Discussion Qualcomm Snapdragon Thread

[poke01]

View attachment 110251
ROFL.

Source

Now it’s Qualcommbench.

 

[FlameTail]

Rumour [DCS];

4.57 GHz.

That's bonkers. It's 0.25 GHz (or +5%) higher than the standard version of 8 Elite (4.32 GHz).

This one would be able to solidly claim the GB6 single-core crown from Apple A18 Pro in Geekbench 6.2 (no SME).

Also 4.57 GHz is just 0.43 GHz away from the 5 GHz mark. Even if the standard version of 8 Elite G2 is not able to hit 5 GHz, I think the high-clock version of 8 Elite G2 will.

 

[FlameTail]

On the other hand if Andrei wanted to do accurate uarch vs uarch comparisons he should have enabled AVX for x64 when doing comparisons on Qualcomm slides as the AVX units belong to the core the same Neon units do (I guess he is somewhat involved with those if the rumours are accurate). But yea, that is a digression

He replies;

The LNL binaries were fully AVX2 and what not enabled, the attendees at summit had full access to the complete build config and see it first-hand

 

[FlameTail]

TechTechPotato podcast about Qualcomm v ARM;

 

[MS_AT]

He replies;

That well might be but then they should improve their messaging as

only lists LLVM (clang) 17.0.6 with -O3 and -flto. Link time optimizations are indifferent to the point we are discussing so it leaves us with O3 only.

https://godbolt.org/z/rnPvW8jnc this is very trivial loop so to ensure compiler will auto vectorize it, you can see that -O3 option is doing vectorization using only xmm registers and does not use fma operations compared to the assembly for the -O3 -mavx512f on the right that is using zmm registers (512b width, avx512 only) and fma operations. [Yes it's using clang 17.0.1 instead of 17.0.6 but this behaviour won't change]

In other words based on the information they made public, one can assume "AVX2 and what not" was not enabled. So for the future they could simply put the link to the full compiler options so they would not bloat the slide if they were afraid about that and nobody would say then these were not enabled.

 

[mvprod123]

Rumour [DCS];
View attachment 110264
4.57 GHz.

That's bonkers. It's 0.25 GHz (or +5%) higher than the standard version of 8 Elite (4.32 GHz).

This one would be able to solidly claim the GB6 single-core crown from Apple A18 Pro in Geekbench 6.2 (no SME).

Also 4.57 GHz is just 0.43 GHz away from the 5 GHz mark. Even if the standard version of 8 Elite G2 is not able to hit 5 GHz, I think the high-clock version of 8 Elite G2 will.

To reach the A18 Pro level, you must overclock the Oryon core as much as possible and provide active cooling🤣. This is reminiscent of Geekerwan's recent Ryzen 9900X review, where the chip was overclocked under nitrogen to beat M4 in SPEC tests. That's ridiculous.

 

[MS_AT]

To reach the A18 Pro level, you must overclock the Oryon core as much as possible and provide active cooling🤣. This is reminiscent of Geekerwan's recent Ryzen 9900X review, where the chip was overclocked under nitrogen to beat M4 in SPEC tests. That's ridiculous.

Well, he could have matched M4 in SPEC(at least fp) without nitrogen by flipping few compiler options but the entertainment value would be lower, wouldn't it?

 

[FlameTail]

To reach the A18 Pro level, you must overclock the Oryon core as much as possible and provide active cooling🤣. This is reminiscent of Geekerwan's recent Ryzen 9900X review, where the chip was overclocked under nitrogen to beat M4 in SPEC tests. That's ridiculous.

Tbh, the way he phrased it, active cooling is optional.

If you look at the ST power curves, the 4.32 GHz 8 Elite actually consumes less power than the Dimensity 9400. We know that Dimensity 9400 phones don't need active cooling.

A 5% overclock would increase the power by 10-20%

But it could even be less than that if they are using chip binning. Bin the high quality chips and push them to 4.57 GHz, while the rest hum along at 4.32 GHz.

 

[mvprod123]

Tbh, the way he phrased it, active cooling is optional.
View attachment 110270
If you look at the ST power curves, the 4.32 GHz 8 Elite actually consumes less power than the Dimensity 9400. We know that Dimensity 9400 phones don't need active cooling.

A 5% overclock would increase the power by 10-20%

But it could even be less than that if they are using chip binning. Bin the high quality chips and push them to 4.57 GHz, while the rest hum along at 4.32 GHz.

This is still called creating special conditions.

 

[Nothingness]

That well might be but then they should improve their messaging as
View attachment 110269
only lists LLVM (clang) 17.0.6 with -O3 and -flto. Link time optimizations are indifferent to the point we are discussing so it leaves us with O3 only.

https://godbolt.org/z/rnPvW8jnc this is very trivial loop so to ensure compiler will auto vectorize it, you can see that -O3 option is doing vectorization using only xmm registers and does not use fma operations compared to the assembly for the -O3 -mavx512f on the right that is using zmm registers (512b width, avx512 only) and fma operations. [Yes it's using clang 17.0.1 instead of 17.0.6 but this behaviour won't change]

In other words based on the information they made public, one can assume "AVX2 and what not" was not enabled. So for the future they could simply put the link to the full compiler options so they would not bloat the slide if they were afraid about that and nobody would say then these were not enabled.

I agree.

OTOH the difference should not be massive. I don't expect a significant amount of vectorized code in SPECint 2017 except for x264.

 

[Meteor Late]

Tbh, the way he phrased it, active cooling is optional.
View attachment 110270
If you look at the ST power curves, the 4.32 GHz 8 Elite actually consumes less power than the Dimensity 9400. We know that Dimensity 9400 phones don't need active cooling.

A 5% overclock would increase the power by 10-20%

But it could even be less than that if they are using chip binning. Bin the high quality chips and push them to 4.57 GHz, while the rest hum along at 4.32 GHz.

I think power would be increased more than that, look at the shape of the curve, for example in the fp test, going from 6 to 8W (33% increase) nets you 5%, maybe 6%? more performance. It's only getting worse if you overclock 5%.
Of course, with binning it's another story, but as it stands, a 5% overclock with the same SoC would be very costly in power consumption.

 

[GTracing]

TechTechPotato podcast about Qualcomm v ARM;

Thanks for sharing. Dr Ian sounds pretty pessimistic about Qualcomm's position in the lawsuit (36:50). I was assuming they'd reach a settlement, but the way he puts it, if this does go to court, Qualcomm needs to win.

Can ARM lose Qualcomm? Yes. Can Qualcomm lose ARM? No.

It's also interesting when they talked about how Qualcomm was required to "destroy" any IP that was created under Nuvia before the acquisition (26:16).

ARM sent notice to Qualcomm that the Nuvia ALA was no longer in effect. The "ALA" being the architecture license. And that all IP made under that ALA had to be destroyed. Now the thing is you can't destroy memories. And I don't mean DRAM. I mean in your head. Right, there are lessons learned. And Qualcomm's position from that point on was essentially to say anything the Nuvia team are doing is denuvo brand new from acquisition. Which was the smart messaging to say I'm pretty sure.

So now if you go dig into the Pacer which is Public Access for court electronic records. There's a few re heavily redacted documents that point to evidence that ARM has that Qualcomm did not actually destroy the IP, regardless of memory or not. There's a fine line between destroying something, and not using it. So and in this case the remedy of the lawsuit is to destroy everything, but also everything derived yes so that would be not just X Elite and 8 Elite but anything after as well.

I'm paraphrasing these quotes btw.

 

[Raqia]

Thanks for sharing. Dr Ian sounds pretty pessimistic about Qualcomm's position in the lawsuit (36:50). I was assuming they'd reach a settlement, but the way he puts it, if this does go to court, Qualcomm needs to win.


It's also interesting when they talked about how Qualcomm was required to "destroy" any IP that was created under Nuvia before the acquisition (26:16).

I'm paraphrasing these quotes btw.

The cure for these contractual disputes is usually monetary considerations.

Even if ARM outright refuses to extend an ALA to Qualcomm, it still has the TLA in place and there is no dispute regarding that. Indeed there'd be a scramble to tape out a replacement at Qualcomm, but they would go on. It also is a sign of bad faith on the part of ARM that they're resorting to blunt force tactics like the lawsuit and contract cancellation instead of more neutral avenues like arbitration. Apple is the other main player to make serious use of the ALA; if this goes to trial, we'll see from discovery if they were un-harassed and if ARM is just trying to squeeze Qualcomm back into the TLA...

 

[FlameTail]

About 70 percent of AI in today’s third-party applications on the smartphone run on the device’s CPU — which includes those where there is a dedicated accelerator available (such as an NPU and GPU). The reasons: CPUs are ubiquitous, flexible, and easy to use for software developers, meaning they can target as many devices as possible with their applications.

Why AI starts on the CPU

The rise of AI inference means more AI workloads are being processed at the edge.

www.theverge.com

That's interesting. Oryon Gen 3 introducing SME (and SVE?), would be actually meaningful to the user experience, not just for winning benchmarks.

 

[FlameTail]

Even if ARM outright refuses to extend an ALA to Qualcomm, it still has the TLA in place and there is no dispute regarding that. Indeed there'd be a scramble to tape out a replacement at Qualcomm, but they would go on.

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.

It also is a sign of bad faith on the part of ARM that they're resorting to blunt force tactics like the lawsuit and contract cancellation instead of more neutral avenues like arbitration.

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

 

[FlameTail]

Revisiting this slide;


The efficiency uplift from Oryon (X Elite) to Oryon-L (8 Elite) is truly wild.

~17W to ~8W.

A 2x efficiency gain in 1 year/1 generation!

Yeah I know it's a Qualcomm slide, but testing by Geekerwan does corroborate those numbers;

	X Elite	8 Elite
SPEC2017 INT	15W	6.5W
SPEC2017 FP	20W	8W

 

[adroc_thurston]

Yeah I know it's a Qualcomm slide, but testing by Geekerwan does corroborate those numbers;

Chefe you're comparing platform power numbers between a phone and a 13" laptop. Do NOT do that.

 

[Doug S]

Apple is the other main player to make serious use of the ALA; if this goes to trial, we'll see from discovery if they were un-harassed and if ARM is just trying to squeeze Qualcomm back into the TLA...

Apple would be un-harassed because they didn't acquire another company with ARM IP developed under a separate ALA that has different terms.

Apple also last year signed a new/extended ALA that covers them through 2040 or so. Who knows, that new ALA may have significantly tighter restrictions around transfer of ARM technology under other licenses via acquisition. Contracts often evolve over time when renewed to cover "lessons learned" (if you've never been a landlord, ask one, they'll have examples of stuff they added to their lease because of things they've learned the hard way)

 

[Raqia]

Apple would be un-harassed because they didn't acquire another company with ARM IP developed under a separate ALA that has different terms.

Apple also last year signed a new/extended ALA that covers them through 2040 or so. Who knows, that new ALA may have significantly tighter restrictions around transfer of ARM technology under other licenses via acquisition. Contracts often evolve over time when renewed to cover "lessons learned" (if you've never been a landlord, ask one, they'll have examples of stuff they added to their lease because of things they've learned the hard way)

We'll see. Qualcomm's CPU IP is now different from the IP that ARM is disputing: after joining Qualcomm and being confronted by ARM's initial complaint, the Nuvia team (from what I have learned from court documents and sources) has allegedly redesigned the relevant blocks in the Oryon design to skirt the matter. Again, ARM's further request for the destruction of the IP isn't a typical cure for these situations...

 

[Panino Manino]

This new chips looks like the real deal.
Huge performance increase with huge energy decrease.

 

[FlameTail]

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

 

[Meteor Late]

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

Do we know if X Elite can sustain 3.8GHz in Cinebench 2024? one thing is announced all core clock speed, another thing is if it's possible to sustain that in all workloads.

 

[FlameTail]

Do we know if X Elite can sustain 3.8GHz in Cinebench 2024? one thing is announced all core clock speed, another thing is if it's possible to sustain that in all workloads.

If that is indeed the case, it's rather terrifying.

X Elite needs 100 Watts to achieve that 1200 score in CB2024 (according to Jeff Geerling).

How much more power can it drink?

 

[mvprod123]

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

M3 Pro/Max can achieve 3200+ pt in ST

 

[poke01]

The X Elite CPU is not very good, its haphazardly put together. The contrast between X elite and 8 elite is stark.