[AT] Eynos 5433 Note 4 (SunSpider dropped by AT)

[lopri]

Edit (2/11/2015): I changed the thread title because the article is worthy of praise just for the sheer amount of time and effort involved - by my guess, but I do not think it is an exaggeration. It is a fascinating technical account on a fascinating chip. I still haven't finished reading it but am trying to. I hope more people read it so that we can have actual discussions on tech, instead of armchair-CEO'ing.

--------
No one should miss that useless benchmark, and I thank AnandTech for living up to expectation, albeit late. Here are key passages from just-published Exynos 5433 article.

http://www.anandtech.com/show/8718/the-samsung-galaxy-note-4-exynos-review

Before discussing the results I'd like to mention that I saw a huge discrepancy between Chrome and the stock browser. In Chrome, which I checked was indeed running the exactly same build version for both the Note 4 and the Alpha, I could see the Alpha consistently outperforming the Note 4 in all the tests. Due to this I deemed Chrome to be extremely unreliable as an apples-to-apples comparison of efficiency and reverted to the stock browser. Here I could see proper performance scaling that we would actually expect from the new core architectures and the clock advantage. (pg.6)

The scores on the Exynos version of the Note 4 are outstanding, beating out all existing devices in our more complex benchmarks. SunSpider is the exception, but seeing the vast discrepancy between browsers and even the Snapdragon version matching Apple's new A8 has made us come to the conclusion that this benchmark has run it's time as a valid test case, and thus we're dropping it from our 2015 test suite going forward. (pg.8, emphasis mine)

The article is incredibly in-depth and at times its writing is confusing, but without a doubt it is the best explanation on big.LITTLE we've got today. Highly recommended read.

 

[ChronoReverse]

It hasn't been a valid test case for a long time. I like how the big reason why they're dropping it seems to be because the Snapdragon matches Apple's A8. What happened to the Anandtech I used to know and love?

 

[lopri]

I think you are reading too much into it. I did not read the vibe that you suggest from the paragraph. Besides, it is 100% better to get rid of it than to keep it perpetuating misunderstandings. And 1000% better than to mix/match scores from Chrome, Safari, and Internet Explorer depending on what was reviewed, which Anand Lal Shimpi used to do.

 

[lopri]

Here is a TechReport's take on the Exynos 5433. Not as thorough as AT's, and there appear to be disagreement in some aspects of performance.

http://techreport.com/review/27539/samsung-galaxy-note-4-with-the-exynos-5433-processor

Per AT's account, there is a meaningful gain to be made if Samsung (or anyone who deploy big.LITTLE arch) pay more attention to Global Task Switching scheme. Apparently even the current implement in the 5433 wakes up too many unnecessary cores too often, resulting in wasted power. This raises the 7420's stake even higher, not just for Samsung but for every OEM that will rely on S810. It will be interesting to watch who can tame this beast (A57+A53) in the most efficient manner.

 

[Commodus]

AnandTech has relied on multiple browsing tests for a while, so the anti-Apple camp's cries of "conspiracy!" don't really hold water. It's a matter of rethinking whether or not a given test accurately reflects overall performance; if it's clear that SunSpider seldom mirrors real life, it should get cut.

 

[lopri]

^ It was not a "conspiracy" because it was too transparent to be called one. Go check out Surface (Pro) reviews, for instance. You will see reviews switching browsers around depending on the results the reviewer desired to present you with. Android was never treated favorably for now-obvious reasons, although it was not difficult to imagine why even then. Thankfully the new writer and the editor saw the practice no longer tenable, and it appears they want more complete and accurate picture of performance measurement going forward. I think it is for the better for everyone, especially users and consumers for false benchmark numbers do not benefit them.

Anyhow, it is amazing just how small those LITTLE chips (Cortex-A7) are.

 

[lopri]

Interesting (and completely different) takes on relatively poor memory bandwidth of big.LITTLE architecture.

http://techreport.com/review/27539/samsung-galaxy-note-4-with-the-exynos-5433-processor/3

The Note 4's relatively weak showing could be the result of bumping up against a power management limit—directed tests sometimes do that—or it could be something else. Most modern CPUs max out their memory bandwidth by pairing a relatively large cache with a predictive pre-fetch mechanism that analyzes access patterns and pulls data from memory before it's needed. The Cortex-A57 cluster in the Note 4 may not be tuned as aggressively as its competition for whatever reason.

http://www.anandtech.com/show/8718/the-samsung-galaxy-note-4-exynos-review/5

ARM's relatively poor memory bandwidth figures have garnered them a poor reputation when it comes to memory performance, but what we are seeing here might be a gross misrepresentation of real-world performance. To understand how these figures come to be, we need to look at how the CPU is wired to the SoC's interconnect and memory controllers. ARM, as opposed to designs by Apple or NVIDIA, uses separate read and write data-ports in its fabric. On the cluster level, this is a dual 128-bit interface (one for reads, one for writes) that connects to matching ports of the SoC's memory controllers via the CCI's (Cache Coherent Interconnect) crossbar architecture. On the Exynos 5430 and 5433, the CCI runs at half the DRAM frequency, meaning 412.5 MHz for the aforementioned SoCs. This results in a maximum physical bandwidth of 6.6 GB/s in each direction.

What most of today's synthetic benchmarks portray is only the bandwidth measured in either direction, giving ARM a distinctive disadvantage. Total achievable bandwidth can reach double these figures. In fact, when we execute simultaneous read and write tests (multithreaded on two CPUs) we benchmark bandwidth numbers reaching the theoretical peaks of the memory controllers at 13.2GB/s. Interestingly, it seems ARM is employing the same setup to the L2 cache as bandwidth there also doubles to up to 25GB/s for the 5430's A15 and 27.5GB/s for the 5433's A57 clusters.

Quite a contrast!

 

[Sweepr]

Impressive in-depth investigation by Andrei and Ryan. Samsung really pulled off a surprise with this chip, few people were expecting a Cortex A57/A53 mobile SoC in 2014. If Samsung had gone Exynos-only by then (like they seem to be doing now with the Galaxy S6) they would have the first 64-bit enabled Android flagship out there (half a year before the Snapdragon S810 devices).

For those who don't visit the CPU section, I posted lots of benchmarks and comparisons here.

 

[lopri]

AnandTech's review makes me really wonder if (4xA53) + (4xA53) design makes any sense at all. If the overhead from GTS (Global Task Switching) is still so great, what is the point of bunching two nearly identical clusters? Seems to me the gain should be atrocious while power consumption can theoretically double.