Exynos Galaxy S9 significantly underperforms in comparison to Snapdragon models - Anandtech

[DrMacintosh]

Apple manages to get significantly higher gains from their chips. You need to look at actual performance irl and you can see the iPhone is always going to output more, and better looking frames. 

 

Thats all thanks to Metal. Apppe can do significantly more with their GPU power compared to anyone else. 

[Trixanity]

45 minutes ago, DrMacintosh said:

Apple manages to get significantly higher gains from their chips. You need to look at actual performance irl and you can see the iPhone is always going to output more, and better looking frames. 

 

Thats all thanks to Metal. Apppe can do significantly more with their GPU power compared to anyone else. 

Holy shit, that's some Koolaid you're drinking: "better looking frames". What gains are we talking about?

The actual performance isn't better on the Apple GPU. That's the whole point. At best you can point to onscreen resolution and make claims based on Android devices having to render at higher resolution.

Vulkan exists on Android but adoption is slower because Google isn't twisting developers' arms to get them to do as told.

 

Fact is that Apple's GPU is bigger but doesn't perform any faster and when it does, it's at higher power. In fact, Adreno 630 is quite small compared to its performance. Most of Qualcomm's components on the SoC are smaller than their competition. Die size is also smaller despite having an integrated modem which Apple's SoCs do not.

 

Besides, Apple can't keep that performance sustained so IRL performance will be worse than benchmarks. That goes for both CPU and GPU. Their chips are designed to be good at burst workloads which is fine for most every day tasks but in any graphical workload? No, not good. Of course you also need to look at how much it throttles which is quite a bit.

[DrMacintosh]

16 minutes ago, Trixanity said:

What gains are we talking about?

Objects on screen, lighting effects, and textures. Metal allows these things in 3D workloads to look significantly better than their Android Counterparts. 

 

There is way way more to graphics than just the FPS that a 3D load is being rendered out as.

 

16 minutes ago, Trixanity said:

At best you can point to onscreen resolution and make claims based on Android devices having to render at higher resolution.

They really don’t. These devices with 1440p dispays render most everything at 1080p and simply upscale it. Games are especially are not running at 1440p. 

 

16 minutes ago, Trixanity said:

Their chips are designed to be good at burst workloads which is fine for most every day tasks but in any graphical workload? No, not good.

That’s every SoC. They are all passively cooled......

 

[suicidalfranco]

'member when Exynos used to be good? I 'member 

[Trixanity]

13 minutes ago, DrMacintosh said:

Objects on screen, lighting effects, and textures. Metal allows these things in 3D workloads to look significantly better than their Android Counterparts. 

 

There is way way more to graphics than just the FPS that a 3D load is being rendered out as.

 

They really don’t. These devices with 1440p dispays render most everything at 1080p and simply upscale it. Games are especially are not running at 1440p. 

 

That’s every SoC. They are all passively cooled......

 

Vulkan would pretty much allow anything that Metal does - perhaps Apple just gives more tools to developers out of the box but other than that it's marketing BS you're eating up.

 

Even so, the Adreno 630 is better than Apple's GPU. Many of the workloads you see Apple win will taper off the more/longer you run it to the point where they'll intersect and eventually drop below the other. However, I can't find evidence that every device scales 1440p down to 1080p for rendering. Samsung does it but it can be disabled. I mean sure it makes sense to do so for performance reasons but that's beside the point.  Also, some games may natively render at a fixed 1080p but still I haven't found evidence that's the case for every or most titles.

 

Anyway, the point is that Adreno is the superior GPU by pretty much all metrics. You can argue that Apple's API evens the playing field but that's just an OpenGL vs Metal discussion more than anything else. Vulkan makes that point moot though. But this is an Adreno vs "Apple GPU" discussion, not a Metal vs Android's graphical APIs discussion.

 

And Qualcomm's SoCs throttle way less. So that's not every SoC. Nice try though. The TDP of Qualcomm's SoCs are smaller meaning it requires less cooling to achieve the same or the same cooling to provide a better sustained profile.

[DrMacintosh]

57 minutes ago, Trixanity said:

Even so, the Adreno 630 is better than Apple's GPU.

https://www.tomsguide.com/us/snapdragon-845-benchmarks,news-26598.html

 

[Trixanity]

2 hours ago, DrMacintosh said:

https://www.tomsguide.com/us/snapdragon-845-benchmarks,news-26598.html

 

You just proved my point. And that's an embarrassingly short list of benchmarks with no real analysis.

 

It doesn't touch on the sustained performance where A11 fails but that would only serve to bring more attention to how Apple hasn't accounted for say a gaming workload. 

 

And Snapdragon 845 even went overboard with the power consumption compared to 835. The 835 barely throttled. 845 throttles a lot more but it still the most competitive perf/W in the business.

[D13H4RD]

5 hours ago, Trixanity said:

You just proved my point. And that's an embarrassingly short list of benchmarks with no real analysis.

 

It doesn't touch on the sustained performance where A11 fails but that would only serve to bring more attention to how Apple hasn't accounted for say a gaming workload. 

 

And Snapdragon 845 even went overboard with the power consumption compared to 835. The 835 barely throttled. 845 throttles a lot more but it still the most competitive perf/W in the business.

Throttling isn’t good to me.

 

 

[Razor01]

10 hours ago, DrMacintosh said:

Apple manages to get significantly higher gains from their chips. You need to look at actual performance irl and you can see the iPhone is always going to output more, and better looking frames. 

 

Thats all thanks to Metal. Apppe can do significantly more with their GPU power compared to anyone else. 

 

Its not due to metal, its due to the fact Cortex A73 is it used in the current Arm chips don't have Neon yet, which  won't come out till Cortex A75 which should be coming out later this year.  Now Apple A11 has their own type of Neon architecture.  Even A10 had this that is why even with half the cores it did well against the 8 core ARM chips.

[Zodiark1593]

47 minutes ago, Razor01 said:

 

Its not due to metal, its due to the fact Cortex A73 is it used in the current Arm chips don't have Neon yet, which  won't come out till Cortex A75 which should be coming out later this year.  Now Apple A11 has their own type of Neon architecture.  Even A10 had this that is why even with half the cores it did well against the 8 core ARM chips.

Uhmm, NEON is a SIMD extension that's been used in nearly all smartphone SoCs since Cortex A8 (2009-ish).

[Razor01]

51 minutes ago, Zodiark1593 said:

Uhmm, NEON is a SIMD extension that's been used in nearly all smartphone SoCs since Cortex A8 (2009-ish).

 

Its hardware driven in A10 and A11, Cortex A8 can do it via software, but its slow.  A73 was supposed to have a Neon based engine, never made it to the final plans, it was delayed till A75.
 

https://developer.arm.com/products/processors/cortex-a/cortex-a75

 

Quote

 

With over 20% more integer core performance and high-performance NEON and FPU engines

 

 

https://www.anandtech.com/show/10347/arm-cortex-a73-artemis-unveiled/2

 

Quote

 

Floating point as well as ASIMD and NEON operations are handled by two pipelines, some of the capabilities we’ll go into detail later on. We find a single branch monitor and two dedicated Load and Store AGUs.

 

 

The pipelines in A73 just didn't have the performance needed for proper acceleration like what Apples A10 and 11 have.

 

There is also a huge cache difference between A11 and the latest snap dragon, where A11 has 8mb L2 cache , snapdragon only has 3 mb L2 cache and no L3 cache. This changes the size of the chips greatly, where A11 is double the size too.  So we can imagine the latency differences between these two chips based on just the cache amounts.

 

The performance is highly architecturally different.  A75 cortex should close the gap but its going to depend on what Apple does next cause A75 isn't coming out till later.

 

This guy explains it better than I do. 

 

This is just catching up to what the A11 chip can do.  Now A11 is a 6 core chip I believe?  So a A75 based 8 core chip should beat it.

[Zodiark1593]

2 hours ago, Razor01 said:

 

Its hardware driven in A10 and A11, Cortex A8 can do it via software, but its slow.  A73 was supposed to have a Neon based engine, never made it to the final plans, it was delayed till A75.

Not sure what PCB you've been smoking. NEON is a hardware driven. It always has been hardware driven. To do so otherwise makes zero sense.

 

What can be done is to expand the capabilities of the NEON engine, which is what Cortex A75 does. Does not mean it was never a thing before. It was introduced in the ARMv7 days. Nvidia even specifically said (for the launch of Tegra 2) that to implement NEON incurs a substantial die size increase. 

 

I recommend reading ARM's own documents on NEON as well.

https://developer.arm.com/technologies

 

http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.dht0002a/BABIIFHA.html

[Trixanity]

6 hours ago, Razor01 said:

 

Its hardware driven in A10 and A11, Cortex A8 can do it via software, but its slow.  A73 was supposed to have a Neon based engine, never made it to the final plans, it was delayed till A75.
 

https://developer.arm.com/products/processors/cortex-a/cortex-a75

 

 

https://www.anandtech.com/show/10347/arm-cortex-a73-artemis-unveiled/2

 

 

The pipelines in A73 just didn't have the performance needed for proper acceleration like what Apples A10 and 11 have.

 

There is also a huge cache difference between A11 and the latest snap dragon, where A11 has 8mb L2 cache , snapdragon only has 3 mb L2 cache and no L3 cache. This changes the size of the chips greatly, where A11 is double the size too.  So we can imagine the latency differences between these two chips based on just the cache amounts.

 

The performance is highly architecturally different.  A75 cortex should close the gap but its going to depend on what Apple does next cause A75 isn't coming out till later.

 

This guy explains it better than I do. 

 

This is just catching up to what the A11 chip can do.  Now A11 is a 6 core chip I believe?  So a A75 based 8 core chip should beat it.

A75 based chips will have L3 cache. Snapdragon 845 is A75 based and not only does it have an L3 cache, it also has an L4 cache. However levels of cache is one thing, actual implementation is another. Type of cache, size of cache and placement of cache are some factors that matter a lot. 

 

Snapdragon 845 has all these caches but it's not all good. The memory controller is further away from the CPU resulting in higher latency to go to memory. The L4 cache will try to hide that latency but it's not a perfect solution. The size of these caches are smaller than ARM recommends for the best performance too. So the performance of the chip isn't as good as it should be and it shows in the benchmarks.

[Zodiark1593]

22 hours ago, DrMacintosh said:

Apple manages to get significantly higher gains from their chips. You need to look at actual performance irl and you can see the iPhone is always going to output more, and better looking frames. 

 

Thats all thanks to Metal. Apppe can do significantly more with their GPU power compared to anyone else. 

Low level api tends to reduce the overhead of making commands to the gpu, reducing cpu load. This typically won't effect shaders. GPUs on Android phones certainly aren't limited by shader capabilities. What this will effect, however, is the number of unique objects that can be drawn to the framebuffer at once.

 

In OpenGL ES, making draw commands seems to be obscenely expensive, so tricks such as instancing (duplicating the same object) are used. Shadow mapping may also incur a hit from draw calls due to the need to render to an offscreen buffer (making it especially expensive on Android).

[Razor01]

4 hours ago, Zodiark1593 said:

Low level api tends to reduce the overhead of making commands to the gpu, reducing cpu load. This typically won't effect shaders. GPUs on Android phones certainly aren't limited by shader capabilities. What this will effect, however, is the number of unique objects that can be drawn to the framebuffer at once.

 

In OpenGL ES, making draw commands seems to be obscenely expensive, so tricks such as instancing (duplicating the same object) are used. Shadow mapping may also incur a hit from draw calls due to the need to render to an offscreen buffer (making it especially expensive on Android).

 

 

I was under the impression there was something wrong with Cortex A73 Neon engine but indeed its the latency that causes the problems.

 

No its not things like instancing that is helping metal.

 

Instancing in Open Gl was there since Open Gl 1.2, its quite an old feature and it was in Open GL ES when it was launched with VBO.  It was improved with OpenGL ES 2.0 though.  And with 3.0 more features were added to it.

 

Actual game performance, is better on A73 than the expected performance from Geek bench tests, that isn't where the problem is with ARM chips, so its highly unlikely due to API's being used.

 

Geekbench is not a good indicator of over all application performance specially gaming performance.  Geekbench is a good performance indicator for OS and web browsing though.

 

If you want to talk about draw calls, yeah that can hurt android as you stated, but draw calls are only part of the problem because these CPU's, programs have to optimized heavily for them.  Even with 6 or 8 cores, you can't possible expect to use more than 2 cores (or total over all 25% of all cores for gaming because unlike PC's there is a heavy burden on power consumption with background applications, that can't be stopped or no need to stop, unlike other devices.

[xriqn]

Snapdragon was always superior