Discussion Qualcomm Snapdragon Thread

[FlameTail]

Negativeonehero:

I have seen many game reviews for the Snapdragon 8 Elite.
It tends to heat up quickly if not properly tuned.
The worst offender is, once again, the "small core only" strategy of Xiaomi devices. The small core's efficiency is actually not good, and the gaming performance is

not much better than some 8 Gen 3 devices.
I'm not saying that the 8 Elite is "a fire dragon", in fact a lot of phones with good tuning have great performance. But just one phone with bad tuning will shift the view of people about the chipset. Xiaomi 15 is a popular one, in fact.

It seems the scheduling strategy is very important.

Data from Geekerwan and others do confirm that the efficiency of Oryon-M is not as good as the A18 E-core or the Dimensity 9400's Cortex A720.

 

[Meteor Late]

Yeah, like I said, Oryon M core is not that impressive, all I see is a core clocked very high to boost MT as much as possible, unlike Apple that clocks its E core very low, there is no way it doesn't clock much higher with its efficiency.

 

[FlameTail]

Jukanlosreve [X]

Why did ARM sue Qualcomm?

In its third-quarter financial report released on the 7th (local time), royalty issues between Arm and Qualcomm became apparent. For the second consecutive quarter, Armv9 royalties as a percentage of revenue have stagnated at around 25%. This is the first time this has occurred since Armv9 royalties were included in the financial statements. Armv9 is the 9th version of the Arm ISA and serves as the foundation for Oryon. If Arm had increased the royalty rate for Oryon from Qualcomm, the results might have been different.

PC royalties are also on a downward trend. While Arm’s share of royalties in the consumer electronics market (including PCs) has increased from 24% to 30% over the past two years, royalty prices have decreased by 23%. This is partially related to Snapdragon X. Due to the low royalty rate, Qualcomm’s manufacturing costs for Snapdragon X are not high, allowing Microsoft to procure Snapdragon X at a low cost and set a significantly lower price for its Copilot PC. This pricing structure is reducing Arm’s royalty revenue.

 

[hemedans]

Yeah, like I said, Oryon M core is not that impressive, all I see is a core clocked very high to boost MT as much as possible, unlike Apple that clocks its E core very low, there is no way it doesn't clock much higher with its efficiency.

While efficient wise is not that Impressive that thing use less area than even Cortex A720, it will be crucial for Qualcomm if they want to use it in midrange and lowend soc.

 

[Raqia]

Yeah, like I said, Oryon M core is not that impressive, all I see is a core clocked very high to boost MT as much as possible, unlike Apple that clocks its E core very low, there is no way it doesn't clock much higher with its efficiency.

Disagree w/ the "higher clock = generally bad" argument given the relative efficiency of the core at those speeds.

There's some cognitive dissonance given the inefficiency of prior higher clocked CPUs like the P4, POWER5 or Bulldozer, but the operating clocks are an implementation detail: the machine was designed to operate at those clocks. Memory latency for instance is a lot less impactful at lower clocks and higher clocked designs will need more aggressive ways to achieve data-locality (better pre-fetchers, more load-store resources etc.).

The point is: a win is a win if the sum of those design decisions still results in better PPA.

 

[Meteor Late]

Disagree w/ the "higher clock = generally bad" argument given the relative efficiency of the core at those speeds.

There's some cognitive dissonance given the inefficiency of prior higher clocked CPUs like the P4, POWER5 or Bulldozer, but the operating clocks are an implementation detail: the machine was designed to operate at those clocks. Memory latency for instance is a lot less impactful at lower clocks and higher clocked designs will need more aggressive ways to achieve data-locality (better pre-fetchers, more load-store resources etc.).

The point is: a win is a win if the sum of those design decisions still results in better PPA.

I agree, I just mean that people compare it with Apple E cores that are weak in comparison, what I'm saying is that Apple is deliberately clocking these cores very low, if you check their architecture diagram and their power consumption graph, I don't see how that's anywhere near what's possible for Apple E cores.

 

[Panino Manino]

Which reminds me, it's still possible to install some application on current gen devices to control governor and clock speed?
I'm the past it was possible, I remember even changing the voltage.

 

[hemedans]

Which reminds me, it's still possible to install some application on current gen devices to control governor and clock speed?
I'm the past it was possible, I remember even changing the voltage.

Gpu yes cpu you can only reduce frequency.

We use konabess, custom kernel, you need root.

 

[Raqia]

I agree, I just mean that people compare it with Apple E cores that are weak in comparison, what I'm saying is that Apple is deliberately clocking these cores very low, if you check their architecture diagram and their power consumption graph, I don't see how that's anywhere near what's possible for Apple E cores.

It may also mean the pipeline length for Apple's E-cores is simply shorter which has trade offs for peak clock (i.e. when some pipeline stage has to take a certain amount of time by design because more gets done relative to a longer pipeline design...). I could be wrong, but I don't remember the pipeline length for the E-cores off the top of my head...

 

[soresu]

Data from Geekerwan and others do confirm that the efficiency of Oryon-M is not as good as the A18 E-core or the Dimensity 9400's Cortex A720.

Still nothing using A725 to compare with?

 

[hemedans]

Still nothing using A725 to compare with?

D8400 is coming soon, problem is many midrange use 4nm, it won't be fair comparison.

 

[soresu]

D8400 is coming soon, problem is many midrange use 4nm, it won't be fair comparison.

Gives a rough estimate at least when you factor in the claimed node differences.

 

[FlameTail]

Hamoa SKUs	Boost clock	Base clock
X1E-00-1DE	4.3 GHz	3.8 GHz
X1E-84-100	4.2 GHz	3.8 GHz
X1E-80-100	4.0 GHz	3.4 GHz
X1E-78-100	-	3.4 GHz
X1P-66-100	4.0 GHz	3.4 GHz
X1P-64-100	-	3.4 GHz

As I have said previously, it is ridiculous how much frequency variation there is between the Snapdragon X1 SKUs.

And with Snapdragon X Elite, the higher SKUs are also rare. The only device with the 4.3 GHz SKU was cancelled (Dev Kit), there is only one device with the 4.2 GHz SKU (Galaxy Book Edge 16) and only a few devices have the 4.0 GHz SKU.

Now it seems Qualcomm has 'fixed' Oryon with the 2nd gen Oryon cores in Snapdragon 8 Elite. They are shipping millions of 8 Elite chips with a clock speed of 4.32 GHz.

So I expect there won't be such frequency variation with the Snapdragon X2 SKUs.

Even AMD has hardly any frequency variation in their latest laptop SKUs;

 

[FlameTail]

Rumour:

"News has come out about the Dimensity 9500 and Snapdragon 8 Elite 2. Both chips are expected to see a 20% increase in single and multi performance thanks to SME. (For the Snapdragon 8E2, the single-core score is 4000 on the GB6.)

By the way, the 8G5 uses a mix of Samsung Foundry SF2 and TSMC N3P"

Source

SF2 is a renamed node, previously known as SF3P.

 

[poke01]

@FlameTail I don’t buy that rumour.

Qualcomm never targets 20%, my take is that that’s the minimum percentage improvement from SME only. The score will go higher as QC will have new cores with higher IPC and clocks.

If the 5.0GHz clock is true then I am expecting 4500-5000 in ST GB6. Qualcomm now achieves 3200 at 4.3GHz. 20% from SME will be 4000. IPC improves will be 10-20% and clocks may increase 15%.

People are going to be shocked next year, I think.

 

[jdubs03]

@FlameTail I don’t buy that rumour.

Qualcomm never targets 20%, my take is that that’s the minimum percentage improvement from SME only. The score will go higher as QC will have new cores with higher IPC and clocks.

If the 5.0GHz clock is true then I am expecting 4500-5000 in ST GB6. Qualcomm now achieves 3200 at 4.3GHz. 20% from SME will be 4000. IPC improves will be 10-20% and clocks may increase 15%.

People are going to be shocked next year, I think.

I think you are very much overestimating the performance increase. I am highly skeptical that there will be another 40+ percent performance gain. Just seems like far too much. Big time kudos if they can, but I just don’t think that is a likely outcome. A score of 4000 would be a 25% increase, that would be impressive by itself.

Also keep in mind the performance jump from the A17 Pro to the A18 Pro was practically the same in GB6 as it was in GB5.

 

[FlameTail]

@FlameTail I don’t buy that rumour.

Qualcomm never targets 20%,

Wot?

8G2 -> 8G3 was only a 15% uplift in Single Core. The 'normal' YoY ST uplift in this industry is 10-15%. Anything more than that is a bonus.

That said though, I do agree I am expecting more than 20% in this scenario. Indeed, the 4000 figure from the same rumour contradicts that, because 3200 -> 4000 is a 25% uplift, not 20%.

my take is that that’s the minimum percentage improvement from SME only.

Looking at A18 Pro scores in GB6.2 and GB6.3, it seems SME only adds about 5% to the score.

For a 20% improvement from SME alone, Qualcomm would have to put a chunges SME block with like 100 TOPS INT8, which I don't just see happening.

The score will go higher as QC will have new cores with higher IPC and clocks.

If the 5.0GHz clock is true then I am expecting 4500-5000 in ST GB6.

Excessive expectations may lead to catastrophic disappointments.
I'll go with the above rumour and set an expectation of 4000. Anything more than that would be a treat.

Also it's interesting to think what this bodes for Snapdragon X2.

 

[Magio]

Also it's interesting to think what this bodes for Snapdragon X2.

I hope at the very least it's on the same node(s) as 8 Elite this time. If it is they'll definitely try to have it compete with M5 (and maybe M5 Pro) tho Windows will be a bottleneck to that. Fingers crossed actual Linux support will be ready near launch for the 2nd gen...

 

[FlameTail]

Rumour:


According to previous rumours, 8 Elite G2 was supposed to be dual sourced between Samsung Foundry (SF2) and TSMC (N3P). If this rumour is true, those plans have been cancelled by Qualcomm.

This is not a good sign about the viability of the Samsung SF2 node.

 

[FlameTail]

David Huang tweet:

This year's test of ARM processors gave me a very bad impression, that is, it is extremely difficult to run at full frequency, as if the entire Fmax is specially customized for digital media to play with dry ice and liquid nitrogen to brush scores. Qualcomm 8E is fine as a mobile phone processor, even Mac Mini can score a full generation higher at low temperatures, so which score should be used?

As I'm playing, I've begun to miss the days when the M1 max would not throttle no matter how hard I pushed it.

Translated from Chinese to English by Google.

https://twitter.com/x/status/1856331643742761102

The translation isn't very good, so it's difficult to fully understand what he's saying.

 

[DZero]

Rumour:

View attachment 111496
According to previous rumours, 8 Elite G2 was supposed to be dual sourced between Samsung Foundry (SF2) and TSMC (N3P). If this rumour is true, those plans have been cancelled by Qualcomm.

This is not a good sign about the viability of the Samsung SF2 node.

So Samsung is about to fail in their semiconductor bussiness?

 

[FlameTail]

So Samsung is about to fail in their semiconductor bussiness?

Doubt it. Not anytime soon.

But the signs aren't good. Samsung lost all their big customers (eg: Nvidia, Qualcomm), and are starting to lose their smaller customers too (eg: Google).

 

[DrMrLordX]

So Samsung is about to fail in their semiconductor bussiness?

As long as they keep getting money from their government, no. It just means they're likely to be their only customer. Or one of the few.

 

[Hitman928]

So Samsung is about to fail in their semiconductor bussiness?

Samsung foundries still push out a ton of volume for memory chips, AFAIK. Their logic foundry efforts have always lagged and struggled for consistent customers, though it's gotten significantly worse with the recent nodes.

 

[LightningZ71]

Samsung is slowly dropping behind Intel and TSMC on what their REAL leading edge node can do. They aren't in the gutter like GlobalFoundries, but they certainly aren't keeping up. They have quite innovative technologies, but, from what we see and hear, their yields are just not up to industry standards and their different nodes struggle to meet their power/performance/density triangle targets.