On the Snapdragon Summit, Chris Patrick, SVP and GM, Cellular 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 Efficiency: Kryo CPU Cluster
Let’s begin with system efficiency, the usual scorching subject for any new technology of processors. Qualcomm has improved all efficiency facets over final 12 months’s Snapdragon 8 Gen 1.
The brand new Kryo CPU cluster is now primarily based on one ARM Cortex X3 (Prime, 3.19 GHz), four efficiency cores (2.8 GHz), and three effectivity cores (2.0 GHz).
Final 12 months’s technology featured one ARM Cortex X2 (Prime, 3 GHz), three Cortex A710 (2.5 GHz), and 4 Cortex A510 (1.8 GHz), suitable with the ARMv9 Gen 1 CPU structure.
The additional efficiency core and the brand new ARMv9 Gen 2 structure drive a lot of the efficiency increase, whereas the minimal enhance in CPU frequency additionally gives an additional nudge.
Qualcomm estimates this new CPU cluster to be 35% quicker and as much as 40% extra power-efficient than the Snapdragon 8 Gen 1, the earlier market chief for flagship Android smartphones.
The 40% power-efficiency quantity is surprisingly excessive, and Qualcomm claims it’s as a result of Kryo CPU’s new microarchitecture. With out a clearer rationalization, I’ll assume it’s primarily based on the perfect case research, and upcoming benchmarks may shed extra gentle too.
The brand new Kryo CPU options an 8MB L3 cache, a 33% enhance over the 6MB L3 cache within the earlier technology. Augmenting the cache dimension helps to keep away from costly DDR reminiscence entry at any time when doable.
The brand new chip provides help for LP-DDR5x reminiscence as much as 4200 MHz and with as much as 16 GB of reminiscence density.
Synthetic Intelligence (AI) Engine – Snapdragon Sensible
One of many high areas “that will likely be essential to Snapdragon 8 Gen 2 is synthetic intelligence. This platform is purpose-built for AI. […] On the coronary heart of our AI expertise is the Qualcomm AI Engine [which] is made from a number of applied sciences together with Adreno GPU, Kryo, Sensing Hub, Reminiscence, and Hexagon” mentioned Francisco Cheng, Senior Managing Director, Product Advertising and marketing, Qualcomm, throughout our briefing.
Throughout all applied sciences concerned in processing AI fashions, the Qualcomm AI Engine eighth technology improved the efficiency by as much as 4.35X and the performance-per-watt by 60% over the seventh technology.
Hexagon, AI co-processor
Qualcomm defines Hexagon as a “fused AI accelerator structure,” that means that the Hexagon logic block is made from a number of varieties of computing items. Qualcomm made a number of enhancements all through Hexagon, and essentially the most notables are the 2X dimension enhance of the Tensor accelerator and the addition of a devoted energy supply system.
The key adjustments in Hexagon have been made to tackle the power to run extra advanced neural networks, notably round transformer networks, to ship, amongst different issues, multi-language translation in real-time.
Qualcomm’s Micro Tile Inferencing is a brand new approach that works by splicing neural networks into ‘tiles’ that will likely be dispatched to accessible compute items in Hexagon. Qualcomm didn’t go into a lot element, however in our expertise, tiling knowledge is usually utilized to extend parallelism (course of the tiles in parallel to maximise {hardware} utilization) and doubtlessly enhance cache coherency and reminiscence bandwidth. Each could make a major impression on general efficiency.
Final 12 months, Qualcomm added the help of blended precision INT8+INT16, (computerized swap between INT8 and INT16 precisions) to its non-automatic blended precision expertise, named “all precisions help” (INT8, INT16, FP16).
Within the new AI Engine, Qualcomm added the help of 4-bit Integer precision (INT4) within the “all precisions” mode, so choose AI algorithms and neural community fashions can course of quicker throughout the Hexagon compute items. On high of that, INT4 permits Hexagon to run extra fashions concurrently (parallel computing).
Every AI mannequin may use a distinct degree of precision, as some AI-driven duties will ship adequate outcomes utilizing an INT4 low degree of precision for quicker processing, whereas others could require larger ranges of precision resembling INT8, INT16, FP16, and FP32 (Floating Level saved on 16 or 32 bits).
Qualcomm mentioned that the brand new AI Engine makes use of INT4 in “cinematic mode” for real-time blurring of the background (video bokeh), with no seen distinction in picture high quality. Beforehand comparable AI fashions used INT8.
4th technology Sensing Hub with a dual-AI processor
The Sensing Hub is the unbiased low-power multi-core logic unit that capabilities when the remainder of the SoC is asleep. It processes in real-time the contextual knowledge streams resembling voice, audio, sensors, and connectivity.
Final 12 months, Qualcomm launched a brand new function for the third technology Sensing Hub, the At all times-on Digicam, which has now turn out to be the At all times Sensing Digicam. The At all times-on Digicam allowed the cellphone to mechanically lock when the person doesn’t take a look at it or get up every time the person look on the sleeping cellphone.
With the brand new function, the digicam can acknowledge and course of a QR code positioned on the optical module when the smartphone is asleep (+show off), with the earlier processor, the cellphone needed to be turned on to handle a QR code.
Addition of a second AI processor with 50% extra reminiscence
Qualcomm has doubled the Sensing Hub’s AI computing capability by including a second AI processor. On high of that, 50% extra inner reminiscence is now accessible. The additional reminiscence permits the system to run new AI fashions which have been beforehand unavailable but in addition higher variations of those that existed earlier than.
Qualcomm talked about 17 AI-based actions tracked by the brand new Sensing Hub, and those which caught our consideration are [human] Exercise Recognition, Voice Activation, and Eye-Monitoring.
Hexagon Direct Hyperlink and Cognitive ISP
Qualcomm launched Hexagon Direct Hyperlink for the primary time, a function that gives devoted bodily communication channels (bi-directional buses, in {hardware} lingo) to different logic blocks such because the GPU, the reminiscence, and the Spectra ISP.
Alongside Hexagon Direct Hyperlink, the chip producer introduced its first Cognitive ISP, a compute unit positioned inside Hexagon that works in reference to the Spectra ISP for quicker processing of the AI-driven imaging options.
Snapdragon Sight: upgraded digicam efficiency with Cognitive-ISP
Cognitive ISP for higher digicam efficiency
The Cognitive ISP, positioned within the Hexagon AI co-processor, communicates in real-time with the Spectra ISP that processes the pixels captured by the digicam. The Cognitive ISP executes the A number of Semantic Segmentation AI mannequin on the photographs despatched by the Spectra ISP and tells the Spectra to use completely different image-enhancing filters on the completely different recognized zones of the picture (as much as eight segments).
The Cognitive ISP can find boundaries and acknowledge objects in a picture resembling faces, our bodies, pores and skin, garments, sky, vegetation, flowers, cats, seashores, buildings, and extra.
This new function, named Realtime Segmentation Filter, allows the system to exactly alter colours, contrasts, and any facets of pictures to every phase of the picture, in real-time. Builders can choose completely different configurations and filters.
In different phrases, the Cognitive ISP is the mind, and the Spectra ISP is the muscle.
The brand new structure permits the digicam system to obtain the segmentation filter rather a lot quicker than earlier than, and processes AI filters in real-time, even earlier than the person presses the seize button, in accordance with Qualcomm.
Graphics and Gaming – Snapdragon Elite Gaming
On the Snapdragon Summit, Chris Patrick, SVP and GM, Cellular Handset, Qualcomm, mentioned that the brand new Adreno graphics processing unit (GPU) brings a 25% enhance in graphics rendering efficiency than the earlier technology whereas delivering 45% higher energy effectivity.
This implies the baseline efficiency will enhance for almost all apps that rely upon the GPU, like video games.
Adreno GPU Show
The Adreno subvisual system additionally contains the show, and Qualcomm introduced the help of HDR Vivid, a brand new HDR normal launched in March 2021 by the China UHD Video Business Alliance (CUVA). HDR Vivid and Adaptive HDR are the most recent additions to the opposite codecs already supported by Adreno GPU Show, together with HDR10, HDR10+, and Dolby Imaginative and prescient HDR.
Adreno GPU for video games
Adreno GPU would be the first cell platform to help Unreal Engine 5’s Metahuman, a framework that enables sport builders to create ultra-realistic people programmatically. Unreal Engine 5’s Metahuman characters have been beforehand solely usable in desktop video games. That’s a testomony to Qualcomm’s graphics software program’s maturity.
The brand new-gen Adreno helps Vulkan 1.3 and delivers a 30% Vulkan efficiency increase in comparison with the earlier technology.
Actual-time hardware-accelerated ray tracing
{Hardware} ray tracing was the star function of the present, because it pushes realism in video games to the subsequent degree. Bringing real-time ray tracing to cell gaming is an enormous deal, as it’s compute-intensive, and solely highly effective gaming rigs can render the extra superior ray tracing strategies at excessive framerate, like world illumination for the entire 3D scene.
On Snapdragon 8 Gen 2, sport builders can combine primary ray tracing strategies resembling comfortable shadows and correct reflections. The standard rendering strategies resembling shadow maps yielded exhausting shadows as a result of they weren’t almost as correct in comparison with ray-traced shadows. Earlier reflections rendering strategies resembling screen-space reflections have been restricted and couldn’t replicate an object that was off-screen. Ray tracing can.
Briefly, ray tracing mimics the sunshine’s habits in the true world by following the trail of sunshine in a 3D world from the digicam by way of the objects within the scene and to an eventual gentle supply.
Dave Durnil, International Head of Gaming & Snapdragon Studios, Qualcomm, confirmed us a cool demo from the favored sport Justice On-line Cellular with the ray tracing on and off. We may clearly see the distinction, with ray tracing, the reflections within the water and on the curved objects have been practical and we may observe the comfortable shadows.
He instructed us that Warfare Thunder Edge and Blue Reflection video games for Android will function Adreno-accelerated ray tracing. How briskly is the brand new GPU? Examine our Snapdragon 8 Gen 2 hands-on benchmarks!
Snapdragon Sound: dynamic spatial audio with head monitoring
Sarah McMurray, Employees Supervisor, Product Advertising and marketing, Qualcomm, launched Spatial Audio with dynamic head-tracking, a brand new function of the Snapdragon 8 Gen 2 that delivers extra practical 3D sound with low motion-to-sound latency. Spatial Audio is a expertise devoted to immersive gaming and multimedia experiences.
Connectivity: 5G AI and Wi-Fi 7 help
5G Modem-RF System
The 5G connectivity’s enhancements revolve round three new options. The primary is the mixing of a “5G AI Processor” compute unit contained in the 5G system. This primarily helps with “tuning” the 5G modem parameters to acquire optimum efficiency in any scenario. It’s the closest you will get to having a (human) modem engineer tweaking issues continuously. AI may additionally assist with sign error correction to keep away from pointless knowledge re-transmission.
Subsequent, the 4X service aggregation additional enlarges the info pipeline accessible for 5G communications, leading to larger peak downloads (a powerful 10 Gbps).
Lastly, it’s doable to make use of two 5G SIM on the identical time, because of the 5G Twin SIM Twin Lively function. It opens the door to much more knowledge parallelism, however you may find yourself paying two subscriptions in fact.
FastConnect Wi-Fi 7 and Bluetooth connectivity system
Snapdragon 8 Gen 2 comes with Wi-Fi 7, the upcoming normal for dwelling connectivity. Qualcomm makes use of extraordinarily superior strategies resembling high-band multi-link to make use of as a lot of the accessible spectrum as doable.
The corporate says its Wi-Fi can attain 5.8 Gbps and, extra importantly, maintain a latency under two milliseconds in a sustained method. Low latency may persuade wired ethernet customers to lastly swap to Wi-Fi.
Filed in . Learn extra about Editorspick, Processors, Qualcomm, Snapdragon, Snapdragon Summit and SoC.
