Samsung outs Exynos 9 Series 9810

[Lodix]

It should be all cores, the power is more or less the same as the 835.

Yeah, I thought so. If the power per core is around the same and the multi score of Geekbench increases similar to the single core... must be it. Great then

 

[Nothingness]

The Geekbench results seem to come from the public score posted in December: http://browser.geekbench.com/v4/cpu/5548051

@Andrei. didn't you have time to run GB4 yourself? Or is that result from you (though it says it's a Samsung platform)?

 

[DeletedMember377562]

XDA-Developers have tested the SD845 as well: https://www.xda-developers.com/qualcomm-snapdragon-845-hands-on-benchmarks-first-impressions/

And they come to the summary that:

The single-core score sees an average increase of 25%, while the multi-core score sees a smaller uplift of 24%. Those figures are around the expected improvements of 25% to 30%.

Andrei is gonna make fun of XDA as well now, I guess. Clearly, they don't "comprehend" benchmarking like Andrei does, and haven't taken "the many implications and factors at play" when dealing with CPU performance improvements...

Forget about the fact that the general performance improvements in benchmarks show 25% increase. Forget that Qualcomm themselves said 25% improvement. Clearly, Andrei knows better than everyone else; even Qualcomm. The performance improvements are what we see in single incidents from Andrei's article...

 

[Andrei.]

The Geekbench results seem to come from the public score posted in December: http://browser.geekbench.com/v4/cpu/5548051

@Andrei. didn't you have time to run GB4 yourself? Or is that result from you (though it says it's a Samsung platform)?

Ryan was at the event but didn't share the detailed scores before flying back and I was writing it up so I looked up the scores from that submission. We verified them to what we ran at the event afterwards and they were within margin of error +- 0.5% and never changed them to that run in the table.

So yeah somebody ran final performance figures back in December. If we had known.

 

[dawheat]

Great article on the SD845 andrei and appreciate the level of detail. Will be very interesting to compare it to the next Exynos as it seems like they'll be different enough to actually be interesting.

Generalako - will you give it a rest? Andrei's explained his reasoning and it's logically sound, even though it proved to be wrong. You've made your point - repeating it 20 times doesn't add value to this thread.

 

[DeletedMember377562]

Repeating does seem necessary, as Andrei refuses to acknowledge he was wrong. I'll keep repeating it until he actually admits he made a mistake -- rather than trying to justify his reasoning back then (which when I questioned him for, he confidently stepped on me for, acting as if he knew what he was talking about). The way I see it the recent XDA test, which uses a long variety of benchmark tests, is the final nail in the coffin in terms of establishing Andrei as an inept and unreliable analyst.

 

[Thala]

I didn't speculate anything actually. I answered your original comment where you pretty clearly stated, as if it were facts, that "[ARM] will introduce the successor to A75 soon", which will bring so and so performance improvement. I answered you back saying none of this was true as there were no basis to make such claims. Ever since then you have still yet to provide me any sources to back up your claims (like any slides or roadmaps from ARM), so that we can possibly look at it and discuss.

I did state a fact, because i do know the roadmap, which apparently is not publicly disclosed yet. Otherwise i would have marked it as speculation. It is that simple. You can take this information or ignore it and continue tapping in the dark.

 

[DeletedMember377562]

If something is a fact, you have to provide actual sources for it. Claiming it "is not publicly disclosed yet" is not a good enough (and also worrisome on your part, as something not being publicly disclosed means they want to keep it under wraps; why are you going around talking about it and its contents?) excuse. I might as well claim that the next ARM core will decrease it's performance by 50% and that this is a fact based on their roadmap....but alas, I can't give any sources since it's not publicly disclosed yet. It is that simple.

You can take this information or ignore it and continue tapping in the dark.

Not any information, really. Just a claim. Unless I actually see the facts for myself, I have no rational reason to believe you. So until ARM actually release the roadmap, so that I can look at it, there's nothing more to be said about the matter.

 

[Thala]

One more comment regarding Geekbench. The single core score is the weighted average of crypto, integer, floating point and memory score. The biggest issues with this the memory score and partially the crypto score. It is inherently wrong, when evaluating an architecture, to penalize due to lower memory bandwidth - which is the case for Snapdragon 845 compared to Snapdragon 835.
That`s like comparing two GPU architectures A and B, and while A wins in benchmarks it has lower memory bandwidth. So A has apparently better memory efficiency, but still in the overall score it is penalized due to lower memory bandwidth - which everyone would agree makes no sense. That's like giving the worse architecture a few extra points due to higher memory bandwidth requirements.

In conclusion, the overall Geekbench score is skewed due to the inclusion of memory bandwidth. Instead the individual integer and floating point score should be taken, when comparing architectures.

Now looking at the numbers from XDA developer, we are seeing an increase of 33% for integer and 44% for floating point workloads for the Snapdragon 845 compared to Snapdragon 835. The quoted 25% are extremely misleading due inclusion of memory score.

 

[DeletedMember377562]

It isn't misleading. We are looking at the actual performance of the SoC. Not of the specific core architecture itself. Memory bandwidth isn't something I can decide to ignore or disclude when using a device, now is it? It's part of the whole picture.

 

[Andrei.]

Now looking at the numbers from XDA developer, we are seeing an increase of 33% for integer and 44% for floating point workloads for the Snapdragon 845 compared to Snapdragon 835. The quoted 25% are extremely misleading due inclusion of memory score.

And this is why AT doesn't post the aggregate GB4 score anywhere and solely focus on the subtests. I also had confirmation that the GB4 memory scores just don't scale after a certain point. The DSU should be able to saturate the 30GB/s of the controllers but that's not reflected here.

It isn't misleading. We are looking at the actual performance of the SoC. Not of the specific core architecture itself. Memory bandwidth isn't something I can decide to ignore or disclude when using a device, now is it? It's part of the whole picture.

You can't do anything with memory bandwidth alone. Its effects are already reflected in the performance of the subtests.

 

[Thala]

If something is a fact, you have to provide actual sources for it.

Lol, and this is because you say so? Provided evidence is no prerequisite/necessary condition for the truth of a statement. You statement is irrational.

I might as well claim that the next ARM core will decrease it's performance by 50%

Of course you could if you want to lie by intention. In addition you provided enough evidence that you are clueless with respect to ARMs roadmap.

You can't do anything with memory bandwidth alone. Its effects are already reflected in the performance of the subtests.

Precisely!

 

[Thala]

The DSU should be able to saturate the 30GB/s of the controllers but that's not reflected here.You can't do anything with memory bandwidth alone. Its effects are already reflected in the performance of the subtests.

Indeed. DSU can be configured to have either 2x128bit ACE interfaces or 1x256bit CHI master interface. Since both are based on AMBA AXI specification, they have independent write and read data channels, which essentially doubles combined read and write bandwidth.

 

[Lodix]

Please don't reply to him...

First benchmark of the Exynos 9810, it is rumoured to be running at 2'6GHZ at single core.

http://browser.geekbench.com/v4/cpu/6974531

 

[HurleyBird]

Repeating does seem necessary

Dude, learn social queues. This is a dumb hill to die on.

 

[Andrei.]

Please don't reply to him...

First benchmark of the Exynos 9810, it is rumoured to be running at 2'6GHZ at single core.

http://browser.geekbench.com/v4/cpu/6974531

Edit: Edit: Stand by....

Yep these seem to be real at first glance. Single core 2.6 and multi 1.9GHz.

 

[Lodix]

Edit: Edit: Stand by....

Yep these seem to be real at first glance. Single core 2.6 and multi 1.9GHz.

Hahaha I hope what you said is true.

 

[Andrei.]

So as expected still a bit short of the A11 but well in range of the A10. IPC still a tad lower than Apple.

But the FP improvements are gigantic, well over 2x! Web browsing workloads should get massive boosts as expected from the huge FP execution back-end.

 

[R0H1T]

Here's the top 2 (MT) scores comparing 845 vs 9810 ~ samsung SM-G960F vs QUALCOMM SDM845 for arm64 - Geekbench Browser

 

[Lodix]

So as expected still a bit short of the A11 but well in range of the A10. IPC still a tad lower than Apple.

But the FP improvements are gigantic, well over 2x! Web browsing workloads should get massive boosts as expected from the huge FP execution back-end.

If the IPC is around same between the A10/A11... what is the clock speed of the latter?

 

[CatMerc]

Race To Idle: The SoC

 

[Andrei.]

If the IPC is around same between the A10/A11... what is the clock speed of the latter?

2.34 and 2.38

 

[Lodix]

2.34 and 2.74ish

Geekbench keeps reporting the wrong frequency then :S

Thanks

 

[Andrei.]

Geekbench keeps reporting the wrong frequency then :S

Thanks

No it's correct, I messed up. I don't remember where I got 2.74 from.... It's 2.38.

Corrected the image accordingly.

 

[Thala]

Interesting scores, in particular score per MHz. Where are the latest Intel cores sitting at? Should be around 1.3 score per MHz for integer right?