Samsung outs Exynos 9 Series 9810

[Andrei.]

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?

I can't filter out overclocked results from stock ones in the GB database, do we have some accurate results at a given frequency?

 

[Nothingness]

I can't filter out overclocked results from stock ones in the GB database, do we have some accurate results at a given frequency?

I guess you can use the results the author of Geekbench puts here: http://browser.geekbench.com/processor-benchmarks

Too bad he didn't provide the full results, but from the overall score you can see if one result seems to be overclocked or fake.

 

[el etro]

I still think at 2.9Ghz with this IPC on final product. Anyway is still a dead end, being tied to buy a Samsung or a Apple phone to have this kind of performance is far from what we want.

 

[Andrei.]

https://browser.geekbench.com/v4/cpu/7016241

If that's running at 4600MHz as it shoud then the IPC is 1.35 integer 1.34 FP, so just a bit above the M3 in integer and a bit below it in FP. But again this is a bit irrelevant given the clock frequencies. The A10 also lost some integer IPC vs the A9 but had a large frequency increase. Zen at 1.02 int and 1.04 FP.

At this point to make comparisons against desktop we should look at SPEC.

 

[Thala]

https://browser.geekbench.com/v4/cpu/7016241

If that's running at 4600MHz as it shoud then the IPC is 1.35 integer 1.34 FP, so just a bit above the M3 in integer and a bit below it in FP. But again this is a bit irrelevant given the clock frequencies. The A10 also lost some integer IPC vs the A9 but had a large frequency increase. Zen at 1.02 int and 1.04 FP.

At this point to make comparisons against desktop we should look at SPEC.

I assume that result is already overclocked as my Skylake 6700K achieves 5120 integer score @4100MHz, which is roughly 1.24 score per MHz. Not sure, why you think this is irrelevant and why you think that SPEC is better when comparing to desktop?

 

[asendra]

https://browser.geekbench.com/v4/cpu/7016241

If that's running at 4600MHz as it shoud then the IPC is 1.35 integer 1.34 FP, so just a bit above the M3 in integer and a bit below it in FP. But again this is a bit irrelevant given the clock frequencies. The A10 also lost some integer IPC vs the A9 but had a large frequency increase. Zen at 1.02 int and 1.04 FP.

At this point to make comparisons against desktop we should look at SPEC.

Well, I gave up on the iPad Pro / A10x review, thats for sure. Hell, we dont even have one for the new iPhones with the A11...

Specially after the original iPad Pro with the a9x review including SPEC results.... I dont understand why dont do it.

 

[Nothingness]

I assume that result is already overclocked as my Skylake 6700K achieves 5120 integer score @4100MHz, which is roughly 1.24 score per MHz.

Your result is lower than the "official" one of 5341 in the Primate Lab DB: https://browser.geekbench.com/processors/1705

The result Andrei showed is similar to the official one.

Not sure, why you think this is irrelevant and why you think that SPEC is better when comparing to desktop?

Though Geekbench 4 is not a toy benchmark, SPEC CPU 2006 and 2017 are heavier and more representative of desktop/workstation workloads.

 

[Thala]

Your result is lower than the "official" one of 5341 in the Primate Lab DB: https://browser.geekbench.com/processors/1705
The result Andrei showed is similar to the official one.

Issue is, that you are quoting single core result for 6700k while i was talking about the integer score, which is lower. Second the one Andrei linked explicitly said something about 5100MHz Turbo. In addition in my calculation i did explicitly divide by the frequency the core was actually running at during the benchmark.

Though Geekbench 4 is not a toy benchmark, SPEC CPU 2006 and 2017 are heavier and more representative of desktop/workstation workloads.

It is questionable if likes of molecular and fluid dynamics, atmosphere modeling and discrete event simulations are particularly representative. Aside from this the integer kernels of SPEC CPU 2017 are very similar to those used in Geekbench.

 

[Nothingness]

Issue is, that you are quoting single core result for 6700k while i was talking about the integer score, which is lower. Second the one Andrei linked explicitly said something about 5100MHz Turbo. In addition in my calculation i did explicitly divide by the frequency the core was actually running at during the benchmark.

Oops sorry, I thought you were quoting the overall score.

It is questionable if likes of molecular and fluid dynamics, atmosphere modeling and discrete event simulations are particularly representative.

They are representative of workstation workloads.

Aside from this the integer kernels of SPEC CPU 2017 are very similar to those used in Geekbench.

The memory footprint of Geekbench is much lower than SPEC CPU 2006/2017. Also SPEC programs are significantly larger (though not large enough IMHO).

 

[Lodix]

I am happy to see that they more than delivered with their promises if this is running final frequencies of 2'6GHz inside the S9. Just curious to see the power and size of the cores compared to the Cortex A75... Ahem haha

 

[eastofeastside]

What does Exynos 9810 outscoring Ryzen 5 mobile single-thread performance mean in terms of a core-to-core comparison of M3 core vs Ryzen mobile?

Is this a major efficiency statement for M3 or do these scores not translate into a meaningful real world comparison?

Exynos 9810:


Single-Core Score 3717

Crypto Score 1915

Integer Score 3735

Floating Point Score 3440

Memory Score 4543

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

Ryzen 5 2500U:


Single-Core Score 3530

Crypto Score 4139

Integer Score 3224

Floating Point Score 3447

Memory Score 4189

https://browser.geekbench.com/v4/cpu/7027425

 

[decaz]

What does Exynos 9810 outscoring Ryzen 5 mobile single-thread performance mean in terms of a core-to-core comparison of M3 core vs Ryzen mobile?

Is this a major efficiency statement for M3 or do these scores not translate into a meaningful real world comparison?

Exynos 9810:
Single-Core Score 3717
(snip)
Ryzen 5 2500U:
Single-Core Score 3530
(snip)

well.....years after years observing Geekbench, this cross-platform benchmark become much irrelevant now. For short, I don't believe 9810 has similar performance to what intel/amd are offering, unless there's some new method to measure or better realworld application that work in multiplatform has similar result.

 

[itsmydamnation]

you also have 2500u on android that looks like:
Single-Core Score 3730
Crypto Score 2872
Integer Score 3889
Floating Point Score 3317
Memory Score 4209

crypto and float goes down but int goes up alot, probably points to compiler playing a part, given windows using MS c++
compiler vs clang/llvm ( i think for android/apple)

https://browser.geekbench.com/v4/cpu/3852138

 

[itsmydamnation]

Also just want to add:

take the data from anadtechs 2200/2400G review and look at the stilts data:

http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...rmance-unveiled.2533111/page-54#post-39301964
https://www.anandtech.com/show/12425/marrying-vega-and-zen-the-amd-ryzen-5-2400g-review/13

If you look at just core power usage:
10.5 watts for 3.7ghz (2200G)
34.28 watt for 3.5ghz 4 cores (2200G)

we can already see the non linear nature of power consumption in relation to clock/voltage (ie . 10.5 / 3.7 * 3.5 * 4 = 39.7)

From here on i will just look at int score.

If we then look at the stilts Fmax-Vmin chart we can see the difference between 3.5 and 3.7 is about 100mV, the difference between 3.7 and 2.9 is about 200mV. So just taking a layman guess at something like 2.9ghz the per core power consumption for a single thread would be around 5 watts maybe even 4.5 at a strech (voltage, quadratic nature etc).

Now the question is what does that do the score, if you look at the 2200/2300/2500u results 200mhz makes 0 difference to geekbench 4. An 1800x between 4.1ghz and 3.5ghz sees a 200 point different.

So a guesstimate score for a 2.9ghz 5watt zen core using the android platform would be 3500-3600. That probably puts Zen a little behind @ 2.9ghz but not much, thats on a process that if you take samsungs marking is about 30% less power efficient. So overall if your making an honest evaluation of per core performance you cant say which is better, its the same ball park.

Now obviously there is more the just the Core, but the two SOC's focus on such different markets at such different power/performance levels i dont see much point trying to compare the uncore/northbridge/southbridge.

 

[Lodix]

You are taking the data that fits you to make Zen look better. Integer performance is "proportional" with frequency. If that score there doesn't scale it is probably because it was not at max frequency, bandwidth starved or something.

Look at the Intel 7700K it scales perfectly.

 

[Thala]

So a guesstimate score for a 2.9ghz 5watt zen core using the android platform would be 3500-3600. That probably puts Zen a little behind @ 2.9ghz but not much, thats on a process that if you take samsungs marking is about 30% less power efficient.

Core benchmarks like Geekbench are over large frequency range almost linear. Only at the top frequencies you may observe some non-linear behavior due to limitations in the memory subsystem. As such you over-estimating performance numbers heavily when assuming an integer score for Ryzen of 3500@2.9GHz. In fact even if i go with your metric, that the score only degrades by 222 points every 600MHz - which i mentioned is unrealistic at lower clocks, the estimate would be 3651-222=3429@2.9GHz. My estimate would be something around 3100-3200 for Ryzen@2.9GHz - certainly not in the same ballpark as Exynos.
Besides an x86 based design achieving similar efficiency to an ARM based design in the same performance class, assuming both design teams know what they are doing, is close to impossible. That's essentially what was also mentioned by Jim Keller, when he was working on Ryzen.

So just taking a layman guess at something like 2.9ghz the per core power consumption for a single thread would be around 5 watts maybe even 4.5 at a strech (voltage, quadratic nature etc).

I cannot follow this calculation. What absolute voltages are you assuming?

 

[itsmydamnation]

Core benchmarks like Geekbench are over large frequency range almost linear. Only at the top frequencies you may observe some non-linear behavior due to limitations in the memory subsystem.

i am hardly a geekbench expert, but i know x86 uarch's quite well. That doesn't really align, on int workload:
your not Dram throughput bound.
your not L1-L2-L3 clock limited (L3 runs at fastest core speed)
your not going to be any more outstanding memory request limited (cache subsystem at core clock)
your not anymore contention limited ( only limited read/write out of a ccx a cycle)
Zen has a really large oooe window

The only thing is IF latency and that disapears as a measurable bottleneck at around 2933mhz DDR memory speed.

So what is the bottleneck?

As such you over-estimating performance numbers heavily when assuming an integer score for Ryzen of 3500@2.9GHz. In fact even if i go with your metric, that the score only degrades by 222 points every 600MHz - which i mentioned is unrealistic at lower clocks, the estimate would be 3651-222=3429@2.9GHz.

Yet at times it is scales just like that, when other CPU benchmarks show more linear performance. That said i haven't look at sub tests

My estimate would be something around 3100-3200 for Ryzen@2.9GHz - certainly not in the same ballpark as Exynos.

https://browser.geekbench.com/v4/cpu/search?dir=desc&q=AMD+Tambourine&sort=score
What does that do for you?

Besides an x86 based design achieving similar efficiency to an ARM based design in the same performance class, assuming both design teams know what they are doing, is close to impossible. That's essentially what was also mentioned by Jim Keller, when he was working on Ryzen.

No he didn't say anything like that, Micheal clake also didn't say anything like that either,
Jim said we can for the same number of transistors have about a 10% bigger OOOE engine with arm then with x86, Micheal Clake said we can deliver Zen level of performance regardless of ISA.

IF you search RWT you will find the wars about ISA covered very well, to me i would summarize the issue as at 4 wide decode x86 spends more transistors on the front end but it doesn't cost you power, uop caches help that limit and save power, over 4 is a big problem. ARM ISA has some nicer load operations.

At that point your done, everything else weak vs strong memory ordering are all just different trade offs for different workloads.

I cannot follow this calculation. What absolute voltages are you assuming?

Im not assuming a specific voltage, the voltage is dynamic with guardbands etc, but what i am talking about is relative change, the stilts data is hard set minimums, i am suggesting that under normal operation a Zen core @2.9 will be using 200mV less then a core @3.6 (2200U boost). I am then using the per core power data from Anandtechs review to give an estimate on per core power usage @2.9.


At this point im more then happy to blame Microsofts compiler

 

[el etro]

Better than Zen IPC at a phone chip. Well Done.
Weird is to know that most of this engineers were before AMD CPU engineers.

 

[Andrei.]

Guys just stop. You're all using scores from all over the place with zero confidence. Thoser Ars 1800X scores are way too low for example. Plus there's some margin between Windows and Linux scores. No point in arguing about architecture stuff if you can't agree even on if a result is running at a certain frequency.

That being said, this seems to be 3.8GHz http://browser.geekbench.com/v4/cpu/4534181 which is 1.18 for integer and 1.12 for floating point.

 

[itsmydamnation]

Guys just stop. You're all using scores from all over the place with zero confidence. Thoser Ars 1800X scores are way too low for example. Plus there's some margin between Windows and Linux scores. No point in arguing about architecture stuff if you can't agree even on if a result is running at a certain frequency.

I dont think we are arguing , i find it interesting. I think M3 looks great, i just dont think its better then Zen at single thread, I also dont think its worse then Zen at single thread. its likely using less power ( what ~4 watts) at peak per core but its also on a better process. Its uncore is likely way more power efficient then RR as well.

I haven't been picking "best" scores or anything like that but if you look at the AMD Tambourine scores they look legit because the multicore score is only 12k , if their wasn't a tdp limit that would be something like 20K ( 4451 *4 * 1.3(smt) minus a whole bunch for memory component). If you look at the multicore scores you will also see that int is alot higher then the FP/crypto which also makes a lot of sense for a 15~25 watt tdp bound platform.

ultimately its a question i think lots of people are interested in .

 

[Andrei.]

I dont think we are arguing , i find it interesting. I think M3 looks great, i just dont think its better then Zen at single thread, I also dont think its worse then Zen at single thread. its likely using less power ( what ~4 watts) at peak per core but its also on a better process.

Better at what? It's clearly going to be better at IPC but worse in absolute terms. **Core** power should be 1.5-2W.

 

[itsmydamnation]

Better at what? It's clearly going to be better at IPC but worse in absolute terms. **Core** power should be 1.5-2W.

How much is clearly?
if you said
~9% for int
~5% for FP


I wouldn't argue with you. That 10nm process gives both power and clocking advantages compared to 14nm that can be spent how Samsung thinks best. Thus at the end of the day if your just looking at the core uarch, its implementation(with what we know so far obviously) but give some weighting factor for the process they come out pretty equal to me.

Is that 1.5-2W peak/95th percentile avg over a benchmark run before temperate becomes an issue, aka single run of geekbench?

but its now 1am and i need to go to bed.......

 

[Asterox]

Better than Zen IPC at a phone chip. Well Done.
Weird is to know that most of this engineers were before AMD CPU engineers.

What would you do with that CPU/IPC performance on Smartphones, maybe some rendering or video editing etc.Marketing and only marketing, "we need more CPU performance but for what aha benchmarks ok".

For everyday use as good example, Snapdragon 430(8 Core Cortex A53) CPU performance is more than enough for any smartphone.

This "old comparison" is an absurd example, in benchmarks Samsung is beast but in everyday normal smartphone use.

 

[Andrei.]

How much is clearly?
if you said
~9% for int
~5% for FP

I wouldn't argue with you.

I thought you were following the thread.

Based on these two since we can be relatively certain about the clocks:
http://browser.geekbench.com/v4/cpu/3301296
http://browser.geekbench.com/v4/cpu/4534181

 

[Gideon]

Yeah, claiming "better than ZEN IPC 100%" is jumping the gun "a bit". At least a while ago samsung invested heavy resources into optimizing for benchmarks. I still remember a A9 Samsung S3 outpacing a Krait Nexus-4 in many popular CPU benchmarks in phone reviews (while Krait essentially had 40%+ more IPC and similar clocks).

Here are some ancient 2700U engineering sample benchmarks from October on Android:

https://browser.geekbench.com/v4/cpu/4534118

Single-Core Score 4323
Crypto Score 5573
Integer Score 4420
Floating Point Score 3976
Memory Score 4313

And even that comparison is borderline worthless because of the X86 vs ARM difference and all the OS changes Samsung throws around.

I would love to see a more detailed comparison on e.g. Ubuntu for both devices, using more than 1 benchmark. Before claims of "Zen killer phones".