Discussion Qualcomm Snapdragon Thread

[The Hardcard]

According to rumours from last year, Qualcomm supposedly had two versions of the 8 Gen 4: (1) Using Cortex cores and (2) Using Oryon cores.

Evidently, they dropped the former and didn't tapeout that chip.

ARM's boldness is certainly surprising. Whatever the outcome of the lawsuit maybe, this ordeal is incredibly damaging to the ARM ecosystem.

For Apple, we know that 3 E-cores ≈ 1 P-core in terms of performance.

	# of cores	# of cores (In terms of P)	Cinebench 2024 Multi Core Score	Contribution of 1 P-core to CB2024 Multi Core score
M3	4P+4E	5.33P	710	133.2
M3 Max	12P+4E	13.33P	1700	127.5
M4	4P+6E	6P	977*	162.8
X Elite	12P	12P	1220	101.6

	1T boost clock speed	All core clock speed	Geekbench 6 Single Core	Cinebench 2024 Single Core
M3	4.05 GHz	3.6 GHz	3150	141
M3 Max	4.05 GHz	3.6 GHz	3150	141
M4	4.5 GHz	4.04 GHz	3800	174*
X Elite	4.3 GHz	3.8 GHz	3200**	133

How is the Apple M3's 'Contribution of 1 P-core to CB2024 Multi Core score' so much higher than X Elite?

From the 1T tests, we know M3 and X Elite score similarly.

Differences in all-core clock speeds doesn't explain it, becauss both M3 and X Elite have an all-core clock speed that's about 10% less than the 1T boost clock speed used in the ST tests.

____

*Leaked score based on M4 Macbook Pro
**X Elite top SKU score in Linux

The operating system stack plays a role in multi threading, from the kernel through the frameworks to the application interface.

Asahi Linux is running on Apple Silicon, but since it requires reverse engineering the chipset, I don’t know how optimized multithreading is. Plus I know they haven’t reverse engineered the M3 yet. That affects GPU, I don’t know if there are CPU issues. I’m also unclear the status of Linux on X Elite.

But the closest you can get to this comparison would be to compile matching Linux kernels and software (maybe SPEC, a raytrace renderer, and a software-based video transcoder ) on the platforms.

I know Phoronix has an extensive benchmark suite that I believe is publicly available. They (he?) has run it on Apple Silicon in the past, but I think just on M1. If someone ever had the time and interest, there would be a much more accurate comparison of SOC and core differences.

 

[FlameTail]

Comparison of Oryon vs Apple CPU Microarchitectures (Block diagrams from Geekerwan)


Oryon-L has 2 less INT units than M4-P (and M3-P too). That would explain it's weaker showing in SPEC2017 INT.

Oryon-M has 1 more INT unit but one less FP unit, compared to M4-E.

There are many more differences, and the diagram for Oryon-L/Oryon-M are not fully fleshed out.

 

[Meteor Late]

There is a good chance that Apple E cores are heavily downclocked and way below their true potential. We don't know if they couldn't be up there with Oryon M at, say, 3.5GHz, and how bad their efficiency would get. But it's quite eye opening that Oryon M even when downclocked to match Apple E core performance, it still needs more power to do so.

 

[soresu]

Comparison of Oryon vs Apple CPU Microarchitectures (Block diagrams from Geekerwan)
View attachment 110349
View attachment 110351
Oryon-L has 2 less INT units than M4-P (and M3-P too). That would explain it's weaker showing in SPEC2017 INT.

Oryon-M has 1 more INT unit but one less FP unit, compared to M4-E.

There are many more differences, and the diagram for Oryon-L/Oryon-M are not fully fleshed out.

M4-E seems well balanced between int and fp.

 

[FlameTail]

Adreno 830 has 3 Slices with 4 CUs each, for a total of 12 CUs.

1 Slice = 4 CUs

Is this a fixed amount? Can they change the size of a Slice?

If they are scaling up to make a bigger GPU (eg: for laptop SoC), it's pretty straightforward to add more Slices.

But if they are scaling it down to make a smaller GPU (eg: midrange smartphone SoC), how will they do it?

 

[soresu]

But if they are scaling it down to make a smaller GPU (eg: midrange smartphone SoC), how will they do it?

Obviously by decreasing the number of slices to 1 or 2, and they can further divide SKUs by scaling clock frequency.

Also this is the CPU sub forum, we should keep this just for Oryon related stuff in Snapdragon (or possible future server SoCs), and make an Adreno thread over on the GPU side.

 

[FlameTail]

Obviously by decreasing the number of slices to 1 or 2, and they can further divide SKUs by scaling clock frequency.

Let's look at last generation's Snapdragon lineup and their iGPUs;

SoC	Adreno Model	Clock Speed	FP32 ALUs
8 Gen 3	750	903 MHz	1536
8s Gen 3 7+ Gen 3	735 732	1100MHz 950 Mhz	768
7 Gen 3	720	935 MHz	256

The 7 Gen 3 has 1/6th the GPU of the 8 Gen 3.

Going by this, it seems to me that anything below the 8s Gen4/7+Gen4 isn't going to have a Sliced GPU (Can you call it a sliced GPU if there is only 1 slice?)

Note that the 8 Elite isn't the only SoC with an Adreno 8 series GPU. There's also the 7s Gen 3 that Qualcomm announced a few months back, which supposedly has the Adreno 810 (1 CU = 128 FP32 ALUs).

Also this is the CPU sub forum, we should keep this just for Oryon related stuff in Snapdragon (or possible future server SoCs), and make an Adreno thread over on the GPU side.

Hmm, isn't it fairly normal for the iGPU to be brought up in SoC/APU discussions?

Also, I doubt having a seperate thread for Adreno will be worth it. Adreno is poorly documented, and Qualcomm seldom divulges any details about it. And nobody is writing in-depth reviews and comparisons of mobile GPU architectures nowadays. I don't see a dedicated thread for Apple GPUs, and those are better documented than Adreno.

 

[FlameTail]

There is a good chance that Apple E cores are heavily downclocked and way below their true potential. We don't know if they couldn't be up there with Oryon M at, say, 3.5GHz, and how bad their efficiency would get. But it's quite eye opening that Oryon M even when downclocked to match Apple E core performance, it still needs more power to do so.

Eyeballing Geekerwan's SPEC2017 graphs... Oryon-M needs about 15%/30% more power to match A18-E in INT/FP respectively.

It's not terrible. I am still impressed that they managed to build a brand new E-core from ground up. Let's see where they take this next generation.

 

[FlameTail]

Snapdragon 8 Elite Has Full Linux Support, Making It Easier To Run PC Games Through Emulation, GPU Side To Require Patches To Ensure Optimum Performance

A Qualcomm executive has said that Linux support for the Snapdragon 8 Elite went live just a few hours after the chipset’s announcement

wccftech.com

Linux support for a mobile SoC? What do you guys make of this news?

 

[Nothingness]

Snapdragon 8 Elite Has Full Linux Support, Making It Easier To Run PC Games Through Emulation, GPU Side To Require Patches To Ensure Optimum Performance

A Qualcomm executive has said that Linux support for the Snapdragon 8 Elite went live just a few hours after the chipset’s announcement

wccftech.com

Linux support for a mobile SoC? What do you guys make of this news?

Little. Qualcomm cancelled the only device I was considering to run Linux, the dev kit. Laptops are different beasts, and running an OS not supported by the vendor laptop can make the device a displeasure to use (this also applies to x86 laptops). And I won't even talk about running Linux on a smartphone.

 

[soresu]

I don't see a dedicated thread for Apple GPUs, and those are better documented than Adreno

Adreno is plenty documented.

It has a fully compliant VK and OGL OSS driver in Mesa.

 

[DZero]

Let's look at last generation's Snapdragon lineup and their iGPUs;

SoC	Adreno Model	Clock Speed	FP32 ALUs
8 Gen 3	750	903 MHz	1536
8s Gen 3 7+ Gen 3	735 732	1100MHz 950 Mhz	768
7 Gen 3	720	935 MHz	256

The 7 Gen 3 has 1/6th the GPU of the 8 Gen 3.

Going by this, it seems to me that anything below the 8s Gen4/7+Gen4 isn't going to have a Sliced GPU (Can you call it a sliced GPU if there is only 1 slice?)

Note that the 8 Elite isn't the only SoC with an Adreno 8 series GPU. There's also the 7s Gen 3 that Qualcomm announced a few months back, which supposedly has the Adreno 810 (1 CU = 128 FP32 ALUs).

Results are leaked.
If those values are true and compared to SD 865, the result is insane.... in few gens Qualcomm put a flagship device on shame.

If the Adreno 810 is being started to use in the SD 4 and 6 series, it would bring a BIG pressure to Mediatek.

 

[hemedans]

Snapdragon 8 Elite Has Full Linux Support, Making It Easier To Run PC Games Through Emulation, GPU Side To Require Patches To Ensure Optimum Performance

A Qualcomm executive has said that Linux support for the Snapdragon 8 Elite went live just a few hours after the chipset’s announcement

wccftech.com

Linux support for a mobile SoC? What do you guys make of this news?

There was earlier report of Qualcomm making X86 to Arm translator for other Os to run windows games.

Qualcomm is laying the foundation for a proper Steam Deck rival

A Qualcomm executive has strongly hinted that the company is laying the groundwork for a Snapdragon-powered Steam Deck rival.

www.androidauthority.com

 

[moinmoin]

Snapdragon 8 Elite Has Full Linux Support, Making It Easier To Run PC Games Through Emulation, GPU Side To Require Patches To Ensure Optimum Performance

A Qualcomm executive has said that Linux support for the Snapdragon 8 Elite went live just a few hours after the chipset’s announcement

wccftech.com

Linux support for a mobile SoC? What do you guys make of this news?

Technically all Android systems use Linux as their underlying kernel. The question is whether the hardware support is being upstreamed or limited to some unupgradeable binary blob.

 

[soresu]

Technically all Android systems use Linux as their underlying kernel. The question is whether the hardware support is being upstreamed or limited to some unupgradeable binary blob.

AFAIK past versions of Android decoupled various features like drivers so that they could be updated through the Play Store.

 

[moinmoin]

AFAIK past versions of Android decoupled various features like drivers so that they could be updated through the Play Store.

The point is that the only real support that counts is open source upstreamed to the official Linux kernel. Everything else is bound to fall by the wayside, meaning hardware support is most likely going to be missing outside of very specific packages and kernel versions supported by the SoC manufacturer.

 

[poke01]

https://twitter.com/x/status/1850599912167153910

https://twitter.com/x/status/1850614037597593678

i just can't...

The 8 Elite GPU is already 21mm2.

 

[DrMrLordX]

Little. Qualcomm cancelled the only device I was considering to run Linux, the dev kit. Laptops are different beasts, and running an OS not supported by the vendor laptop can make the device a displeasure to use (this also applies to x86 laptops). And I won't even talk about running Linux on a smartphone.

If there were a market for it, someone could produce an SBC out of 8 Elite and go from there. It would be a killer basis for a miniPC, especially as a competitor to some of those Alder Lake-N PCs. It'll probably cost too much to hit the same price points though.

 

[mvprod123]

https://twitter.com/x/status/1850599912167153910

https://twitter.com/x/status/1850614037597593678

i just can't...

The 8 Elite GPU is already 21mm2.

 

[Ghostsonplanets]

Kurnal published die shot of of 8 Elite

124 mm²
View attachment 109918

CPU core area (estimates)
Oryon-L = 2.1 mm²
Oryon-M = 0.9 mm²

I think

Snapdragon 8S Elite/Plus:
2 Phoenix L @ 3.x GHz + 6 Phoenix M @ 2.xGHz
1024 ALUs Adreno 825
8MB L2 cache + 8MB L2 cache + 4MB SLC + 4MB GMEM

Target = ~100mm²


Snapdragon 7 Plus:
2x Phoenix M @ 3.x GHz + 6x Phoenix M @ 2.x GHz
768 ALUs Adreno 820
8MB L2 cache + 2MB SLC + 2MB GMEM

Target = ~70 - ~80mm²

 

[FlameTail]

https://twitter.com/x/status/1850599912167153910

https://twitter.com/x/status/1850614037597593678

i just can't...

The 8 Elite GPU is already 21mm2.

The problem with overclocking the GPU by 2x is that it significantly degrades the efficiency.

The 2.2x (680 MHz -> 1.5 GHz) overclock on the 8 Gen 2 GPU to make the X Elite GPU, caused the power to increase by 4x (6W -> 25W)

So we can assume a 2x overclock (1100 MHz -> 2.2 GHz) of the 8 Elite GPU would also cause the power to increase by 4x. In fact, the power consumption increase may be even higher by a 5x or 6x factor, because now the baseline is much higher (1.1 GHz).

I'll speculate 3 configurations for the X Elite Gen 2's GPU.

Config A	12 CU	2200 MHz	22 mm²
Config B	16 CU	1650 MHz	29 mm²
Config C	24 CU	1100 MHz	44 mm²

Config A has the same 12 CU GPU as the 8 Elite, but running at 2x the clock speed.

Config B has 33% more CUs than 8 Elite, and runs at 1.5x the clock.

Config C has 2x the CUs compared to 8 Elite, but runs at the same clock speed as it.

Now let me pull up the grand 3DMark Steel Nomad Light table!

All score and power numbers are from Geekerwan except for;

(1) Ones with asterisks are estimates
(2) Ones that are italicised are speculation

PPW = Performance Per Watt
PPA = Performance Per Area

SoC	GPU Clock​	3DMark Steel Nomad Light Score​	3DMark Steel Nomad Light Power​	PPW​
Lunar Lake	2 GHz	3300	35W	94.28
Strix Point	2.9 GHz	3200	40W	80
M3	1.34 GHz	3500	25W	140.0
M4	1.47 GHz	3850	30W	128.3
8 Gen 2	680 MHz	1100	6W	183.33
8 Gen 3	903 MHz	1700	8W	212.5
8 Gen 4 [LP5X- 8533]	1100 MHz	2600	11W	236.36
8 Gen 4 [LP5X- 10700]	1100 MHz	2700	11W	245.45
X Elite	1.5 GHz	2400	25W*	96.00*
Config A	2.2 GHz	5400	55W*	98.18*
Config B	1.65 GHz	5400	33W*	163.63*
Config C	1.1 GHz	5400	22W*	245.45*
Apple M5		5000	35W	142.85

I guess Apple M5 will score about 5000 points in SNL, a 30% uplift from the M4. The Apple M4 brought only a 10% uplift over the M3, so I think it's reasonable to expect a larger uplift from M5.

Back to the hypothetical X Elite Gen 2 GPU configs;

A has the best PPA but the worst PPW. If 8 Elite Gen 2 had this GPU, it would be like the C Elite GPU situation all over again. It won't be competitive against Apple M5, in terms of efficiency.

C has the best PPW but the worst PPA. This make for a stellar laptop GPU, but the area would be rather large. So we know Qualcomm is unlikely to use something like this.

B offers the best balance of PPW and PPA, and hence I think it would be the most suitable configuration.

 

[GTracing]

PPA = Performance Per Area

FYI, the acronym PPA already stands for Power-Performance-Area. It's probably best not to use it for something so similar unless you're looking to confuse people.

 

[Magio]

The problem with overclocking the GPU by 2x is that it significantly degrades the efficiency.

The 2.2x (680 MHz -> 1.5 GHz) overclock on the 8 Gen 2 GPU to make the X Elite GPU, caused the power to increase by 4x (6W -> 25W)

So we can assume a 2x overclock (1100 MHz -> 2.2 GHz) of the 8 Elite GPU would also cause the power to increase by 4x. In fact, the power consumption increase may be even higher by a 5x or 6x factor, because now the baseline is much higher (1.1 GHz).

I'll speculate 3 configurations for the X Elite Gen 2's GPU.

Config A	12 CU	2200 MHz	22 mm²
Config B	16 CU	1650 MHz	29 mm²
Config C	24 CU	1100 MHz	44 mm²

Config A has the same 12 CU GPU as the 8 Elite, but running at 2x the clock speed.

Config B has 33% more CUs than 8 Elite, and runs at 1.5x the clock.

Config C has 2x the CUs compared to 8 Elite, but runs at the same clock speed as it.

Now let me pull up the grand 3DMark Steel Nomad Light table!

All score and power numbers are from Geekerwan except for;

(1) Ones with asterisks are estimates
(2) Ones that are italicised are speculation

PPW = Performance Per Watt
PPA = Performance Per Area

SoC	GPU Clock​	3DMark Steel Nomad Light Score​	3DMark Steel Nomad Light Power​	PPW​
Lunar Lake	2 GHz	3300	35W	94.28
Strix Point	2.9 GHz	3200	40W	80
M3	1.34 GHz	3500	25W	140.0
M4	1.47 GHz	3850	30W	128.3
8 Gen 2	680 MHz	1100	6W	183.33
8 Gen 3	903 MHz	1700	8W	212.5
8 Gen 4 [LP5X- 8533]	1100 MHz	2600	11W	236.36
8 Gen 4 [LP5X- 10700]	1100 MHz	2700	11W	245.45
X Elite	1.5 GHz	2400	25W*	96.00*
Config A	2.2 GHz	5400	55W*	98.18*
Config B	1.65 GHz	5400	33W*	163.63*
Config C	1.1 GHz	5400	22W*	245.45*
Apple M5		5000	35W	142.85

I guess Apple M5 will score about 5000 points in SNL, a 30% uplift from the M4. The Apple M4 brought only a 10% uplift over the M3, so I think it's reasonable to expect a larger uplift from M5.

Back to the hypothetical X Elite Gen 2 GPU configs;

A has the best PPA but the worst PPW. If 8 Elite Gen 2 had this GPU, it would be like the C Elite GPU situation all over again. It won't be competitive against Apple M5, in terms of efficiency.

C has the best PPW but the worst PPA. This make for a stellar laptop GPU, but the area would be rather large. So we know Qualcomm is unlikely to use something like this.

B offers the best balance of PPW and PPA, and hence I think it would be the most suitable configuration.

Another possibility is they have more than one configuration and that X Elite Gen 2 will have both an M4/5 and an M4/5 Pro competitor. The leaked Dell roadmap mentioned an "Oryon V2" and "Oryon V2 UI", the latter of which targeting a 40W TDP and shipping in the next gen XPS 14 which will also feature an Intel PTL-P SKU with 12 Xe3 cores. And the year after has Oryon V3 appearing in an 80W flavor in the XPS 16 which could be a Max competitor.

Maybe Oryon V2 UI is just a higher TDP version of Oryon V2, I'm not actually sure what the rumors are on that front yet, but if instead its another chip then your option C with higher clocks would fit right in there.

 

[ikjadoon]

The Court docket shows a hearing this Wednesday (30 October); this might be the earlier motions re: expunging certain expert testimony and / or summary judgments. Though, perhaps something fresh regarding the recent Qualcomm ALA cancellation notice.

Unfortunately, it often takes weeks to get details and PACER doesn't have any documents (yet).

 

[The Hardcard]

Another possibility is they have more than one configuration and that X Elite Gen 2 will have both an M4/5 and an M4/5 Pro competitor. The leaked Dell roadmap mentioned an "Oryon V2" and "Oryon V2 UI", the latter of which targeting a 40W TDP and shipping in the next gen XPS 14 which will also feature an Intel PTL-P SKU with 12 Xe3 cores. And the year after has Oryon V3 appearing in an 80W flavor in the XPS 16 which could be a Max competitor.

Maybe Oryon V2 UI is just a higher TDP version of Oryon V2, I'm not actually sure what the rumors are on that front yet, but if instead its another chip then your option C with higher clocks would fit right in there.

Given the 12 full Oryon cores, I see the X Elite as M3 Pro competition. Whatever the V2 replacement for that chip is, beating the base M GPU is not a worthy goal if they plan to not support discrete GPUs. They are cutting off a large section of the market that would appreciate the multithreaded performance if it only has base thin and light GPU power.