At the Snapdragon Summit, Chris Patrick, SVP and GM, Mobile Handset, Qualcomm, unveiled the Snapdragon 8 Gen 2, Qualcomm’s flagship SoC for premium smartphones, and the successor of the Snapdragon 8 Gen 1 and the Snapdragon 8+ Gen 1.
System Performance: Kryo CPU Cluster
Let’s start with system performance, the standard hot topic for any new generation of processors. Qualcomm has improved all performance aspects over last year’s Snapdragon 8 Gen 1.
The new Kryo CPU cluster is now based on one ARM Cortex X3 (Prime, 3.19 GHz), four performance cores (2.8 GHz), and three efficiency cores (2.0 GHz).
Last year’s generation featured one ARM Cortex X2 (Prime, 3 GHz), three Cortex A710 (2.5 GHz), and four Cortex A510 (1.8 GHz), compatible with the ARMv9 Gen 1 CPU architecture.
The extra performance core and the new ARMv9 Gen 2 architecture drive most of the performance boost, while the minimal increase in CPU frequency also provides an extra nudge.
Qualcomm estimates this new CPU cluster to be 35% faster and up to 40% more power-efficient than the Snapdragon 8 Gen 1, the previous market leader for flagship Android smartphones.
The 40% power-efficiency number is surprisingly high, and Qualcomm claims it is due to the Kryo CPU’s new microarchitecture. Without a clearer explanation, I will assume it is based on the best case study, and upcoming benchmarks might shed more light too.
The new Kryo CPU features an 8MB L3 cache, a 33% increase over the 6MB L3 cache in the previous generation. Augmenting the cache size helps to avoid expensive DDR memory access whenever possible.
The new chip offers support for LP-DDR5x memory up to 4200 MHz and with up to 16 GB of memory density.
Artificial Intelligence (AI) Engine – Snapdragon Smart
One of the top areas “that will be critical to Snapdragon 8 Gen 2 is artificial intelligence. This platform is purpose-built for AI. […] At the heart of our AI experience is the Qualcomm AI Engine [which] is made of several technologies including Adreno GPU, Kryo, Sensing Hub, Memory, and Hexagon” said Francisco Cheng, Senior Managing Director, Product Marketing, Qualcomm, during our briefing.
Across all technologies involved in processing AI models, the Qualcomm AI Engine 8th generation improved the performance by up to 4.35X and the performance-per-watt by 60% over the 7th generation.
Hexagon, AI co-processor
Qualcomm defines Hexagon as a “fused AI accelerator architecture,” meaning that the Hexagon logic block is made of several types of computing units. Qualcomm made multiple improvements throughout Hexagon, and the most notables are the 2X size increase of the Tensor accelerator and the addition of a dedicated power delivery system.
The major changes in Hexagon were made to address the ability to run more complex neural networks, particularly around transformer networks, to deliver, among other things, multi-language translation in real-time.
Qualcomm’s Micro Tile Inferencing is a new technique that works by splicing neural networks into ‘tiles’ that will be dispatched to available compute units in Hexagon. Qualcomm did not go into much detail, but in our experience, tiling data is typically utilized to increase parallelism (process the tiles in parallel to maximize hardware utilization) and potentially increase cache coherency and memory bandwidth. Both can make a significant impact on overall performance.
Last year, Qualcomm added the support of mixed precision INT8+INT16, (automatic switch between INT8 and INT16 precisions) to its non-automatic mixed precision technology, named “all precisions support” (INT8, INT16, FP16).
In the new AI Engine, Qualcomm added the support of 4-bit Integer precision (INT4) in the “all precisions” mode, so select AI algorithms and neural network models can process faster across the Hexagon compute units. On top of that, INT4 allows Hexagon to run more models simultaneously (parallel computing).
Each AI model could use a different level of precision, as some AI-driven tasks will deliver good enough results using an INT4 low level of precision for faster processing, while others may require higher levels of precision such as INT8, INT16, FP16, and FP32 (Floating Point stored on 16 or 32 bits).
Qualcomm said that the new AI Engine uses INT4 in “cinematic mode” for real-time blurring of the background (video bokeh), with no visible difference in image quality. Previously similar AI models used INT8.
4th generation Sensing Hub with a dual-AI processor
The Sensing Hub is the independent low-power multi-core logic unit that functions when the rest of the SoC is asleep. It processes in real-time the contextual data streams such as voice, audio, sensors, and connectivity.
Last year, Qualcomm introduced a new feature for the 3rd generation Sensing Hub, the Always-on Camera, which has now become the Always Sensing Camera. The Always-on Camera allowed the phone to automatically lock when the user does not look at it or wake up each time the user glance at the sleeping phone.
With the new feature, the camera can recognize and process a QR code placed on the optical module when the smartphone is asleep (+display off), with the previous processor, the phone had to be turned on to manage a QR code.
Addition of a second AI processor with 50% more memory
Qualcomm has doubled the Sensing Hub’s AI computing capacity by adding a second AI processor. On top of that, 50% more internal memory is now available. The extra memory allows the system to run new AI models which were previously unavailable but also better versions of the ones that existed before.
Qualcomm mentioned 17 AI-based activities tracked by the new Sensing Hub, and the ones which caught our attention are [human] Activity Recognition, Voice Activation, and Eye-Tracking.
Hexagon Direct Link and Cognitive ISP
Qualcomm introduced Hexagon Direct Link for the first time, a feature that provides dedicated physical communication channels (bi-directional buses, in hardware lingo) to other logic blocks such as the GPU, the memory, and the Spectra ISP.
Alongside Hexagon Direct Link, the chip manufacturer announced its first Cognitive ISP, a compute unit located inside Hexagon that works in connection with the Spectra ISP for faster processing of the AI-driven imaging features.
Snapdragon Sight: upgraded camera performance with Cognitive-ISP
Cognitive ISP for better camera performance
The Cognitive ISP, located in the Hexagon AI co-processor, communicates in real-time with the Spectra ISP that processes the pixels captured by the camera. The Cognitive ISP executes the Multiple Semantic Segmentation AI model on the images sent by the Spectra ISP and tells the Spectra to apply different image-enhancing filters on the different identified zones of the image (up to eight segments).
The Cognitive ISP can locate boundaries and recognize objects in an image such as faces, bodies, skin, clothes, sky, plants, flowers, cats, beaches, buildings, and more.
This new feature, named Realtime Segmentation Filter, enables the system to precisely adjust colors, contrasts, and any aspects of photography to each segment of the image, in real-time. Developers can select different configurations and filters.
In other words, the Cognitive ISP is the brain, and the Spectra ISP is the muscle.
The new architecture allows the camera system to receive the segmentation filter a lot faster than before, and processes AI filters in real-time, even before the user presses the capture button, according to Qualcomm.
Graphics and Gaming – Snapdragon Elite Gaming
At the Snapdragon Summit, Chris Patrick, SVP and GM, Mobile Handset, Qualcomm, said that the new Adreno graphics processing unit (GPU) brings a 25% increase in graphics rendering performance than the previous generation while delivering 45% better power efficiency.
This means the baseline performance will increase for nearly all apps that depend on the GPU, like games.
Adreno GPU Display
The Adreno subvisual system also includes the display, and Qualcomm announced the support of HDR Vivid, a new HDR standard launched in March 2021 by the China UHD Video Industry Alliance (CUVA). HDR Vivid and Adaptive HDR are the latest additions to the other formats already supported by Adreno GPU Display, including HDR10, HDR10+, and Dolby Vision HDR.
Adreno GPU for games
Adreno GPU will be the first mobile platform to support Unreal Engine 5’s Metahuman, a framework that allows game developers to create ultra-realistic humans programmatically. Unreal Engine 5’s Metahuman characters were previously only usable in desktop games. That’s a testament to Qualcomm’s graphics software’s maturity.
The new-gen Adreno supports Vulkan 1.3 and delivers a 30% Vulkan performance boost compared to the previous generation.
Real-time hardware-accelerated ray tracing
Hardware ray tracing was the star feature of the show, as it pushes realism in games to the next level. Bringing real-time ray tracing to mobile gaming is a big deal, as it is compute-intensive, and only powerful gaming rigs can render the more advanced ray tracing techniques at high framerate, like global illumination for the whole 3D scene.
On Snapdragon 8 Gen 2, game developers can integrate basic ray tracing techniques such as soft shadows and accurate reflections. The traditional rendering methods such as shadow maps yielded hard shadows because they were not nearly as accurate compared to ray-traced shadows. Previous reflections rendering methods such as screen-space reflections were limited and could not reflect an object that was off-screen. Ray tracing can.
In short, ray tracing mimics the light’s behavior in the real world by following the path of light in a 3D world from the camera through the objects in the scene and to an eventual light source.
Dave Durnil, Global Head of Gaming & Snapdragon Studios, Qualcomm, showed us a cool demo from the popular game Justice Online Mobile with the ray tracing on and off. We could clearly see the difference, with ray tracing, the reflections in the water and on the curved objects were realistic and we could observe the soft shadows.
Snapdragon Sound: dynamic spatial audio with head tracking
Sarah McMurray, Staff Manager, Product Marketing, Qualcomm, introduced Spatial Audio with dynamic head-tracking, a new feature of the Snapdragon 8 Gen 2 that delivers more realistic 3D sound with low motion-to-sound latency. Spatial Audio is a technology dedicated to immersive gaming and multimedia experiences.
Connectivity: 5G AI and Wi-Fi 7 support
5G Modem-RF System
The 5G connectivity’s improvements revolve around three new features. The first is the integration of a “5G AI Processor” compute unit inside the 5G system. This essentially helps with “tuning” the 5G modem parameters to obtain optimal performance in any situation. It is the closest you can get to having a (human) modem engineer tweaking things constantly. AI could also help with signal error correction to avoid unnecessary data re-transmission.
Next, the 4X carrier aggregation further enlarges the data pipeline available for 5G communications, resulting in higher peak downloads (an impressive 10 Gbps).
Finally, it is possible to use two 5G SIM at the same time, thanks to the 5G Dual SIM Dual Active feature. It opens the door to even more data parallelism, but you might end up paying two subscriptions of course.
FastConnect Wi-Fi 7 and Bluetooth connectivity system
Snapdragon 8 Gen 2 comes with Wi-Fi 7, the upcoming standard for home connectivity. Qualcomm uses extremely advanced techniques such as high-band multi-link to use as much of the available spectrum as possible.
The company says its Wi-Fi can reach 5.8 Gbps and, more importantly, sustain a latency below two milliseconds in a sustained manner. Low latency could persuade wired ethernet users to finally switch to Wi-Fi.