Samsung Unveils Exynos 7872 Hexa-core Arm Cortex A73/A53 Mobile Processor

Samsung has introduced a new hexa-core Exynos 5 series processor with Exynos 7872 featuring two Cortex A73 cores, four Cortex A53 cores, a Mali-G71 MP1 GPU, and an LTE modem that supports Category 7 with 2CA for 300Mbps downlink speed and Category 13 with 2CA for 150Mbps.

The new processor claims to bring premium performance to mid-tier device with a doubling of single core performance compared to its predecessor, which should be Exynos 5430 or Exynos 5422 octa-core Cortex A15A/7 processor.

Samsung Exynos5 7872 processor specifications:

  • CPU – Dual-core Cortex-A73 @ up to 2.0GHz, and Quad-core Cortex-A53 @ up to 1.6GHz
  • GPU – Arm Mali-G71 MP1
  • Memory I/F – LPDDR3
  • Storage I/F – eMMC 5.1, SD 3.0
  • Display – Up to WUXGA (1920×1200)
  • LTE Modem – LTE Cat.7 2CA 300Mbps (DL) / Cat.13 2CA 150Mbps (UL)
  • Connectivity – Wi-Fi 802.11n Dual-band, Bluetooth 5, FM Radio
  • GNSS –  GPS, GLONASS, BeiDou, Galileo
  • Camera –  Rear 21.7MP, Front 21.7MP
  • Video – Full HD 120fps encoding and decoding with HEVC (H.265), H.264 and VP8, and decoding with VP9
  • Process – 14nm FinFET Process

One of the first phone based on the processor will be Meizu M6s with a 5.7″ HD+ display (18:9 aspect ratio), 3GB of RAM and 32GB / 64GB flash, a 16MP rear camera, 8MP front-facing camera, and a 3,000 mAh battery. It would also be a nice (and affordable) update to ODROID-XU4 board based on Exynos 5422 SoC, we’ll have to wait and see.

Support CNX Software - Donate via PayPal or become a Patron on Patreon

19
Leave a Reply

avatar
18 Comment threads
1 Thread replies
6 Followers
 
Most reacted comment
Hottest comment thread
9 Comment authors
crashoverridejonscravitwillym][skobenjamin Recent comment authors
  Subscribe  
newest oldest most voted
Notify of
Memeka
Guest
Memeka

I think 5422 originally was top tier, not middle. And it’s like 3 generations old now

tkaiser
Guest
tkaiser

Anything known about Exynos 7872’s wired IO capabilities? Stuff like USB Hi-Speed or maybe even SuperSpeed?

tkaiser
Guest
tkaiser

cnxsoft :
That’s a mobile chip, so they did not provide much info for this part.

Yeah, I know. But given the orientation on wireless capabilities it would be possible that this SoC lacks any performant USB implementation. At the bottom of the Meizu M6s there seems to be an USB-C port but USB-C with only Hi-Speed data lines exposed (and maybe alternate modes for eg. external display) is a valid use case.

tkaiser
Guest
tkaiser

I really hope the information provided there (USB 2.0 only) is wrong: https://www.devicespecifications.com/en/model/e5e9481c

blu
Guest
blu

I was wondering when Samsung were to introduce the successor to the 54xx series. Overall, a very conservative choice but also one that does not need to be anything else.

willy
Guest
willy

Memeka :
Yeah it’s got a bit better CPU (2 core A73 vs 4 core A15)

I’d be *very* cautious on this statement. In my various tests, I never got an A72 to run faster than an A17, the A15 is supposedly a bit faster than A17 on most workloads, and the A73 is not expected to be faster than A72 but more energy efficient. For my use cases (gcc), a dual-A72 is half as fast as a quad-A17, and I hardly imagine it would be much different here. At least the CPU supports more RAM 🙂

benjamin
Guest
benjamin

I wish someone made a cheapish implementation of a73 cores. Like using 28nm fab and maybe a pair of a73 cores and a pair of a53 cores to go along. And a small gpu to keep the die size down.

blu
Guest
blu

Memeka :
I know they’re both Exynos5, but i don’t understand why anyone would think it succeeds 5422. Yeah it’s got a bit better CPU (2 core A73 vs 4 core A15) but it’s got worse GPU (1 core G71 vs 6 core T628), no USB3 so I would say it’s just a different category.
When 5422 was introduced, it was top tier SoC, this one is not…

You’re right this is not top tier by any stretch of the imagination, at least GPU-wise, but at the same time, I’m not sure 5422 was top-tier either — it did not use leading fabnode and did not stay as a flagman SKU for long — 5433 came shortly after. So it seems to me 5422 was not meant to be a top-tier SKU, just the top-CPU SKU of the armv7 lineup.

m][sko
Guest
m][sko

@willy My HiKey960 is faster than Odroid XU4 but not much. I would expect more. But problem is that both run on 2GHz
Kirin 960 should be able to run 2.8GHz but HiKey960 has ram module on top of CPU soc => no way (cooling problem)
and 2GHz is max for 5422 with active cooling/big heat sink

So ARM, AMD and Intel marketing about faster CPU is all about it can run on faster frequency.

willy
Guest
willy

m][sko :
@willy My HiKey960 is faster than Odroid XU4 but not much. I would expect more

Just curious, on what type of workload ? In my case the RK3399 in the H96max can give me similar build times as the RK3288 in the MiQI, but it varies a lot. The reason is that there are 6 cores in the RK3399, 2 fast (A72) and 4 slow (A53). The A53 are almost exactly half the performance of the A72 on this use case, resulting in approximately the same performance as if I had 4 A72… except that I don’t have this, so every time a long task is dispatched to an A53 (which happens 2/3 of the times), it in fact takes twice as long as it would take on the A17.

So in the end it depends a lot on your workload, but having half the number of fast cores and twice the number of slow ones often results in a big headache to try to guess if a workload will be faster or slower. And sometimes it can really be twice as slow.

For my use case (build farm), I got the best results by *not* using the A53 cores and keeping only the two A72 ones. Otherwise the A53 almost always increase build time. I ended up only using the A53 on very small files and when all other nodes are already in use :-/ I think I hate Asymmetric Multi-Processing.

m][sko
Guest
m][sko

@willy
My use case was x264 encoder
just search Hikey960 benchmarks on 96boards forum
But I disabled small cores as x264 can use 4 threads.

willy
Guest
willy

@m][sko
OK so finally your observations there more or less match mine, because you found that 4xA73 run only slightly faster than 4xA15 on this specific workload. A real gain coming from armv8 is crypto, but not everyone needs it 🙂

blu
Guest
blu

@willy
The real ARMv8 gain in some compute cases is the notably improved ASIMD instruction set. While NEON has been an advanced, well-thought-out SIMD ISA from its inception, ARMv8 ASIMD (v2) introduced fp64, some much-needed horizontal ops, and even further orthogonalization of the instruction set. That, along with the re-design of the reg-file layout, helps compilers (e.g. autovectorizers) and SIMD coders a lot.

My pet peeve about ARMv8 devices being “stuck” in ARMv7 land by some vendors has always been about the above, otherwise I don’t give a flying figment if my scalar int code will be 64-bit or not ; )

jonscravit
Guest
jonscravit

@cnxsoft
Samsung learned from Intel

crashoverride
Guest
crashoverride

@willy

For my use case (build farm), I got the best results by *not* using the A53 cores and keeping only the two A72 ones. Otherwise the A53 almost always increase build time. I ended up only using the A53 on very small files and when all other nodes are already in use :-/ I think I hate Asymmetric Multi-Processing.

I saw this same behavior on RK3399. Pinning all tasks to the big cores always yielded better results. My theory is that this is due to the kernel/scheduler rather than the hardware itself. The basis for this is comparison to A53 workloads on S905.

willy
Guest
willy

@crashoverride
In my case it’s not a scheduler issue but just a hardware limitation : the A53 cores in this CPU simply are the slowest cores of all the farm, so each time they pick a task there’s a risk that everything around finishes then waits for this A53 to complete its task 😉

But even if I were running on 4xA72, it would not be faster than the 4xA17 I already have on the MiQi, which is why for me this CPU is pointless in the end :-/

I suspect ARM doesn’t see well how to improve their current cores’ performance by now. Just take a look at the pipelines on the last high-end versions. A57, A72, A73, A75 : 2-wide decode, then 3-wide, then 2-wide again “because it was enough”, then 3-wide again “because it’s better”. It gives a clue that they’re having difficulties improving the IPC on these chips while keeping the thermal envelope low. I’m not blaming them, it’s a real challenge, but going back and forth between two approaches over the last 4 generations isn’t going to convince a lot of people that the last choice is the best one. However they’ve improved branch prediction and speculative execution, but these days it’s not always seen as an improvement 😉