HiSilicon Kirin 970 Android SoC Review - Performance and Efficiency

    Today, there are only two vertically integrated mobile device OEMs that have full control over their chips: Apple and Huawei, and one of them is more integrated than the other. This is Huawei, which among other things has a modem of its own design. Huawei's semiconductor division, HiSilicon, over the past few years, has remained the only company that has been able to do what the others found too tough: enter the high-end market with solutions that can compete with the current business leader, Qualcomm.

    I remember the release of Honor 6 with the recently emerged (and then little known) SoC (System-On-Crystal, single-chip system) Kirin 920. This was the first Huawei device with an integrated SoC that we reviewed. This, like the next generation, the Kirin 930, suffered from immaturity and had serious problems, such as a voracious memory controller and just an unusable camera processing pipeline (ISP / DSP). Kirin 950 was, in my opinion, a turning point for HiSilicon, since the product corrected past flaws, came out really impressive, and drew close attention to the semiconductor industry market.

    Over the past few years, we have seen greater consolidation in the mobile semiconductor industry. Companies such as Texas Instruments, which were once key players, no longer offer mobile SoC products. We see companies like Nvidia trying to occupy a large part of the market, but they are constantly failing. MediaTek tried to implement a high-end SoC with a line of Helio X chipsets with even less success, so I even had to suspend development in this segment in order to focus on more profitable P-series hardware.

    Even Samsung LSI, having a relatively good product of the flagship Exynos series, has so far failed to win the trust of its own mobile unit. Instead of using Exynos as the exclusive key component of the Galaxy series, Samsung turned to an external supplier and used Qualcomm's Snapdragon SoC. Based on the foregoing, it is safe to say that the production of competitive high-tech SoC and semiconductor components is indeed a difficult business.

    The Kirin 960 introduced last year was rather ambiguous: although the SoC showed noticeable improvements compared to the Kirin 950, it looked pale against the background of the success of competing flagships: SoC from Samsung and Qualcomm, since both of them had a significant advantage in the process. And the next release of Huawei flagships with the new generation of SoCs in the fourth quarter intersected with the release of Apple, in contrast to the usual for Qualcomm and Samsung first quarter.

    Thus, when we compare Kirin with Snapdragon and Exynos, we see a product that every now and then is late for a party in terms of introducing new technologies, such as new technological process and IP. Kirin 970 is also from this rank: as a 10 nm SoC based on the Cortex-A73, it lags behind Qualcomm and Samsung in terms of processor nodes, yet it came out too early regarding the ARM release schedule, which prevented him from implementing CPU cores based on DynamiQ, A75 and A55. This suggests that the Kirin 970 has only a few months of equal technical performance with the Snapdragon 835 and Exynos 8895, before we see the new Snapdragon 845 and Exynos 9810 products in the normal spring update cycle.
    Nevertheless, today's review is dedicated to Kirin 970 and its achievements, and also provides an opportunity to analyze the current position of SoCs used in Android devices.

    Kirin 970 did not show major IP improvements, as it continues to use the same ARM CPU as Kirin 960. The new SoC did not even increase the frequency of the CPU clusters, and we see the same 2.36 GHz for the A73 cores and 1.84 GHz in A53 cores. When ARM initially launched the A73, we saw a positive intention to increase the frequency to 2.8 GHz on TSMC 10 nm, which seems to have failed. This indicates the ever-increasing complexity of increasing the frequency in the mobile SoC, since the impact of updates to the processor node is becoming less and less.

    But the Kirin 970 processor demonstrates a significant revision and improvement of the graphics processor. We see the first implementation of the ARM Mali G72 in a 12-cluster configuration, and a 50% increase in the number of cores compared to the G71-MP8 in Kirin 960. The new GPU operates at a much lower frequency (746 MHz compared to 1033 MHz Kirin 960). In the Matt Humrick processor review, abnormally high Mali G71 average power levels were found, which heated up the Mate 9 smartphone case, so I hope the architectural improvements of the new G72, along with a wider configuration and lower frequency, will result in to a significant improvement over its predecessor.

    The new modem in the Kirin 970 now uses 3GPP LTE Release 13 and supports download speeds up to 1200 Mbps thanks to a carrier combination up to 5x20 MHz with 256-QAM, which makes the Kirin modem equivalent to Qualcomm X20, which will be integrated into Snapdragon 845.
    The biggest hype associated with Kirin 970 was around the embedded neuromorphic processor. NPU, like HiSilicon calls it, is a device of a new generation. It consists of specialized blocks designed to accelerate the "output" of the convolutional neural network (CNN). Around this news, there is already talk of "artificial intelligence" in the smartphone, but the correct term is machine learning or deep learning. The built-in hardware acceleration units from various manufacturers do not actually perform deep learning, but rather improve the performance (output) of neural network models, while model training will continue to be performed either in the cloud or by other SoC units, for example, GPU, but this is still a first look, but we will properly consider NPU in a specialized article.

    SoC matrix scheme, courtesy of TechInsights Mate 10 teardown

    As mentioned above, one of the major improvements to the Kirin 970 is the transition to the TSMC 10FF processor node. While it is believed that Samsung will use the 10 nm process for a very long time - we will see two more SoC generations released at 10LPE and 10LPP - TSMC uses a different approach and sees its 10FF processor node only as transitional to the expected 7FF node, which due later in 2018. Thus, the only mobile products of TSMC 10FF today are MediaTek X30 and Apple 10X, released in small volume last summer, as well as Apple A11 and HiSilicon Kirin 970, released in the third to fourth quarter, that is, 2-3 quarters later, as Samsung began mass production of the Snapdragon 835 and Exynos 8895.

    The expectations of HiSilicon from the new processor node are small. This is a fairly modest increase in efficiency by only 20% with the same performance of processor clusters, which is below 30% of the ARM predictions stated in previous forecasts. A very meager improvement in power was probably one of the reasons HiSilicon decided not to increase the processor frequency on the Kirin 970, and instead focused on reducing power consumption and reducing TDP compared to Kirin 960.

    The new SoC has noticeably reduced the size of the matrix from 117.72 mm² to 96.72 mm², although it has 50% more GPU cores, as well as new IP blocks, such as NPU. Our colleagues from TechInsights published a detailed comparison of the block sizes between Kirin 960 and Kirin 970, and we see a decrease in block size by 30-38% for apples-to-apples IP. The quad-core Cortex-A73 cluster now has a size of just 5.66 mm², which contrasts sharply with Apple, which uses twice the area of ​​silicon in its dual-core large processor cluster.

    Thank you for staying with us. Do you like our articles? Want to see more interesting materials? Support us by placing an order or recommending to friends, 30% discount for Habr's users for a unique analogue of the entry-level servers that we invented for you:The whole truth about VPS (KVM) E5-2650 v4 (6 Cores) 10GB DDR4 240GB SSD 1Gbps from $ 20 or how to share the server? (Available with RAID1 and RAID10, up to 24 cores and up to 40GB DDR4 RAM).

    Dell R730xd 2 times cheaper? Only we have 2 x Intel Dodeca-Core Xeon E5-2650v4 128GB DDR4 6x480GB SSD 1Gbps 100 TV from $ 249 in the Netherlands and the USA! Read about How to build an infrastructure building. class c using servers Dell R730xd E5-2650 v4 worth 9000 euros for a penny?

    Also popular now: