It isn't integrated anymore? Interesting decision I guess. I wonder if FC7900 is bundled with Snap 8E or OEMs can choose one but not the other.This confirms that the Fastconnect 7900 isn't integrated into the 8 Elite SoC (which is made on 3nm), but that it's a discrete chip.
That make sense because IMR is very bandwidth demanding and was abandoned more than a decade ago by Nvidia and AMD.Does Adreno 830 exclusively use Binned Direct mode?
View attachment 110196
Vince suggests that Adreno 830 could be using TBIM (Tile Based Immediate Mode Rendering), which is a hybrid of the tradional mobile TBDR (Tile Based Deferred Rendering) and traditional desktop IMR (Immediate Mode Rendering).
Fastconnect chips were always discrete;It isn't integrated anymore? Interesting decision I guess. I wonder if FC7900 is bundled with Snap 8E or OEMs can choose one but not the other.
The Qualcomm® FastConnect™ 7800 Mobile Connectivity System is an advanced 14nm Wi-Fi and Bluetooth®....
The FastConnect 6900 is an advanced. 14nm 2x2 Mobile Connectivity system
A18-P is ~18% faster than Oryon-L.SPEC numbers extracted from Geekerwan's graphs.
A18-P A18-E Oryon-L Oryon-M Clock speed 4.04 GHz 2.2 GHz 4.32 GHz 3.53 GHz SPEC INT 10.7 3.3 8.9 5.2 SPEC FP 16.0 5.0 14.0 8.0 Core size 3.0 mm² 0.8 mm² 2.2 mm² 0.9 mm²
Weird. Your own line of reasoning shows it performing disproportionately better. Not a long enough context window?Conclusion: Apple CPUs are not disproportionately faster in Geekbench 6 Multi-core.
Snapdragon 8 Elite
2 × Phoenix-L @ 4.32 GHz [12 MB sL2]
6 × Phoenix-M @ 3.53 GHz [12 MB sL2]
Snapdragon 7 Gen 4 (Speculation)
2 × Phoenix-L @ 3.9 GHz [8 MB sL2]
2 × Phoenix-M @ 3.2 GHz [4 MB sL2]
Will a quad core CPU be unsuitable for Android today?
Even $150 phones have '8-core CPU' nowadays.
Edit: The figure I used for Oryon-M was wrong. It's 60% of Oryon-L, not 40% of Oryon-L. This changes my conclusion.Weird. Your own line of reasoning shows it performing disproportionately better. Not a long enough context window?
Hmm, I wonder if you should compare to SPEC nt instead of extrapolating from 1t. Does anyone test that?Edit: The figure I used for Oryon-M was wrong. It's 60% of Oryon-L, not 40% of Oryon-L. This changes my conclusion.
You can read my analysis again. Sorry for the error.
Unless they have a Small Phoenix core, that does not make sense. I expect 2 Phoenix L and 6 M but at much lower clock and with the name of Snapdragon 7 Plus.Snapdragon 8 Elite
2 × Phoenix-L @ 4.32 GHz [12 MB sL2]
6 × Phoenix-M @ 3.53 GHz [12 MB sL2]
Snapdragon 7 Gen 4 (Speculation)
2 × Phoenix-L @ 3.9 GHz [8 MB sL2]
2 × Phoenix-M @ 3.2 GHz [4 MB sL2]
Will a quad core CPU be unsuitable for Android today?
Even $150 phones have '8-core CPU' nowadays.
I think AMD would be happy to negotiate a better deal than ARM in the event of an Intel acquisition to keep their bread and butter going; doubt they'll want to pivot to ARM given their new found zeal for licensing revenues at any cost: ARM's architectural license is already in place w/ Qualcomm, so this current, rather weak law suit was their best bet out of few to re-negotiate pricing.Yeah Qualcomm isn’t getting an x86 license unless something very very crazy happens
GB6's "MT" is not MT and as such shouldn't be used for multi core performance comparisons like you attempted there.8 Elite is only about 15% faster on average than A18 Pro in GB6 MT.
Why? How? What's the reason for thus disparity?
My question wasn't "How much faster actually is 8 Elite than A18 Pro in MT?"GB6's "MT" is not MT and as such shouldn't be used for multi core performance comparisons like you attempted there.
Snapdragon 8 EliteUnless they have a Small Phoenix core, that does not make sense. I expect 2 Phoenix L and 6 M but at much lower clock and with the name of Snapdragon 7 Plus.
GB6 doesn't scale linearly with more cores. You can compare other CPUs and see similar results. For example there's only a 20% difference between the 8 core 16 thread 1700X vs 4 core 8 thread 1500X. https://browser.geekbench.com/v6/cpu/compare/8462630?baseline=8458189My question wasn't "How much faster actually is 8 Elite than A18 Pro in MT?"
I am asking "Why does A18 Pro (and previous Apple CPUs) perform so well in the Geekbench 6 Multicore test?"
What quality is there in those CPUs that make them perform disproportionately better? Or is Apple handing money under the table to Primate Labs, to boost the scores of their SoCs in Geekbench 6? Thus making the 'Geekbench 6 = Applebench' theory a reality (I wasn't a fan of the theory in the first place).
Yeah but some of the tests almost do scale linearly (in particular the rendering one). Not all, far from it, and that's why the global score can be misleading.GB6 doesn't scale linearly with more cores. You can compare other CPUs and see similar results. For example there's only a 20% difference between the 8 core 16 thread 1700X vs 4 core 8 thread 1500X. https://browser.geekbench.com/v6/cpu/compare/8462630?baseline=8458189
I want to revisit the topic of Apple SoCs performing disproportionately better in Geekbench 6 Multi Core test.
Now that Geekerwan has published his review of Snapdragon 8 Elite.
We can compare A18 Pro and 8 Elite.
A18-P is ~18% faster than Oryon-L.
Oryon-M is ~60% faster than A18-E.
Oryon-M has ~40% of the performance of Oryon-L. A18-E has ~27% of the performance of A18-P.
I will take Oryon-L as the 100% baseline.
A18-P = 118%
A18-E = (118 × 27%) = 32%
Oryon-L = 100%
Oryon-M = (100 × 60%) = 60%
A18 Pro
= 2P + 4E
= 2 (118%) + 4 (32%)
= 364%
8 Elite
= 2L + 6M
= 2 (100%) + 6 (60%)
= 560%
So theoretically MT performance of 8 Elite should be 53% higher.
Geekbench 6.3 Multi
A18 Pro = 9000 points
8 Elite = 10500 points
8 Elite is only about 15% faster on average than A18 Pro in GB6 MT.
Why? How? What's the reason for thus disparity?
A18 Pro does have SME, and Geekbench 6.3 boosts the score for SME. But this boost is only about 10% in ST. And consider that there is one SME block per cluster, so SME performance doesn't linearly scale up with core count. That means the MT boost by SME of A18 Pro to Geekbench 6 Multicore test would be <10%.
This is nothing new. Ever since Geekbench 6 released, Apple A CPUs have been performing disproportionately better in it than their Android series counterparts. So much so, that there is a running joke that Geekbench 6 = Applebench.
But we are hardware enthusiasts, and I want to get to the truth of this.
Back in the day, I thought that one reason for the disparity might be Apple's superior 2-tier (L1/L2) cache hierarchy. Whereas Qualcomm used ARM's inferior 3-tier (L1/L2/L3) hierarchy. But now Snapdragon 8 Elite also employs a similar cache subsystem as Apple. Snapdragon 8 Elite even has more cache than Apple A18 Pro (24 MB L2 vs 20 MB L2).
I want answers.
Edit: The figure I used for Oryon-M was wrong. It's 60% of Oryon-L, not 40% of Oryon-L. This changes my conclusion.
Edit2 : Grammar and Formatting.
Since Qualcom will have problems with ARM I expect this...Snapdragon 8 Elite
2 × Phoenix-L @ 4.32 GHz [12 MB L2]
6 × Phoenix-M @ 3.53 GHz [12 MB L2]
Snapdragon 8s Gen 4 / 7+ Gen 4 (Speculation)
2 × Phoenix-L @ ~4.0 GHz [8 MB L2]
6 × Phoenix-M @ ~2.5 GHz [8 MB L2]
Snapdragon 7 Gen 4 (Speculation)
2 × Phoenix-M @ ~3.5 GHz [6 MB L2]
6 × Phoenix-M @ ~2.5 GHz [6 MB L2]
Using Phoenix-M as the prime core in 7 Gen 4 wouldn't be bad.
We know from Geekerwan's 8 Elite review that Phoenix-M has 60% of the peak performance of Phoenix-L.
Phoenix-L @ 4.3 GHz scores ~3250 in Geekbench 6 ST. So Phoenix-M @ 3.5 GHz would score about 2000 points.
Now the 7 Gen 4 configuration that I speculated above has half the L2 cache, so let's deduct 10%.
That gives 1800 points in GB6 ST for Snapdragon 7 Gen 4 (Speculated).
That's quite respectable performance for a midrange SoC, and a large uplift from the Snapdragon 7 Gen 3 (~1200 points).
What Andrei said is raising an interesting dillema does the end user care about uarch comparison or the actual performance in context of Geekbench version (6.2 vs 6.3). I mean I understand that introducing SME will have less impact in android ecosystem due to fragmentation, but I guess that vertically integrated Apple is already making use of it due to having much better control over software stack.
At least this gives some context to Qualcomm slides where 6.2 was used, as from uarch vs uarch it's more fair seeing SME unlike Neon is not part of the core.
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
And yet people draw conclusions after seeing variations of a few percents.Benchmark results are always a snapshot in time. Remember the differences in GB results for Apple depending on which OS version is used, i.e. the regression with iOS 18.0 compared to developer betas during the summer, and the improvement people running 18.1 betas reported? I imagine you see the same sort of stuff with different Android versions. That kind of stuff swamps out small things like using jemalloc to get a better SPEC result.
You'll never get a "fair" comparison because next week an OS update could increase or decrease performance levels for one, a new compiler version might provide a boost in SPEC, and so forth.
Sure, that is why you won't be able to provide 100% fair comparison, but it doesn't mean you cannot strive to reach perfection I don't expect every one will suddenly start to use the same OS version and compiler settings to run SPEC comparisons, but it would be definitely nice if they could provide informations about environments they used. Bonus points for being consistent with what they do.Benchmark results are always a snapshot in time. Remember the differences in GB results for Apple depending on which OS version is used, i.e. the regression with iOS 18.0 compared to developer betas during the summer, and the improvement people running 18.1 betas reported? I imagine you see the same sort of stuff with different Android versions. That kind of stuff swamps out small things like using jemalloc to get a better SPEC result.
You'll never get a "fair" comparison because next week an OS update could increase or decrease performance levels for one, a new compiler version might provide a boost in SPEC, and so forth.
Side question, what is the threshold the improvement is deemed meaningful? 0.01%, 0.05% or everything is weighted against the cost and if 0.01% improvement is "free" it will go in anyway?I think I already explained how this is done during CPU development: extracts of benchmarks/apps or instruction traces are run through simulation (accurate model, RTL simulation, FPGA, emulator) in a completely controlled platform. That's the only way to measure things accurately when you're expecting improvements from a change in the range of 0.1 or 0.2% (yes, some tweaks only bring that).
Hmm. I think GB6 MT is best fit for phones and thin laptops. Where you're not running embarrassingly parallel workloads.GB6 isn’t the best to compare MT, plus in a phone you have a lot of variables.
Flame is talking about why it’s not scaling well, one reason could be not maintaining clocks and GB6 not meant for parallel MT.Hmm. I think GB6 MT is best fit for phones and thin laptops. Where you're not running embarrassingly parallel workloads.
It is less relevant for high end desktops and is entirely meaningless for servers.
After a few iterations of the same microarchitecure everything that brings performance with acceptable power and area is picked, because the uarch usually has reached a balanced point where more performance requires larger changes and these are done earlier in the design cycle. So, yes, if the cost is low, any improvement goes in.Side question, what is the threshold the improvement is deemed meaningful? 0.01%, 0.05% or everything is weighted against the cost and if 0.01% improvement is "free" it will go in anyway?