Every year, Qualcomm releases the newest iteration of Snapdragon, the company’s smartphone platform. This year’s launch event, the Snapdragon Technology Summit, took place in Maui, Hawaii last week. One thing I noticed was that there was a stronger focus this year on connectivity and AI than there has been in the past. I don’t think Qualcomm is just posturing to align with industry trends. Rather, it is trying to define itself as the go-to company OEMs turn to for best-in-class connectivity and AI capabilities. Let’s take a closer look.
The Snapdragon 855 is Qualcomm’s first 7nm SoC. It features an integrated 4G LTE Snapdragon X24 modem, and is the company’s first platform with the ability to support multiple-gigabit 5G, with its optional Snapdragon X50 modem. This means the 855 will have some of the fastest cellular connectivity on earth. In addition to blazing fast cellular connectivity, the Snapdragon 855 has a Wi-Fi 6 ready capability, which translates to multi-gigabit connectivity with the upcoming next generation of Wi-Fi technology (it’s worth noting that this standard is not fully certified yet, so the Wi-Fi solution might not be fully Wi-Fi 6 compatible with all features when it is finalized). In addition to current 2.4 Ghz and 5.8 GHz Wi-Fi, the Snapdragon 855 also includes the next generation of WiGig with its 60 GHz 802.11ay solution. All of these capabilities make the Snapdragon 855 the first mobile chip for 5G, 2 Gbps LTE, Wi-Fi 6 and 60 GHz 802.11ay—an undisputed connectivity powerhouse.
Traditional processor performance (CPU and GPU)
The Snapdragon 855 has an 8-Core Kryo 485 processor, with a cluster of 4 high-performance A76-based cores (including a single high frequency core for single threaded applications) and 4 low power A55-based cores. Though the new Snapdragon 855 doesn’t have a much higher clock speed than the previous generation, Qualcomm claims a 45% performance boost on the CPU. A lot of this is due to the larger caches and major architectural changes. This represents the biggest leap for a Snapdragon’s CPU performance, ever. The Adreno 640 is the fastest GPU that Qualcomm has ever made for smartphones, with a purported 20% performance improvement over the Snapdragon 845.
All these performance figures from Qualcomm are for sustained performance, rather than peak performance—the metric by which many of its competitors measure their performance numbers. Since people run smartphones for hours at a time, sustained performance is vastly more important than peak performance (which might last a few seconds or minutes). This has always been Qualcomm’s approach to performance. It deserves kudos for providing real world performance figures, not just something someone threw in the freezer and benchmarked for a few minutes.
Adding More AI and CV Performance
The Hexagon 690 and Spectra 380 are two parts of the SoC where Qualcomm inserted additional intelligence. The Hexagon 690 was redesigned from the ground up to include new Tensor cores in addition to the improved scalar and vector cores. This allows for applications to target the right type of core for the right inference job. In addition to the new Tensor accelerator, Qualcomm claims 1.2x scalar and 2x vector extensions performance improvements over the previous generation. Qualcomm also put 50% more ALUs in the Adreno GPU to improve FP32 and FP16 performance, as well as support for INT 16 and INT 8 in the Hexagon cluster. Even the CPU gained new dot product instructions for FP32 and INT8 instructions.
Qualcomm also added fixed accelerated CV capabilities into its Spectra 380 ISP which makes it the first CV-ISP. This accelerates some of the most popular CV functions of the ISP while reducing power consumption. All these improvements yield a 3x AI performance improvement over the previous generation of Snapdragon platforms, and 2x the competition, according to Qualcomm. Qualcomm seriously beefed up the AI capabilities of this smartphone platform, and as a result I suspect that we’ll see a lot more utilization of AI in next year’s phones.
The Snapdragon 855 features the newest 4th generation of Qualcomm’s AI Engine, which incorporates the CPU, GPU, Hexagon and all types of cores inside of the Hexagon. This includes support for runtime software frameworks like Tensorflow Lite for mobile devices, Google NN API, and the Neural Processing SDK. The NN frameworks that Qualcomm supports are also important because they determine how many different frameworks can be applied by different developers like TensorFlow, Pytorch, ONNX, Caffe2, Mxnet, Chainer, PaddiePaddie, and Cognitive Toolkit.
Based on Qualcomm’s benchmarks, the total AI performance of the Snapdragon 855’s AI Engine is 7+ TOPS, or 7 trillion operations per second. Google is already validating Qualcomm’s claims with its own benchmarks for performance and energy efficiency. Google reports that it reduced latency by 3x and improved efficiency by 3.7x. These kinds of improvements not only make AI perform faster, but use less power. This makes implementing AI in applications even more of a no-brainer for OEMs and application developers. Qualcomm also showed a couple really impressive live demos with some of its AI partners, including AI noise reduction during phone calls and aggressive noise reduction. Both of those demos surpassed my expectations, and I hope to see them adopted by OEMs.
Devices for next year
Qualcomm and their partners didn’t pre-announce any devices at the Snapdragon Tech Summit where the Snapdragon 855 was unveiled, however OnePlus was the only smartphone vendor to present and announced that they would be the first to market with the chip. If OnePlus launches before 5G networks are expected to launch, we could see something like a OnePlus 7 with the Snapdragon 855 and a OnePlus 7T later in the year with 5G capability later in the year, following the company’s annual cadence.
One interesting thing about the Snapdragon 855 is that it’s the first time Qualcomm has released a flagship phone chip that’s really only going to be used in phones. In the past, chips like the Snapdragon 835 found its way into VR HMDs and Windows PCs, but now Qualcomm has specific SoCs for those purposes. I see this as a good thing because it allows Qualcomm to focus on certain capabilities that are beneficial to smartphone OEMs and their users rather than building a single chip for so many markets. I expect we could still see some HMD makers build devices with the 855, but that’s not the purpose of this platform.
All these major improvements to AI and connectivity demonstrate Qualcomm’s priorities and where it believes that its OEMs will benefit the most next year. With such massive improvements to connectivity and AI, next year’s Android phones should feel noticeably faster than even this year’s phones. Battery life was not discussed much, but we will probably have to wait for 5G networks roll out to see real-world performance and battery life measurements. It’s being hinted at that we’ll see early Snapdragon 855 devices without 5G connectivity so we might see Snapdragon 855-based devices before the big 5G networks light up in the first half of 2019.