The smartphones SoC market doesn’t change much in recent years. Qualcomm is the absolute leader with a 41% market share, according to the CounterPoint Research data for the Q3, 2017. And though it’s down by 1% in comparison to last year’s result, it’s still the highest position. Qualcomm is followed by Apple (21%), MediaTek (18%), Samsung (8%), HiSilicon (6%), Spreadtrum (5%), and others (1%). The thing is that Apple’s chips are used in iPhones only. The same is true for HiSilicon as well. But Huawei has designed and announced the world’s first mobile chip with NPU, the Kirin 970. So there is every reason to think its market share will change in 2018. Samsung has brought the Exynos 9810 to the Galaxy S9 recently. This chip yields Qualcomm’s Snapdragon 845 chip by 8% only. And finally, a few days back MediaTek uncovered its own AI mobile chip, the Helio P60. As you can see, all big players have already announced their latest flagships SoCs. We just have to find out which of them is better. Though it will be very difficult to understand which of them leads the list, we can find out their advantages, at least.
What’s a SoC?
Not that much earlier we have been using a special term to describe the ‘brain of a digital device or a computer’. It was the CPU (Central Processing Unit). A regular CPU has been accompanied with various chips like GPU (Graphics Processing Unit), memory controllers, specialized video and audio chips (DSP), and many more. As for now, they all are integrated into one chip. That’s called a system-level integration. Therefore, the term of SoC (System-on-a-Chip) means a single chip including a lot of different functions.
SoC = CPU + GPU + memory controller + DSP + network radio + …
At the beginning of the article, we also introduced the main players in the smartphones chip makers. The only one we didn’t pay attention to is Xiaomi. Yes, this manufacturer is known for multiple types of products, but it’s begun making its own mobile chips named Surge.
As for graphics cards, there are 3 major manufacturers in face of ARM, Qualcomm, and Imagination. The former one comes in with Mali GPUs, the second vendor makes Adreno series GPUs, and the third one comes with PowerVR GPUs.
The next important portion is the network radio. Recently, on February 14, Qualcomm demonstrated brand new technologies such as spectrum sharing and URLLC (ultra-reliable low latency communication). They should boost the speeds with the recently approved 5G NR standards. Thanks to the X24 gigabit LTE modem, the download speed reaches up to 2Gbps. Before that, the fastest downlink speed belonged to Huawei Kirin 970 with 1.17Gbps. Although, the previous-gen Snapdragon 835 should provide a 1.2Gbps speed theoretically.
Well, once we know the main characteristics of a SoC, that’s the right time to dive into details of each of the newly launched flagship SoCs. We have no preferences concerning these chips. That’s why we put them in an announcement time order.
Huawei Kirin 970
Generally, there are two truly vertically integrated mobile OEMs who have full control over their silicon, Apple, and Huawei. But the latter is more integrated due to in-house modem development. Huawei’s semiconductor division, HiSilicon, is the only one that could break in into the high-end market. The Kirin 970 is the best example of that. It runs on a 10nm process node and adopts DynamiQ and A75 and A55 based CPU cores. Thus it is way better than the Snapdragon 835 and Exynos 8895. Of course, the next-gen Kirin SoC will be more competitive to the SND845 and Exynos 9810. But at this moment, the flagship chip of the Chinese manufacturer is the Kirin 970, and we should compare it to the current models.
Actually, the Kirin 970 doesn’t come with a complete IP overhaul. It continues using the same central processing unit IP from the ARM that was used in the Kirin 960. Moreover, the core frequency remains the same as well, 2.36GHz for the A73 cores and 1.84GHz for the A53 cores. But due to some improvements, the performance was improved by 20%.
However, the GPU is not identical to the previous-gen model. This is the first implementation of ARM’s Mali G72 in a 12-cluster configuration, a 50% increase in core count over the Kirin 960’s G71-MP8 setup. This graphics card runs at a lower frequency of 746MHz in comparison to the Kirin 960’s 1033MHz. So the architectural improvements, a wider and lower clocked configuration, and the new manufacturing process bring many improvements over its predecessor.
The chip comes with a built-in upgraded self-research dual-camera ISP supporting AI scene recognition, face recognition, sports scene detection, while night shooting performance has been improved as well. The video support reached to HDR10 for the first time, and it also supports support 4K @ 60fps video decoding, and 4K @ 30fps video encoding.
The Huawei Kirin 970 is the world’s first built-in independent NPU (Neural network unit) AI mobile computing platform. AI only allows devices to use the whole power but also to use cloud technology for bringing more massive information and services. So it’s quite reasonable why Huawei is currently focusing on collaboration AI with cloud computing.
Every user has different habits when interacting with a phone. AI will learn them, understand you, and help you by providing more accurate information and services. But it will also need too much information. The latter should be stored somewhere. The cloud storing is the most convenient solution. Mobile AI = On-Device AI + Cloud AI.
Actually, cloud storages have been used actively in recent years. We know there are too many issues related to them. They mainly refer to the real-time performance, stability, privacy, etc. But seems Huawei has found the better way of using this technology. Unlike the server-side AI design, the Kirin 970 chose a highly energy-efficient heterogeneous computing architecture to dramatically increase AI’s computing power to respond to the data center. HiSilicon has integrated the HiU mobile computing architecture with the dedicated hardware processing unit named the NPU (Neural Processing Unit). That’s why the AI performance is way better than the one provided by ‘regular’ SoCs.
At last, the Kirin 970 uses more advanced 4.5G LTE technology to support the world’s highest LTE Cat.18 connectivity and to achieve the industry’s highest 1.2Gbps download rate thanks to up to 5x20MHz carrier aggregation with 256-QAM. There are also built-in TEE and inSE security engines with higher security options.
Samsung Exynos 9810
Samsung announced its next flagship chip at CES 2018 in January. It is aiming at Apple’s A11 Bionic and other top-notch chips. It is manufactured using Samsung’s latest 10nm LPP FinFET technology. Actually, this is the fastest processor ever made by Samsung. It’s even faster than the Exynos 8895 found on the Samsung Galaxy S8. As the manufacturer claims, the performance is boosted by 10%, while the power efficiency is increased by 15%.
As for architecture, the Samsung Exynos 9810 is powered by the company’s third-gen custom CPU cores clocked at 2.9 GHz. They are accompanied by four ARM’s latest Cortex-A55 cores optimized for energy efficiency and clocked at 1.9GHz. The CPU has a wider pipeline and improved cache memory. That leads to notable performance improvements. This resembles ARM’s cutting-edge DynamIQ cluster configuration. The latter introduces the option of a shared L3 cache between cores and a private L2 cache for each core. But the South Korean manufacturer didn’t disclose much about it. So we don’t know whether it is using an architecture license for its M3 core or it has developed another in-house interconnect system or a hybrid design with separate clusters. Anyway, the M3 offers a two-fold single-core performance increase over the M2. By the way, the company doesn’t miss the chance to talk about this. Generally, the single-core performance is boosted by 2 times, the multi-core performance by 40%, and the GPU performance by 20%. As for a regular user, this means the apps will open faster, multitasking will be smoother, and the battery will serve longer.
Everything has been changed in the Exynos 9810. This refers to the graphics card. There is an integrated ARM Mali-972 GPU that supports for 10-bit HEVC and VP9 codecs. Therefore, the 4K videos will occupy less storage space, but thanks to clever post-processing, can contain more details. Interestingly, the core count has dropped from 20 to 18, but the clock speed has been improved a lot up from 546MHz to somewhere in the 700MHz.
Like all of the current flagship chips, the Samsung Exynos 9810 supports higher resolution photography, performs better in low-light shots with less image noise. It features a dual-ISP, which supports up to 4 four separate image sensors.
The Exynos 9810 uses a neural network-based deep learning for 3D face scanning as well as recognizing objects and people in photos. This is almost the same technology found on the Kirin 970. But the manufacturer doesn’t risk to call it AI. Instead, it says the Exynos 9810 introduces ‘sophisticated features to enhance user experiences with neural network-based deep learning’. This simply means it utilizes both hardware and software to speed up these workloads. Something like this we have seen on the Exynos 8895, which packed in a Vision Processing Unit (VPU). The latter was used specifically for analyzing data taken from the phone’s camera. Seems it is used in the 9810 as well. Thanks to it, the Samsung Galaxy S9 (the only smartphone with this chip) supports image recognition, depth sensing, and 3D hybrid face detection features. Of course, the facial recognition and iris scanning aren’t anything new. But with the launch of the Exynos 9810, the scanning speed and accuracy have been improved a lot.
At last, the Exynos 9810 packs in a 1.2 Gbps LTE modem, which comes with a Category 18 LTE supporting for 6x carrier aggregation (uplink speeds up to 200Mbits/sec.), 4×4 MIMO, 256-QAM, and eLAA technology.
Qualcomm Snapdragon 845
The Snapdragon 845 is a big move in terms of SoC architectures as it’s the first to employ ARM’s DynamiQ CPU cluster organization. As AnandTech explains, the DynamiQ enables the various different CPU cores within a SoC to be hosted within the same cluster and cache hierarchy, as opposed to having separate discrete clusters with no shared cache between them (with coherency instead happening over an interconnect such as ARM’s CCI). This major transition is probably the largest to date that we’ve seen in modern mobile smartphone ARM consumer SoCs.
Qualcomm’s latest flagship chip is based on the second-gen 10LPP process from Samsung. Say. Its predecessor, the Snapdragon 835 was manufactured on 10LPE. But Samsung had promised a 10% improved performance at the same power level or a reduced power consumption of 15% at the same performance.
As for architecture, the Snapdragon 845 is the world’s first SoC based on a DynamIQ big.LITTLE CPU organization. Plus, it comes with ARM’s Cortex A75 and Cortex A55 IPs. The Kryo 385 gold/performance cluster runs at up to 2.8GHz. Thus this chip is 14% better than the Snapdragon 835. As for the silver/efficiency cluster performance, it runs at 1.8GHz, which is a bit slower than the A53’s on the Snapdragon 835.
Of course, Qualcomm has brought a new generation of the Adreno GPU, Adreno 630. This switch marks the transition consisting of larger architectural changes. Say, the performance is increased by 30%, while the power-efficiency by 30%.
The Snapdragon 845 provides an upgraded multimedia performance with a VR support. As the manufacturer claims, it’s now able to drive dual-2400x2400p120 displays for VR headsets. As for the video recording capabilities, the SND845 supports for a maximum video encoding framerate at 4K. Qualcomm advertised support for Rec.2020 wide gamut color space recording, meaning we now see full support for HDR10 HEVC recording.
Additionally, the newest 280 ISP comes with an interesting feature called multi-frame noise reduction (MFNR), which is used in Google’s HDR+ proprietary processing. It captures numerous pictures in fast succession. Thereafter, it applies an algorithm for noise reduction in higher quality fashion compared to traditional single-frame noise reduction, which can introduce blurriness.
Qualcomm like Samsung doesn’t hurry to say this is an AI mobile chip. But actually, it is. The Snapdragon 845 comes with an upgraded DSP 685. The improvements refer to the power and performance, especially for AI and imaging tasks. Thanks to them the SND845 has achieved up to 3x increase in performance compared to the Snapdragon 835. However, traditional DSP architecture chips are not optimized well for neural network processing in comparison to the SoCs with a dedicated NPU. Probably, this is the main reason why Samsung and Qualcomm are more humble.
Qualcomm’s current flagship chip integrates the X20 modem, which elevates the LTE UE Category to 18 as it’s now capable of 5xCA as opposed to 4xCA on the X16 modem integrated into the Snapdragon 835. Theoretically, the download speed is 1.2Gbps when in a 5x20MHz downstream carrier aggregation mode
MediaTek Helio P60
One of the biggest advantages of the Helio P60 is the use of the 12nm process node from TSMC. As the Taiwanese manufacturer claims, it will continue using the 12nm manufacturing process for all chips launched in 2018. But it will switch to the impressive 7nm process node in 2019. Thus, there are no plans concerning the 10nm process node. Actually, the P60 is the first mobile chip of the manufacturer to come with TSMC’s new 12nm 12FFC process node. The latter is a shrink of 16FFC and. Therefore, it’s reasonable why it uses the same tools. This manufacturing process allows competing with the 10FF thanks to the power efficiency.
The Helio P60 is the first P-line chip to come with ‘big’ microarchitecture cores. It uses four Cortex A73 cores at up to 2.0GHz and other four power-efficient A53 cores clocked at the same frequency. It uses, so-called, a big.LITTLE configuration. So the tasks that require more power can use the A73 cores, while those that don’t can use A53, preserving battery life. They deliver a total of 70% CPU performance boost if compared to the previous P23 and P30, while the power consumption is improved by 25%.
The Helio P60 integrates the latest generation ARM Mali G72 MP3 GPU, albeit with three cores and clocked at 800MHz. thanks to it a good gaming performance is provided, undoubtedly. However, that’s a big move from the previous G71 to the newest G72. That’s why the gaming efficiency is improved by 25%.
The SoC also consists of a Triple-core ISP and a dual-core APU (AI Processing Unit). So it can power all the AI specific features, as well as AR and VR. First, as you can see, there are too many camera improvements. The three ISPs support dual cameras of up to 20MP + 16MP, or a single 32MP sensor. A single camera at 16MP resolution is capable of recording videos at up to 90fps, which means that you’ll be able to record 4K video at a high frame rate on a device that’s not too expensive.
Second, the Helio P60 is the first SoC featuring a multi-core AI processing unit. So in this sense, we can state the P60 will bring AI performance to the mid-range market. Yes, many think the P60 is a mid-range chip. But as it comes too many high-end features, we decided to put it next to the real flagship chips. Till the date, MediaTek hasn’t disclosed the IP provider but say’s it’s a dual-core implementation offering 2x140GMACs. However, the chip maker is going to offer a custom software stack called NeuroPilot AI technology, which will be compatible with Google’s NN API and frameworks such as TensorFlow, TF Lite, Caffe, and Caffe2.
The presence of APU works identically to the Kirin 970’s NPU. It provides things like object recognition and behavior-informed performance enhancement. As the company claims, the APU doubles the efficiency handled by the GPU. Generally, this new technology enables original equipment manufacturers to deliver AI accelerated features, maximized performance, energy efficiency, facial recognition, real-time beautification, general AI and machine learning tasks, and photography enhancements to more devices without the price markup.
In terms of network radio, the Helio P60 comes with the same modem as the P30. Thus it supports LTE up to Category 7 with 3x20MHz CA and 300Mbps download speeds and Category 13 with 2x20MHz CA for up to 150Mbps upload speeds.
At last, the P60 scores 1524 points and 5871 points in a single-core and multi-core GeekBench tests, respectively. The first devices to sport this chip will appear on the market in early Q2 2018
Again, we should note the MediaTek Helio P60 is not a high-end chip like the rest of SoCs we were talking about. But it’s the first AI chip of the chip maker, and it enters a competition want it or not. So in this sense, the P60 and Kirin 970 are more attractive, as they have a dedicated NPU. But Qualcomm is an absolute leader in the market, and its new chip will appear on more devices. Therefore, the manufacturer has more chances to learn what users want and want to bring to the market. Samsung’s Exynos-series chips can be found on a few models. Moreover, it’s assumed the Exynos 9810 won’t appear on other devices but the Samsung Galaxy S9. So we can make only one conclusion – AI is gaining momentum, and the chip makers just have to switch to it. But which of them will provide a better performance still remains a big question.