM but that is definitely close.

-Two run average on Safari was 33.5.

I’ll test again when iOS 18.2 comes out.

Chrome on iOS doesn't have anything to so with Android Chrome since third party browser engines are banned on iOS. Although now it probably should be possible at least on EU.

 

[igor_kavinski]

https://videocardz.com/newz/geekom-qs1-pro-mini-pc-specs-leaked-featuring-qualcomm-x1e-80-100-processor

DDR5-5600 is gonna be ugly for the iGPU. Let's hope that some of the cores can be disabled to extract more performance out of the remaining ones in memory intensive workloads.

 

[FlameTail]

DDR5-5600 is gonna be ugly for the iGPU.

i wonder if that's a typo or not. LPDDR5-5600 is 4 year old technology.

The memory controller in X Elite supports upto LPDDR5X-8448.

Let's hope that some of the cores can be disabled to extract more performance out of the remaining ones in memory intensive workloads.

What's that logic?

 

[The Hardcard]

i wonder if that's a typo or not. LPDDR5-5600 is 4 year old technology.

The memory controller in X Elite supports upto LPDDR5X-8448.

What's that logic?

It is probably not a typo, but a need to control costs, Top spec memory is expensive and in demand. Likely few discounts.

They are already at a price that will be a tough sell outside of people who really want Qualcomm. There is not even a battery claim here, why would people pay $200 to $300 more than this over a Hawk Point mini PC.

 

[igor_kavinski]

What's that logic?

Fewer cores would be better able to work fully with the limited bandwidth of DDR5-5600 than all the cores causing bandwidth contention and bogging down the multithreaded computing effort.

I think this is one reason why some games are seeing better framerates with Lion Cove cores disabled. They are so hungry that they gob up the limited bandwidth but don't contribute enough to the computing but with them disabled, the Skymonts are able to work at their best with just one lonely Lion Cove playing with itself.

 

[FlameTail]

Rumour: Adreno 840 in Snapdragon 8 Elite Gen 2 will have 18 CUs.

https://twitter.com/x/status/1863240380311126458

That's 50% more CUs than the Adreno 830 in 8 Elite. A whopping 2304 FP32 ALUs.

Does not necessarily mean that peak performance will be 50% higher, as that depends on the clock speed, but the efficiency uplift from widening the GPU by 50% will be terrific.

I just realised what the "18m" in Reve's tweet meant. It means 18 MB of GMEM!

So Adreno 840 will add 6 CUs and also 6 MB of GMEM, keeping the 1:1 CU:GMEM ratio introduced with Adreno 830. This aspect of the Adreno 8 GPU architecture fascinates me. This is more cache than even some laptop SoC iGPUs.

Widening the GPU by 50% is also going to add ~10 mm² to the die size. CPU block will also be bigger due to the addition of SME, as well larger cores themselves.

It looks like 8 Elite Gen 2 is on track to become the largest Snapdragon 8 mobile chip ever!

Will surely exceed 140 mm², and maybe even touch 150 mm²!

 

[adroc_thurston]

This is more cache than even some laptop SoC iGPUs

That's not cache.

 

[FlameTail]

This is interesting;

No plans for discrete GPUs?

 

[Ghostsonplanets]

This is interesting;
View attachment 112792
No plans for discrete GPUs?

dGPU means they would need to re-engineer their whole GPU arch and also do tons of R&D and support towards discrete features.

Non-starter given their core biz is mobile.

 

[FlameTail]

dGPU means they would need to re-engineer their whole GPU arch and also do tons of R&D and support towards discrete features.

Non-starter given their core biz is mobile.

I meant pairing up Nvidia/AMD/Intel dGPUs to Qualcomm Snapdragon X SoCs.

 

[igor_kavinski]

The Elite chip smokes Luna Lake in CB R24 and wins in battery life by 3%.

 

[SpudLobby]

The Elite chip smokes Luna Lake in CB R24 and wins in battery life by 3%.

I have been pointing this out about the X Elite for months now. True of the Plus chips that are cheaper too: even those 8 core parts at least match or beat Lunar in MT no problem with barely any ST difference in practice, and match it in battery life. And this is the JV roster from QC looking at V2 already. Panther Lake will be a regression in a few aspects and Intel’s fundamental problems remain with area, energy, idle power, scaling — that’s why they have to have LPE Skymont. Can’t use the regulars because they sacrifice low load efficiency with a bunch of extra ring power.

Just makes everything complicated and is why they’ll have 3 very much distinct core cluster types again. QC doesn’t have this issue and can already take it to Intel with a homogenous 12C part. Only downhill from here for Intel really. Energy and responsiveness gaps will widen.

 

[SpudLobby]

The Lenovo ThinkPad T14s Gen 6 provides a very interesting comparison when it comes to battery life, both the ThinkPad T14s Gen 6 Qualcomm and ThinkPad T14s Gen 5 Intel we tested had the same battery capacity (58 Wh) and screen type (1,920 x 1,200 low power). The verdict is clear: While the T14s Gen 6 AMD easily beats the Intel Core Ultra model in our Wi-Fi test (150 cd/m² screen brightness), it is not even close to the Snapdragon X Elite model in endurance.

Ryzen HX 360: 847 minutes
Snapdragon X Elite: 1349 minutes.

44% more at the same laptop, display, battery size. In a Wi-Fi browsing test so CPU use & some intermittent lower load/idle.

In practice most people won’t get 22 hours, but it’s the principle here. Even turning brightness up etc you might be looking at the difference between 6 and 9 hours of real use. Pretty pathetic from AMD.

 

[SpudLobby]

https://www.notebookcheck.net/Lenovo-ThinkPad-T14s-Gen-6-laptop-review-The-AMD-version-returns-with-the-Ryzen-AI-7-Pro-360.923414.0.html

 

[igor_kavinski]

Pretty pathetic from AMD.

They probably couldn't foresee that Snapdragon would eat their lunch in battery life yet deliver acceptable performance. Nuvia was going for servers then QC bought them and diverted the SoC to WinARM which didn't give AMD or Intel enough time to take reactive measures. Intel kinda got lucky thanks to targeting M series level of performance/watt and beats Elite X in Load battery test easy peasy. Zen 6 or Sound Wave APU should be a better comparison.

 

[SpudLobby]

They probably couldn't foresee that Snapdragon would eat their lunch in battery life yet deliver acceptable performance. Nuvia was going for servers then QC bought them and diverted the SoC to WinARM which didn't give AMD or Intel enough time to take reactive measures. Intel kinda got lucky thanks to targeting M series level of performance/watt and beats Elite X in Load battery test easy peasy. Zen 6 or Sound Wave APU should be a better comparison.

lol we hear this like every year. 4 years since M1 and Qualcomm is much closer than AMD is on ST curves and battery life, and we know that’s their C+ implementation. Intel’s Lunar was similar but a Herculean effort and too expensive, with meh throughput.

And we of course know Oryon V2 is a huge gain and V3 is what’s next.

Intel beating Qualcomm in a load battery test is irrelevant without knowing the performance. It’s a different kind of test. Ok, it lasts 120 minutes instead of 60 and your workload takes twice as long etc.

You can easily change that too by just changing the power policy and I suspect at a 15-35W profile in an average fan laptop Intel gets blown out or at least matched.

 

[igor_kavinski]

And we of course know Oryon V2 is a huge gain and V3 is what’s next.

I'm sure that AMD management has set Sound Wave APU's target as beating both Mediatek and QC. Otherwise, they may simply not manufacture it at volume and just use the available fab capacity for Zen 6. If Intel can push x86 as far as they did with Lunar Lake, it gives me hope that AMD can do something special with their ARM SoC. Their engineers are definitely more brilliant than Intel's.

 

[FlameTail]

I'm sure that AMD management has set Sound Wave APU's target as beating both Mediatek and QC. Otherwise, they may simply not manufacture it at volume and just use the available fab capacity for Zen 6. If Intel can push x86 as far as they did with Lunar Lake, it gives me hope that AMD can do something special with their ARM SoC. Their engineers are definitely more brilliant than Intel's.

You think the Sound Wave APU will contain a custom ARM core designed by AMD? I doubt it.

Anyways, ARM isn't happy now that everyone started designing their own cores.

View attachment 112871View attachment 112872
Now there are 3 vendors with their own set of custom ARM CPU cores (Big + small).

Apple
Qualcomm
Hauwei

A remarkable development. One year ago, it was only Apple.

According to the big gChips leak in October, Google is also working on custom ARM cores (Orion, Orion-E) for their 2027 Pixel phone bound Tensor G7 SoC.

It's also interesting to compare the cache setups. Apple, Qualcomm and Google's designs all have big shared L2 caches. Whereas Huawei's Taishan core has a cache hierarchy similar to stock ARM cores (pL2, sL3).

We are witnessing a renaissance of custom ARM core designs.

I guess ARM the company isn't very happy about this, because ALAs have lower royalty rates than TLAs.

They all had various reasons to pursue custom ARM cores. Apple of course needs no explaining. They have been making custom ARM for a decade+.

For Qualcomm, it was so that they can;
(1) Create CPU cores that are more powerful and efficient than stock ARM cores.
(2) Reduce the royalty rates being paid to ARM.

As for Huawei, they had no other choice really. They can't license the big ARM cores (Cortex X, A7xx), because those are designed in America, and Huawei is under heavy US sanctions.

Not sure what Google's objectives might be. Will have to see when their custom core comes to market.

 

[The Hardcard]

Ryzen HX 360: 847 minutes
Snapdragon X Elite: 1349 minutes.

44% more at the same laptop, display, battery size. In a Wi-Fi browsing test so CPU use & some intermittent lower load/idle.

In practice most people won’t get 22 hours, but it’s the principle here. Even turning brightness up etc you might be looking at the difference between 6 and 9 hours of real use. Pretty pathetic from AMD.

I disagree. I think the key mainstream metric is getting through the day. “Oh I forgot to bring my charger, will it last until I get back.” The AMD is doing that for most people. Not that the greater battery won’t turn some to the Qualcomm, but the impact past one day is not nearly as great.

AMD and Intel clearly have a lot of work to do, they can’t sit on this, but neither are in danger of a marketshare collapse.

 

[FlameTail]

I just realised what the "18m" in Reve's tweet meant. It means 18 MB of GMEM!

So Adreno 840 will add 6 CUs and also 6 MB of GMEM, keeping the 1:1 CU:GMEM ratio introduced with Adreno 830. This aspect of the Adreno 8 GPU architecture fascinates me. This is more cache than even some laptop SoC iGPUs.

Widening the GPU by 50% is also going to add ~10 mm² to the die size. CPU block will also be bigger due to the addition of SME, as well larger cores themselves.

It looks like 8 Elite Gen 2 is on track to become the largest Snapdragon 8 mobile chip ever!
View attachment 112753
Will surely exceed 140 mm², and maybe even touch 150 mm²!

Bigger SoCs cost more to manufacture, and Qualcomm is passing on the cost (with margins added on top) to OEMs.

8+G1 to 8 Elite, the price has doubled!


This is putting pressure on OEMs. You can see it with the recently unveiled Snapdragon 8 Elite phones. Some aspects such as cameras are a downgrade compared to the predecessor.

This year’s Qualcomm Snapdragon 8 Elite is a prime example; it offers significant leaps in benchmark performance but is said to be considerably more expensive than its predecessor. Likewise, MediaTek’s Dimensity 9400 commands a higher price tag too, yet you’ll struggle to tell a phone packing these chips apart from last year’s model when it’s in hand. Money thrown at higher benchmark scores could be spent on better cameras, newer battery tech, or anything else on your wishlist. Returning to the ROG, the baseline model has dropped the telephoto camera, no doubt to keep up with the spiraling costs of top-tier performance.

With reports that the next-gen Snapdragon 8 Elite 2 will be “significantly” more expensive once again, something has to give; be it higher price tags or sacrificing other key specs. But is that a trade-off any of us are really willing to make?

That raises the question, when will Qualcomm move off this trajectory?

Density scaling with new process nodes has slowed to a crawl, so the only way to make SoCs with significantly better performance is by making them bigger. Since cost-per-transistor scaling is also dead, bigger SoCs cost more to make.

 

[SpudLobby]

I mean OEMS can just pass this on to consumers. Kind of tired of this absurd sticky thinking. Just force them to accept the costs lol, don’t give them outs like with Intel and Panther’s optional PMIC setup etc - PC OEMs and fabless designers are worst about this, whereas in smartphones and tablets we see much more competitive dynamics from displays (early on) to power management.

And secondly the growth in silicon content — the fact that they’re able to charge this and people keep buying $700-1200 smartphones — is just an outgrowth of the value provided even if it costs more than it would in a Moores law utopia. Smartphones are much more capable as pocket computers now than in 2015. Not the same as the 2015 -> 2007 gap, but still notable. It’s your camera, GPS, music player, social media device, and beyond that gaming is a thing for many however silly it seems.

I’d also point out I use my iPhone for casual web browsing more often than I did in 2016, just because it’s a bit more enjoyable and fluent in doing so before I make the handoff to a PC. (Which granted is superior but for some stuff it isn’t always necessary, this uber casual threshold has shifted I think).

 

[LightningZ71]

Google moves a LOT of Pixel phones with notably slower SoCs. I have an 8, and I'm not dissatisfied with ANY aspect of it's SoC performance. I find that it is more than fast enough to do any web browsing that I do, run any of my frequently used apps, and process any of my photos or videos.

I'm of the opinion that, in the VAST majority of cases, users would be happier with far more modest CPU core upgrades and more focus on the companion logic blocks. Keep package size in check, focus on the aspects of core performance that are under the most pressure, and make the biggest improvements in things like the DSP, camera logic, and aspects of the NPU that are actually being used.

I'm not even upset that Tensor might see no generational performance improvement in the pixel 10, so long as the cameras continue to improve, the radios continue to behave better, efficiency improved, and Gemini gets more features. If Google can do all that while holding the line on pricing, they will sell as many as they make.

 

[DZero]

Google moves a LOT of Pixel phones with notably slower SoCs. I have an 8, and I'm not dissatisfied with ANY aspect of it's SoC performance. I find that it is more than fast enough to do any web browsing that I do, run any of my frequently used apps, and process any of my photos or videos.

I'm of the opinion that, in the VAST majority of cases, users would be happier with far more modest CPU core upgrades and more focus on the companion logic blocks. Keep package size in check, focus on the aspects of core performance that are under the most pressure, and make the biggest improvements in things like the DSP, camera logic, and aspects of the NPU that are actually being used.

I'm not even upset that Tensor might see no generational performance improvement in the pixel 10, so long as the cameras continue to improve, the radios continue to behave better, efficiency improved, and Gemini gets more features. If Google can do all that while holding the line on pricing, they will sell as many as they make.

Even more with Huawei and their Hisilicon Kirin.