Like clockwork, Qualcomm has announced its Snapdragon 845 system on chip (SoC) and high-end Snapdragon platform which will power numerous high-end smartphones and possibly always-connected laptops throughout 2018.
As always, a major architecture change brings new performance, efficiency, and features. After a very comprehensive briefing at the Snapdragon Summit 2017, here is our overview of Snapdragon 845, split into different areas of interest: Camera, Performance, Battery life, Connectivity, Immersion, AI, and Security.
Let’s start with the best understood, most desirable feature for many users: the camera. With Snapdragon 845, camera performance will reach new levels, both in terms of raw capabilities, but also in image quality.
First, the camera will be capable of capturing 4K Ultra HD Premium HDR video. This is a mouthful, but it is similar to the feature-set used in Hollywood movies (1 Billion colors + HDR). Obviously, this is not exactly Cinema-quality, but in general, this is a huge improvement in video recording (16.7M colors, no HDR). Today’s best devices can “play” this kind of content, but Snapdragon 845 phones will be able to create that content.
“A HUGE IMPROVEMENT IN VIDEO RECORDING”
If you are unaware, Ultra HD Premium is not a technical standard, but rather a marketing umbrella that federates different TV makers around a certain quality level for 4K, HDR and color rendering. You will hear a lot about “Color Volume”. That is a 3D representation of the colors that devices can record (camera) and reproduce (display). The color volume can be visualized by using color gamut (reproducible colors) and luminance data (how bright tings are). The image below shows the larger color volume of Snapdragon 845 (Rec. 2020) vs, Snapdragon 835 (Rec. 709). This is leading-edge stuff.
Secondly, the camera ISP (Image Signal Processor) and overall subsystem are so fast that it is now possible to capture 720p video at 480 FPS. This will allow video creators to capture amazing slow-motion videos without having buffering limitations often encountered with this kind of extreme situations. For example, Sony has an extreme 960 FPS slow-motion recording, but only for a very short time. And yes, you can capture 1080p at 240FPS and 4K UHD HDR at 60FPS.
Still photos also benefit from the same resources allocated to image processing. Snapdragon 845 can process 16 Megapixel photos at a rate of 60 FPS (talk about bursting!). You might think of this as a video, but photos are treated differently because they are meant to be seen and saved as individual frames, while videos are an animated stream.
Snapdragon’s performance is due to various high-performance computing units, some with a distinct purpose such as the ISP (Image Signal Processor) or the GPU (graphics unit) and others with a slightly more general role such as the DSP (Digital Signal Processor) and of course the CPU cluster.
Snapdragon 845 is manufactured using Samsung’s second-generation 10nm process, which is an optimization of last year’s introductory 10nm semiconductor node. There are very small improvements in performance and efficiency derived from this.
Kryo 385 CPU
The CPU cluster is divided into two groups of four cores: four high-performance cores (ARM A75 derivatives, 2.8GHz max) to process workloads as fast as possible. There are also four other low-power cores (ARM A55 derivatives, 1.8 GHz max) to handle less intensive tasks with the best possible power-efficiency. You can read our article about ARM’s big-LITTLE architecture to understand more about heterogeneous CPU core clusters.
“DOUBLE-DIGITS CPU PERFORMANCE INCREASE: WORTHY OF ATTENTION”
Qualcomm did some custom work on the CPU cores and therefore named it Kryo 385. Qualcomm often tunes a specific core design or can build CPU cores from the ground up if necessary. This time, Qualcomm has done some custom work on the A75 and A55 designs from ARM but did not rebuild the cores from the ground up.
For Snapdragon 845, the engineers have added a 2MB L3 cache along with a 3MB “System Cache.” The main difference is that the L3 cache is only accessible to the CPU cluster, while the System Cache is accessible to other computing units as well (GPU, DSP, etc…). In both cases, the caches can drastically reduce the latency of getting information (from RAM) along with memory bus traffic.
With all of this, Qualcomm is expecting a ~25% CPU performance boost from last year’s Snapdragon 835 chip (see image above for details). I have to say, double-digits CPU performance increase: that is worthy of attention.
Adreno 630 GPU (Graphics Processing Unit)
As its name implies, this is the unit responsible for 3D gaming and VR performance. A GPU can also be used for artificial intelligence and other non-graphics workloads, but it’s fair to say that 3D graphics is the main purpose here.
At the moment, Qualcomm has not revealed the inner-workings of the GPU, so we don’t know exactly how many GPU cores there are, texture units, etc.… however a 30% graphics performance improvement should come out of this new design.
Additionally, Qualcomm also claims a 30% power-efficiency increase, which means that Snapdragon 845 can perform the same rendering while using 30% less energy as its predecessor. That is a rather large increase in both efficiency and speed.
Snapdragon 845 can support 2K per eyes in VR applications to improve one of VR’s pain point: the blockiness of the in-headset graphics. Incidentally, Qualcomm has also developed a Foveated rendering framework to help developers boost performance. I’m not 100% clear on what engines are supported, but the principle is sound can significantly accelerate pixel-level rendering. I previously explained how Foveated rendering works on Desktop PC, and this is the same principle, but in one sentence, Foveated rendering consists in reducing the detail away from your gaze point.
“THESE ARE CUTTING EDGE RENDERING TECHNIQUES”
To accelerate rendering, Qualcomm has added “multiview rendering,” a very effective technique that can reduce geometric (per-vertex) operations by nearly 50% in VR apps. The principle is simple: to achieve stereo rendering, the same scene is being rendered TWICE from slightly different points of views (LEFT+RIGHT eyes). Multiview rendering allows some computation to be shared between the LEFT/RIGHT frames because they are extremely similar (although not identical). There are other applications of similar multi-projection techniques, and they have proven to be highly efficient.
The graphics and extended reality framework of Snapdragon 845 allow it to support 6 Degrees of Freedom (DoF) and SLAM (Simultaneous Location And Mapping). SLAM is used to sense where you are in the room and track your motion in space by looking at your position relative to the environment. This is the foundational technology for 6 Degrees of Freedom without wires. Although some of this was available before, it is now fully baked into the Snapdragon platform and will be available to more OEMs.
The compound optimizations offer a very high potential for boosting VR (and AR) performance. These are cutting edge rendering techniques previously available only on desktop computers, or in a small number of engines. Qualcomm could democratize this further by supporting them in the Snapdragon Platform.
Hexagon 685 DSP
The DSP (Digital Signal Processor) is a specialized computing unit that can perform vector math on large quantities of data with extreme power-efficiency. It can be used for a vast array of things such as, but not limited to 2D image processing/effects, depth sensing for face-recognition, artificial intelligence inference and more.
People often think of the GPU when it comes to massive math workloads, and it is not a bad reflex. However, the DSP and the GPU are complementary tools that serve different situations. It is impossible to generalize completely, but DSPs are often more appropriate to use when you need to quickly process relatively small quantities of data. GPUs can be great at churning super-massive quantities of data, but they also require more setup and have more overhead. In reality, it’s not DSP vs. GPU — both can be extremely useful depending on the app.
Hexagon 685 is 3X faster than last year’s Hexagon 682. Without a doubt, Qualcomm must have increased the number of math units but hasn’t yet revealed the implementation details. This kind of performance increase matches the arrival of Google’s Neural API for Android 8.1+.
With battery life being the #1 concern for users, Power optimizations are always at the top of the list for mobile chip engineers. There are several ways to increase the power efficiency, and without a major manufacturing node change, it had to come from the architectural and software efforts.
The software can be optimized to offload work (when possible) to the most efficient sub-systems we discussed previously.
The addition of multiple cache layers reduces RAM access by as much as 75%, thus helping save power as well. The hardware has multiple clock domains and voltage domains, which means that the chip has several power islands that can use the most optimum amount of power at any given time. This leads to continuous savings if the tuning is done properly.
Snapdragon 835 was 50% more power-efficient as Snapdragon 821. A large chunk of this was due to the manufacturing node change. This time, the delta is not as large, but Eliane Fiolet checked power draw tests comparing Snapdragon 845, 835 and 660 and for the exact same task (4K video playback, 30FPS) Snapdragon 835 uses ~8% more power while Snapdragon 660 draws ~15% more power as Snapdragon 845. Not bad at all!
Earbuds battery savings
This time, Bluetooth (BT) will have the ability to broadcast to multiple devices at once under the Qualcomm TrueWireless name. You may not know, but BT wireless earbuds are currently set up to have one of them as the “master” that will orchestrate the communication with the phone, and with the other earbud. The master essentially does 2X more work as the other earbud, causing an imbalance in power consumption between the pair. That situation can be fixed by having the phone talk to both at the same time.
Last year’s world-class X16 LTE modem is replaced by a more advanced X20 model that can reach 1200 Mbps in theoretical peak speed (vs. 1000 before). That is of course if your wireless carrier supports it locally.
Even without an advanced network, Qualcomm uses MIMO (multiple inputs multiple output) techniques with up to four antennas that lead to large performance increases due to sheer parallelism on existing networks.
Qualcomm has added support for 60 GHz WiFi AD (802.11ad) which is a 4.6 Gbps short range WiFi connection ideal for large data transfer from within your network. Today many people still use wired Ethernet which is considered to be faster and more reliable. However, this technology could, in theory, replace wires within a 30-feet range. Keep in mind that the range is typically a distance without any physical obstacles that could disturb the signal.
Extended Reality tasks such as Virtual Reality (VR) and Augmented Reality (AR) are two of the most challenging tasks that the GPU can handle. In both cases, there are other units at play. Typically, the GPU would render the 3D scene or 3D overlay.
At the same time, the Hexagon 685 DSP could treat signals coming from the real world such as the room’s environment while the Spectra 280 ISP processes incoming data from the cameras, maybe to spot AR beacons. Qualcomm’s real technique for peak performance and power efficiency is to offload work to the best-suited, most power-efficient, unit – sometimes at the same time. If you think of it, game console programming is sometimes exactly like that. It’s about maximizing your hardware utilization.
Artificial Intelligence (A.I)
As of late, several silicon vendors have announced their flavor of “Neural Processors” with various marketing claims. However, the same OEMs have been dodgy when asked for details about the specs and capabilities of said neural computing units. To Qualcomm’s point, “neural processors” for mobile are almost always “DSP” units being re-marketed.
“A WIDE ARRAY OF AI TOOLS”
As we have seen earlier, the Qualcomm Hexagon 685 DSP is the 3rd generation that has been used in an AI context. The 300% performance increase from last year’s model shows that Qualcomm’s customers are taking AI very seriously and demanding more horsepower for that kind of workload.
However, Qualcomm’s AI effort doesn’t rely on a single unit. The DSP is surely the central nexus, but the embedded Adreno GPU supports FP16 numbers, and the CPU has been optimized to support 8-bit operations that are commonly used in AI. With these changes, developers have a wide array of AI tools at their disposal to optimize different types of workload.
Maybe the most obscure but important aspect of Snapdragon chips is security. First, it is important to know that security is not equal for all chips and that standards can be pretty loose (if there are any outside of payment). Not so long ago, some phones stored the fingerprint sensor data in the public area of the phone’s storage!
Previous generations of Qualcomm Snapdragon chips as a trusted execution zone which was secure. However, with the recent apparition of hacks for that kind of trusted environment, Snapdragon 845 has a new “Secure Processing Unit” which is more isolated from the rest of the SoC’s sub-system.
It is an island that has its memory, CPU and power gates. It can also generate random numbers internally (this is the base of all encryption) and its cryptographic engine. Services that need a very high level of security will go through this unit. Things like biometric data, for example, should transit through this unit to be processed so that only some kind of digital key (or Hash) is accessible by apps and services. A 3rd party should never see the biometric data.
Biometric data is not like a password. You cannot change your biometric data, and once stolen, you can never take it back, so it’s best to protect it as much as possible.