The Kirin 990 SoC

The Kirin 990 is an unusual SoC this year for the fact that there’s two variants of the chip. Huawei and HiSilicon have opted to hedge their bets in terms of silicon manufacturing with TSMC and create two versions of the chipset: The regular 4G variant is created using the existing TSMC N7 manufacturing process which was also used for the previous generation Kirin 980 processor, whilst the 5G variant of the Kirin 990 is a completely altogether different chip using TSMC’s new N7+ manufacturing node which uses EUV lithography.

Huawei Kirin 990 Family
AnandTech Kirin 990
5G
Kirin 990
(4G)
Kirin 980 Kirin 970
CPU 2xA76 @2.86G
2xA76 @2.36G
4xA55 @1.95G
2xA76 @2.86G
2xA76 @2.09G
4xA55 @1.86G
2xA76 @2.60G
2xA76 @1.92G
4xA55 @1.80G
4xA73 @2.36G
4xA53 @1.80G
GPU G76MP16
600MHz
G76MP16
600 MHz
G76MP10
720 MHz
G72MP12
850 MHz
NPU 2 + 1
Da Vinci
1 + 1
Da Vinci
2
Cambricon
1
Cambricon
Modem Balong 5G 4G 4G 4G
DRAM LPDDR4-4266

+ LLC
LPDDR4-4266

+ LLC
LPDDR4X-4266 LPDDR4X-3733
Die Size 113.31 mm2 ~90 mm2 74.13 mm2 96.72 mm2
Transistors 10.3b ~8.0b 6.9b 5.5b
Mfc. Process N7+ N7 N7 10FF

Although the two chipsets are two different pieces of silicon, going as far as needed different physical designs due to rule incompatibility between the N7 and N7+ nodes, the two chipsets still essentially share a lot of common features and are very similar to each other in terms of the IP they use.

Both chips follow a 2+2+4 CPU cluster setup, with two high performance A76 cores clocking in at 2.86GHz, two mid-performance cores clocking in at 2.09GHz on the 4G variant and 2.36GHz on the 5G variant, and four Cortex A55 cores at 1.86GHz for 4G and 1.95GHz for the 5G chipset. HiSilicon continues to employ a 4MB L3 cache in for the cluster.

On the GPU side, the chip uses a very large implementation of the Mali G76. At 16 cores, it’s the largest configuration of this GPU we’ve seen, and is clocked in at 600MHz.

Another auxiliary IP difference between the two variants is the fact that the 5G version employs a stronger NPU. We find a dual-core variant of the new DaVinci microarchitecture, while the 4G chipset makes due with a single core version of the IP.

Naturally, the single biggest difference between the 4G and 5G variants is implied in the names of the parts: The 5G variant comes with an additional 5G modem that isn’t found in the 4G version.


Die shot source: TechInsights - Labelling & Custom contrast: AnandTech

TechInsights has published the die shot of the new Kirin 990 5G and we took a closer look at the IP blocks that we can identify. Unfortunately, while there isn’t a public die shot of the Kirin 990 4G out there, we did have a chance to have a look at it and cross-compare the differences to the 5G version in order to better pin-point blocks such as the NPU and modem.

The Kirin 990 5G comes in at 113.31mm² which isn’t actually too big considering that HiSilicon is dedicating a lot of area to the massive GPU as well as the bigger NPU this generation. The CPU cluster at the middle of the SoC is self-explanatory and we’re able to identify the A76 as well as A55 cores. It’s interesting to see that the approach of using the same CPU IP for the “big” and “middle” cores in fact doesn’t really result in much of a physical size difference between the two groups – the biggest differences here are purely in the transistor-level implementation. This comes in contrast to Qualcomm’s 1+3 CPU configuration where the one performance cores is actually larger than the remaining three cores, partly due to the doubled L2 cache.

The DaVinci NPU would have been hard to identify weren’t it for the fact that we saw similar block structures on the 4G and 5G variants- with the difference that the 5G variant has two identical adjacent blocks while the 4G variant only has one. It’s a quite sizeable piece of the SoC so it does show that the company puts a lot of investment into AI.

The modem was the trickiest part to identify – the highlighted labelling and blocks in the image above is just my best estimate of what’s seen on the SoC. The 4G variant of the SoC completely lacks the block labelled as the 5G modem, so just by pure elimination of other possibilities we have to come to the conclusion that this is indeed the modem IP. The block is roughly a little over 15mm² in size, which while sizeable in terms of a SoC block, isn’t all that big in the grand scheme of things, especially when compared to discrete modem sizes in excess of 50-60mm².

The oddity here is that I still believe that there’s two modem subsystems and that the 4G modem as presented on the 4G Kirin 990 is still present on the 5G variant. While relocated in position, the marked group of IP must be the 4G modem IP as there’s no other difference between the 4 and 5G variants of the chip; the 5G variant merely adds in more IP blocks rathe than changing things around too much. The only major lackings on the 5G version is that I’m identifying far less interface I/O to what I believe are PCIe PHYs. Possibly because this variant of the chip has a 5G modem integrated, there’s no practical need for these additional lanes.

Finally, a big important addition to the Kirin 990 is the adoption of a new last level cache, or how we like to call it, a system level cache. The block can be identified at the centre of the chip and judging by the physical size, is 3-4MB large. HiSilicon this year is said to have paid a lot of attention to the memory subsystem and improved is performance quite a lot.

If we look at the memory latency results of the CPU, we indeed see that the company has been able shave off ~10ns of latency and achieve sub-100ns memory latency, making the new chip the fastest (Along with the Exynos 9820) in this regard, although Apple’s large caches are a different beast altogether.

It’s to be noted that we don’t see the SLC cache in the CPU’s memory latency results, evidently HiSilicon is bypassing the cache on the high-performance CPUs in order to get better latency to DRAM. We weren’t able to confirm if the middle or little cores have access to the SLC – I imagine they do as that’s also how Samsung and Qualcomm have implemented their power management policies for their SLCs.

Given that the SoC doesn’t contain any new IP, we’ll spare the readers the effort of going into detailed SPEC results. However, the overview is still interesting to look at in terms of the results. As a reminder, this is the 4G variant of the Kirin 990 which uses the same manufacturing node as the Kirin 980, as well as the same CPU IP. The differences between the two chips is the improved physical design that HiSilicon was able to implement, as well as differences of the new memory subsystem.

The overall improvements here are quite impressive – the new chip sports higher performance all whilst using less energy and power. The interesting thing here is that most of these improvements in efficiency aren’t due to the better CPU implementation; there’s definitely improvements here, but they are vastly overshadowed by the improvements we’re seeing in the memory subsystem. Benchmarks which have little memory pressure and thus mostly reside on the CPU caches have the smallest improvements in terms of energy efficiency, whilst the big memory heavy benchmarks such as 429.mcf or 462.libquantum are seeing perf/W improvements of up to 27% due to both very large performance improvements as well as actual lower power usage.

So whilst HiSilicon wasn’t able to implement the new Cortex A77 CPU IP this year into the Kirin 990, the new chip shouldn’t be mistaken as simply having the same or similar performance to the Kirin 980 or Snapdragon 855 just because they have the same A76 CPU IP. The new chip is definitely faster and more efficient than the predecessors.

Introduction & Design System Performance
POST A COMMENT

59 Comments

View All Comments

  • s.yu - Thursday, November 28, 2019 - link

    "a unique 90° curved screen on its sides – a first not only for Huawei but for the smartphone market overall"
    ...IIRC NEX 3 with the same screen curve design was launched 3 days before this model.
    But the comments regarding the ergonomics certainly push me further towards the ROGP2.

    "The new sensor is a native 16:9 unit" really surprises me as I thought it was a 3:2, since it shoots stills in 3:2, in this case the 4:3 crop in night mode would be even more of a compromise cutting out a larger portion of the sensor, essentially highlighting how severe the falloff and in cam shadow lifting is.
    But this doesn't explain the poor DR of the new unit.

    If I were shooting an UWA I'd probably prefer a narrower one, as the ~12mm FoV is overdramatic and incredibly difficult to master, but this isn't the main issue as I don't believe a significantly relevant proportion of users of these ultimately consumer devices actually understand photography enough to have that in mind.

    The night mode on the UWA still has severe smearing that hasn't been resolved, especially obvious in the shot of the theater, whose exacerbation in edges and corners under a certain lighting threshold is observable in the 2nd sample on the stairs, so what I observed from GSMA's very few night samples turns out to be universal, that the UWA night mode on this big big sensor is actually worse than Samsung's tiny tiny sensor with a far wider FoV. Peculiar.

    So the conclusion here is: When not much FoV is needed, DR isn't too extreme, and lighting is either sufficient enough not to need night mode or lacking enough that Samsung's tiny sensor tends to capture no data, this new UWA has the advantage.

    A minor detail regarding "The ISO51200 of the sensor in the regular mode is able to get better results than the image stacking of ISO1600 pictures in night mode." The ISO1600 is probably calculated from the accumulated shutter durations of the burst, I don't believe it's likely that a burst could be captured without vibration under this lighting with each shot at ISO1600.

    Finally, I was expecting a sprinkler test of the "7680fps", and preferably a test on effective video resolution, shame this wasn't included.
    Reply
  • Andrei Frumusanu - Thursday, November 28, 2019 - link

    You are right about the 3:2, I was absentmindedly writing 16:9.

    In regards to the slow-mi, Ian did an article few months back:
    https://www.anandtech.com/show/14897/a-day-with-th...

    The capture is just 2000fps interpolated 4x, I investigated this in the software.
    Reply
  • s.yu - Thursday, November 28, 2019 - link

    Yeah I know, just that a running tap was too slow, while a sprinkler test would yield more as to how the interpolation is done, e.g. the NPU technically should be able to afford some object tracking resulting in much better interpolation than mere averaging or something.

    Also a comparison of resolution between all the slow-mo that on paper appear to be 720P could help assess actual quality, my second point.
    Reply
  • Dazzler86 - Thursday, November 28, 2019 - link

    Eventhough the DR is not up to the unrealistic level of Samsung, its subtle enough & flagship grade.

    Whining about Night Mode? The Mates doesnt even need to use Night Mode to beat other competition, its still a low light king, as stated by the review.

    Funny that you only stated negativity about the camera phone but ignores its pros that beat others. Best UWA implentation, best detail preservation on a smartphone camera, ever.

    I dont care about your other post about their politics, but i know i made the right choice, choosing Huawei for a point & shoot camera along with my main Leica Q & Fuji X-E3.
    Reply
  • s.yu - Thursday, November 28, 2019 - link

    Oh I know a Huawei fanboy when I see one alright.

    "Eventhough the DR is not up to the unrealistic level of Samsung, its subtle enough & flagship grade."

    So more DR than Huawei equals unrealistic? lol. Apple's UWA output also has more DR and is quite realistic. The processing of the two Samsung variants is also different, and not entirely less realistic than for example P30P.

    "The Mates doesnt even need to use Night Mode to beat other competition"
    Rather that it would lose to others where it might have won without night mode?

    The Mate, which on auto:
    https://images.anandtech.com/galleries/7356/Mate30... beats Samsung's night mode(for as wide as its FoV would cover): https://images.anandtech.com/galleries/7356/S10E_2...
    loses handily to Samsung's night mode with its own night mode: https://images.anandtech.com/galleries/7356/Mate30...
    with much narrower FoV, much more severe smearing etc.
    This inconsistency, a strong argument against night mode use in many nighttime conditions, is a liability, because the user is forced to guess when night mode works better and when it's far worse, and the penalty for making a mistake in judgement is a heavy one.

    Andrei left out a lot of individual analysis between samples, but clearly there are too many catches for it to be declared outright "king", which I summarized as "When not much FoV is needed, DR isn't too extreme, and lighting is either sufficient enough not to need night mode or lacking enough that Samsung's tiny sensor tends to capture no data, this new UWA has the advantage."

    Note that I acknowledge the advantage, where there is one, only I also acknowledge the disadvantage, where there is one, unlike you who does nothing but brag.

    "Best UWA implentation"
    is, again, false. You don't say that a 21/2 is better than a 12/5.6, because the drastic difference in FoV makes then incomparable, and the UWA, already ~18mm in stills, a 1.5x crop compared to Samsung and Apple's UWA, is a ~21mm equiv. FoV in night mode's forced 4:3 crop, so no matter how much cleaner it is, you can't call it something along the lines of "the best". Your claim is akin to "my 16-35/2.8 is better than your 12-24/4", or "my 50/0.95 is better than your 75/1.4", which is nothing but ignorant fanboy talk. You're lucky this is Anandtech, or else you'd be laughed at commenting like that under DPR.

    Also some things I left unsaid as I already covered some of Mate30P's performance in the comments of the Pixel 4 review, I felt no need to repeat them.

    "Leica Q"
    You bought an expensive toy that will go to waste in your hands. Your X-E3 purchase does expose your taste for JPG filters over material performance gain though, not surprising for a Huawei fanboy.
    Reply
  • SydneyBlue120d - Thursday, November 28, 2019 - link

    Thank You Andrei for your great review!

    After all this years of Android it is really disappointing that still at the end of 2019 no one seems to be able to beat the iPhone 11 PRO display accuracy.
    Reply
  • phils1969 - Thursday, November 28, 2019 - link

    Anyone else had an issue with the Nano Memory card not being detected? Dual sim works so the slots fine, but 1 sim and 1 NM card and its not seeing the NM card, the wifes Mate 20 Pro sees it fine so its not the card? Reply
  • Calista - Thursday, November 28, 2019 - link

    If anything this shows we really need properly split Alphabet/Google from Android. Android being open source, it belongs to the world. But that's not much use if it's still locked to a single company in a country more than happy to be both judge, jury and executioner when it comes to world politics. All the major cellphone vendors should work together to replace the Google services with open alternatives. It's easy to pick on Huawei, but what's the next target? Samsung? Oppo? LG? Sony? Some other company not from USA? Reply
  • liquid_c - Saturday, November 30, 2019 - link

    In the pfw chart, the iPhones get 2 subsections (cold/peak and hot) while the other devices, don’t. Why is that? Reply
  • AidenP - Saturday, December 7, 2019 - link

    Looks decent (except for those extra curved sides) ,and the specs are up there,it seems, I think at 500-600 Euros would be a realistic value offer , but for now I will stick to my iPhone 8,for at least a few more years . Reply

Log in

Don't have an account? Sign up now