Qualcomm’s AI Research team is showing off some of the on-device AI capabilities of their latest chips.
First, Qualcomm is showing a demo of Large Language and Vision Assistant (LLaVa), a 7B parameter LMM that accepts multiple types of data inputs, running offline on an Android phone. In the demo, a person asks a chatbot what ingredients are shown in an image of some fruits and vegetables, then asks it to create a recipe with those ingredients and estimate the calorie count of the recipe.
Qualcomm is also showcasing a demo of Low Rank Adaptation (LoRA) on an Android phone. Running Stable Diffusion with a lightweight LoRA adapter enables changing the tone/personality of the model without needing multiple fine-tuned models. LoRA is not only useful for language vision models (LVMs) but also large language models (LLMs) as well to create things like more personalized assistants.
Finally, Qualcomm is also showcasing a 7B parameter LMM that can accept text and audio inputs running offline on a Windows PC. In the demo, a person uploads a sound byte and asks the chatbot to provide a detailed description of the audio, the environment it might’ve been recorded in, what the vibe is like, and what era the music is from. Another demo compares the performance of GIMP with a Stable Diffusion plugin on a Snapdragon X Elite PC versus a PC running an Intel Core Ultra 7 - 155H CPU.
First, Qualcomm is showing a demo of Large Language and Vision Assistant (LLaVa), a 7B parameter LMM that accepts multiple types of data inputs, running offline on an Android phone. In the demo, a person asks a chatbot what ingredients are shown in an image of some fruits and vegetables, then asks it to create a recipe with those ingredients and estimate the calorie count of the recipe.
Qualcomm is also showcasing a demo of Low Rank Adaptation (LoRA) on an Android phone. Running Stable Diffusion with a lightweight LoRA adapter enables changing the tone/personality of the model without needing multiple fine-tuned models. LoRA is not only useful for language vision models (LVMs) but also large language models (LLMs) as well to create things like more personalized assistants.
Finally, Qualcomm is also showcasing a 7B parameter LMM that can accept text and audio inputs running offline on a Windows PC. In the demo, a person uploads a sound byte and asks the chatbot to provide a detailed description of the audio, the environment it might’ve been recorded in, what the vibe is like, and what era the music is from. Another demo compares the performance of GIMP with a Stable Diffusion plugin on a Snapdragon X Elite PC versus a PC running an Intel Core Ultra 7 - 155H CPU.
🔥45👍13❤2🥰2
Qualcomm just announced the FastConnect 7900, their latest flagship connectivity chip for mobile devices. The most intriguing part about the FastConnect 7900 is that it integrates Wi-Fi, Bluetooth, and UWB in a single chip. This will help increase adoption of UWB, which to date is only available on select Android devices.
UWB, if you aren’t aware, stands for ultra-wideband and is a short range wireless communication technology that enables high precision locating. It’s commonly used for pinpointing lost items or to facilitate digital car keys.
The FastConnect 7900 also supports Bluetooth Channel Sounding, which is another high precision locating technology that enables estimating the distance between two Bluetooth LE devices. It’s much more precise than using signal strength measurements between two Bluetooth devices to determine location (RSSI).
Of course, it also supports key Wi-Fi technologies like Wi-Fi 7 and High Band Simultaneous, but it also has some new “AI-enhanced” Wi-Fi features. Qualcomm says the chip uses AI to “understand” the context of how you’re using your Wi-Fi connection without accessing any data or doing any deep package inspections so it can optimize Wi-Fi settings to best serve what you’re doing.
Qualcomm says the FastConnect 7900 will be fabricated on a 6nm process and will launch commercially in H2 2024.
UWB, if you aren’t aware, stands for ultra-wideband and is a short range wireless communication technology that enables high precision locating. It’s commonly used for pinpointing lost items or to facilitate digital car keys.
The FastConnect 7900 also supports Bluetooth Channel Sounding, which is another high precision locating technology that enables estimating the distance between two Bluetooth LE devices. It’s much more precise than using signal strength measurements between two Bluetooth devices to determine location (RSSI).
Of course, it also supports key Wi-Fi technologies like Wi-Fi 7 and High Band Simultaneous, but it also has some new “AI-enhanced” Wi-Fi features. Qualcomm says the chip uses AI to “understand” the context of how you’re using your Wi-Fi connection without accessing any data or doing any deep package inspections so it can optimize Wi-Fi settings to best serve what you’re doing.
Qualcomm says the FastConnect 7900 will be fabricated on a 6nm process and will launch commercially in H2 2024.
😍39👍15🤔5👏3👌1
Qualcomm just unveiled the Snapdragon X80, the company’s 7th-gen 5G modem-RF system that’ll ship in commercial devices in H2 2024 (likely in flagships powered by the next-gen Snapdragon 8 Gen 4 chipset).
The Snapdragon X80 is the first 5G modem with fully integrated NB-NTN (narrowband non-terrestrial network) satellite communications support, a 6-antenna architecture for smartphones, and 6X carrier aggregation.
Qualcomm says it has a tensor accelerator to power a few “AI” features, such as AI-based mmWave range extension which extends the range for fixed-wireless access applications. Other improvements include a 30% boost in location accuracy, reduction of 20% for best-cell selection time, 30% faster link acquisition, and 60% faster CPE service acquisition for mmWave connections.
The Snapdragon X80 is the first 5G modem with fully integrated NB-NTN (narrowband non-terrestrial network) satellite communications support, a 6-antenna architecture for smartphones, and 6X carrier aggregation.
Qualcomm says it has a tensor accelerator to power a few “AI” features, such as AI-based mmWave range extension which extends the range for fixed-wireless access applications. Other improvements include a 30% boost in location accuracy, reduction of 20% for best-cell selection time, 30% faster link acquisition, and 60% faster CPE service acquisition for mmWave connections.
👍40😱6👏4🤔2❤1🤮1😍1
This media is not supported in your browser
VIEW IN TELEGRAM
This media is not supported in your browser
VIEW IN TELEGRAM
This media is not supported in your browser
VIEW IN TELEGRAM
This media is not supported in your browser
VIEW IN TELEGRAM
This media is not supported in your browser
VIEW IN TELEGRAM
This media is not supported in your browser
VIEW IN TELEGRAM
Google’s latest Android Feature Drop for February 2024 is here, bringing you 9 new features to help you stay productive. These include:
1) You can now chat with Gemini right in the Google Messages app on Android! You can ask Gemini to do things like draft messages, come up with a joke, and more. This is available for beta users in English.
2) Android Auto can now use AI to summarize notifications. AI will also suggest relevant replies and actions, so you can just tap a button to send a message, share your ETA, or start a call. (This was announced at Galaxy Unpacked last month but should be rolling out now.)
3) Lookout’s Image Q&A feature is now rolling out globally in English! This feature left closed beta last September and was available in the US, UK, and Canada. Image Q&A uses AI to auto-generate descriptions of images.
4) Last October, Google launched Lens in Maps (formerly Search with Live View). This feature uses AI and augmented reality to help you find information about nearby stores. Now, this feature has been upgraded with screen reader support, so you can point your phone’s camera at your surroundings and have TalkBack read the place’s information out loud.
5) You can now add handwritten annotations to documents in Google Docs from your phone or tablet using your finger or stylus. You can choose from a selection of markup tools like different pen colors and highlighters.
6) Spotify Connect devices will finally appear in Android’s media output switcher! Google announced this way back at CES 2023, but it’s finally going to roll out starting this week.
7) The Fitbit app now pulls in your health and fitness data from other apps that support Android’s Health Connect platform. This means you can see your data from services like AllTrails, Oura Ring, and MyFitnessPal, all within the Fitbit app. In the “You” tab, you can see data from connected apps next to your Fitbit data. In the “Today” tab, you can see data like exercise, steps, calories burned, and more from Health Connect-compatible apps.
(1/2)
1) You can now chat with Gemini right in the Google Messages app on Android! You can ask Gemini to do things like draft messages, come up with a joke, and more. This is available for beta users in English.
2) Android Auto can now use AI to summarize notifications. AI will also suggest relevant replies and actions, so you can just tap a button to send a message, share your ETA, or start a call. (This was announced at Galaxy Unpacked last month but should be rolling out now.)
3) Lookout’s Image Q&A feature is now rolling out globally in English! This feature left closed beta last September and was available in the US, UK, and Canada. Image Q&A uses AI to auto-generate descriptions of images.
4) Last October, Google launched Lens in Maps (formerly Search with Live View). This feature uses AI and augmented reality to help you find information about nearby stores. Now, this feature has been upgraded with screen reader support, so you can point your phone’s camera at your surroundings and have TalkBack read the place’s information out loud.
5) You can now add handwritten annotations to documents in Google Docs from your phone or tablet using your finger or stylus. You can choose from a selection of markup tools like different pen colors and highlighters.
6) Spotify Connect devices will finally appear in Android’s media output switcher! Google announced this way back at CES 2023, but it’s finally going to roll out starting this week.
7) The Fitbit app now pulls in your health and fitness data from other apps that support Android’s Health Connect platform. This means you can see your data from services like AllTrails, Oura Ring, and MyFitnessPal, all within the Fitbit app. In the “You” tab, you can see data from connected apps next to your Fitbit data. In the “Today” tab, you can see data like exercise, steps, calories burned, and more from Health Connect-compatible apps.
(1/2)
❤64👍30🥱14🔥1
Mishaal's Android News Feed
Google’s latest Android Feature Drop for February 2024 is here, bringing you 9 new features to help you stay productive. These include: 1) You can now chat with Gemini right in the Google Messages app on Android! You can ask Gemini to do things like draft…
8) Passes from Google Wallet can now appear on your Wear OS smartwatch. This means you can access boarding passes, gym memberships, loyalty cards, and more from your watch. You can choose which passes to hide or display. (This has already started rolling out over the past month, as I first reported, but should now be available more widely.)
9) Finally, Google Maps for Wear OS now supports public transit directions, meaning you can use your smartwatch to look up departure times for buses, trains, or ferries and then get compass-guided navigation directions to your destination. You can also mirror directions from your phone to your watch if you prefer looking things up on your phone.
(2/2)
9) Finally, Google Maps for Wear OS now supports public transit directions, meaning you can use your smartwatch to look up departure times for buses, trains, or ferries and then get compass-guided navigation directions to your destination. You can also mirror directions from your phone to your watch if you prefer looking things up on your phone.
(2/2)
❤58👍27💩5👏2
This media is not supported in your browser
VIEW IN TELEGRAM
Wear OS’s new hybrid interface promises a big boost in power and performance for smartwatches.
Today, OnePlus unveiled the OnePlus Watch 2 powered by Google’s Wear OS 4 operating system. The OnePlus Watch 2 has a dual-chipset architecture that takes advantage of Wear OS’s updated hybrid interface to achieve a stunning up to 100 hours of regular use while in Smart Mode. (I can attest to the OnePlus Watch 2’s battery life being stellar, but that’s for a separate post!)
Most Wear OS smartwatches have a dual-chipset architecture wherein a more powerful applications processor (AP) is paired with an ultra low-power co-processor microcontroller unit (MCU). The AP handles complex tasks while less complex tasks are offloaded to the MCU. Google has managed to heavily optimize the interface between the MCU and AP, allowing for very seamless, intelligent switching. So for example, the AP can be put to sleep while the MCU handles sensor data processing.
Wear OS’s hybrid OS interface supports APIs for Display, Health Services, and Notifications. The best part for developers is that these optimizations are available without needing to do any code changes!
OnePlus uses the hybrid interface’s notification API to offload processing of bridged notifications on the MCU so the AP doesn’t need to be woken up. Users can read and dismiss these notifications while the watch is powered by the MCU, and this can include advanced features like quick replies or remote actions. App developers just need to use the Wear OS Notification API.
Things like precise workout tracking, automatic sports recognition, and health data monitoring can be offloaded to the MCU thanks to Wear OS’s updated hybrid OS interface. Apps that use the Health Services API can get periodic access to sensor data. The MCU batches sensor data to optimize power use.
Lastly, watch faces built on the Watch Face Format - a declarative XML format to configure the appearance and behavior of watch faces - can be rendered on the MCU thanks to the updated hybrid OS interface.
Today, OnePlus unveiled the OnePlus Watch 2 powered by Google’s Wear OS 4 operating system. The OnePlus Watch 2 has a dual-chipset architecture that takes advantage of Wear OS’s updated hybrid interface to achieve a stunning up to 100 hours of regular use while in Smart Mode. (I can attest to the OnePlus Watch 2’s battery life being stellar, but that’s for a separate post!)
Most Wear OS smartwatches have a dual-chipset architecture wherein a more powerful applications processor (AP) is paired with an ultra low-power co-processor microcontroller unit (MCU). The AP handles complex tasks while less complex tasks are offloaded to the MCU. Google has managed to heavily optimize the interface between the MCU and AP, allowing for very seamless, intelligent switching. So for example, the AP can be put to sleep while the MCU handles sensor data processing.
Wear OS’s hybrid OS interface supports APIs for Display, Health Services, and Notifications. The best part for developers is that these optimizations are available without needing to do any code changes!
OnePlus uses the hybrid interface’s notification API to offload processing of bridged notifications on the MCU so the AP doesn’t need to be woken up. Users can read and dismiss these notifications while the watch is powered by the MCU, and this can include advanced features like quick replies or remote actions. App developers just need to use the Wear OS Notification API.
Things like precise workout tracking, automatic sports recognition, and health data monitoring can be offloaded to the MCU thanks to Wear OS’s updated hybrid OS interface. Apps that use the Health Services API can get periodic access to sensor data. The MCU batches sensor data to optimize power use.
Lastly, watch faces built on the Watch Face Format - a declarative XML format to configure the appearance and behavior of watch faces - can be rendered on the MCU thanks to the updated hybrid OS interface.
👍54🔥22
The OnePlus Watch 2 is here, and it’s quickly become my favorite Wear OS smartwatch after only a few days of use! Its battery life has been STELLAR - it lasted me 75 hours during the first cycle! This is thanks to its huge battery (500mAh versus 306mAh in the Pixel Watch 2) and also Google’s new Wear OS hybrid OS interface.
I’ll share my full impressions on tomorrow’s episode of the Android Faithful podcast, but here’s a rundown of the key specs and features:
- Size: 47.0 x 46.6 x 12.1mm
- Weight: 49g (without strap)
- Build: Stainless steel chassis, 2.5D sapphire crystal face, IP68 rating, MIL-STD-810H certified
- Colors: Black Steel or Radiant Steel
- Strap: 22mm fluor rubber
- Display: 1.43” AMOLED round display, 60Hz refresh rate, up to 1000 nits peak brightness, 466x466 resolution (326 ppi)
- Chipsets: Qualcomm Snapdragon W5 Gen 1 + BES2700
- Positioning: Dual-frequency L1+L5 GNSS
- Sensors: Accelerometer, gyroscope, 8-channel optical heart rate, optical pulse oximeter, geomagnetic, light, barometer
- Connectivity: Wi-Fi + Bluetooth (no LTE)
- Battery: 500mAh. Up to 100 hours in Smart Mode. Up to 12 days in Power Saver Mode.
- Charging: VOOC fast charging (1 to 100% in 1 hour)
- RAM/ROM: 2GB + 32GB
- Software: Wear OS 4 based on Android 13 (W5 Gen 1) + custom RTOS (BES2700). 2 years of updates. Can be paired with Android 8.0+ devices only through the OHealth app.
With the launch of the OnePlus Watch 2, the OHealth app has been updated to support Android’s Health Connect by platform. This lets you share health data collected by OHealth with other apps that support Health Connect, like Fitbit. OHealth includes tracking modes for > 100 sports, detailed sleep tracking analysis, and more.
I’ll share my full impressions on tomorrow’s episode of the Android Faithful podcast, but here’s a rundown of the key specs and features:
- Size: 47.0 x 46.6 x 12.1mm
- Weight: 49g (without strap)
- Build: Stainless steel chassis, 2.5D sapphire crystal face, IP68 rating, MIL-STD-810H certified
- Colors: Black Steel or Radiant Steel
- Strap: 22mm fluor rubber
- Display: 1.43” AMOLED round display, 60Hz refresh rate, up to 1000 nits peak brightness, 466x466 resolution (326 ppi)
- Chipsets: Qualcomm Snapdragon W5 Gen 1 + BES2700
- Positioning: Dual-frequency L1+L5 GNSS
- Sensors: Accelerometer, gyroscope, 8-channel optical heart rate, optical pulse oximeter, geomagnetic, light, barometer
- Connectivity: Wi-Fi + Bluetooth (no LTE)
- Battery: 500mAh. Up to 100 hours in Smart Mode. Up to 12 days in Power Saver Mode.
- Charging: VOOC fast charging (1 to 100% in 1 hour)
- RAM/ROM: 2GB + 32GB
- Software: Wear OS 4 based on Android 13 (W5 Gen 1) + custom RTOS (BES2700). 2 years of updates. Can be paired with Android 8.0+ devices only through the OHealth app.
With the launch of the OnePlus Watch 2, the OHealth app has been updated to support Android’s Health Connect by platform. This lets you share health data collected by OHealth with other apps that support Health Connect, like Fitbit. OHealth includes tracking modes for > 100 sports, detailed sleep tracking analysis, and more.
👍57🤬7🤔1😱1
Mishaal's Android News Feed
The OnePlus Watch 2 is here, and it’s quickly become my favorite Wear OS smartwatch after only a few days of use! Its battery life has been STELLAR - it lasted me 75 hours during the first cycle! This is thanks to its huge battery (500mAh versus 306mAh in…
Fun fact: the OnePlus Watch 2 can be easily bootloader unlocked.
Just like with the Pixel Watch 2, the charging puck for the OnePlus Watch 2 supports USB data transfer. Unlocking it is as easy as toggling "enable OEM unlocking" in Developer Options, rebooting to the bootloader, then sending
The recovery interface isn't optimized for the watch, but the physical buttons do work as selectors.
I'd imagine most of you won't unlock the bootloader (I just wanted to see if it worked), but the fact that the puck supports USB data is nice because it also makes ADB access much easier.
I'll be sharing my full impressions of the OnePlus Watch 2 on tomorrow's episode of the Android Faithful podcast!
Just like with the Pixel Watch 2, the charging puck for the OnePlus Watch 2 supports USB data transfer. Unlocking it is as easy as toggling "enable OEM unlocking" in Developer Options, rebooting to the bootloader, then sending
fastboot flashing unlock.The recovery interface isn't optimized for the watch, but the physical buttons do work as selectors.
I'd imagine most of you won't unlock the bootloader (I just wanted to see if it worked), but the fact that the puck supports USB data is nice because it also makes ADB access much easier.
I'll be sharing my full impressions of the OnePlus Watch 2 on tomorrow's episode of the Android Faithful podcast!
👍55👀28🔥8❤6
This media is not supported in your browser
VIEW IN TELEGRAM
You can now set the Microsoft Copilot app as your default assistant app on Android!
With version 27.9.420225014, available in beta, you can change the default assistant app to Copilot. This lets you launch Copilot from any screen by swiping diagonally from the corner or long-pressing the power button.
However, it seems Copilot isn't taking full advantage of this, as invoking it just launches the main activity. It doesn't even automatically listen for input yet, nor does it use the APIs that would let it grab a screenshot of the current on-screen content. I wouldn't be surprised if Microsoft is working on these aspects, though.
H/T Bartok Adrian Gabriel on Telegram
With version 27.9.420225014, available in beta, you can change the default assistant app to Copilot. This lets you launch Copilot from any screen by swiping diagonally from the corner or long-pressing the power button.
However, it seems Copilot isn't taking full advantage of this, as invoking it just launches the main activity. It doesn't even automatically listen for input yet, nor does it use the APIs that would let it grab a screenshot of the current on-screen content. I wouldn't be surprised if Microsoft is working on these aspects, though.
H/T Bartok Adrian Gabriel on Telegram
👍73🤮52🎉8🆒3❤1
After being stuck on the July 2023 Google Play System Update for so long, the January 2024 GPSU is finally rolling out for Samsung Galaxy devices like the Galaxy S24!
Meanwhile, the February 2024 GPSU is slowly rolling out for some Pixel users.
Neither the January nor the February 2024 GPSU has any really interesting changes, but the March 2024 GPSU will have a pretty significant change. More on that soon 👀
(Thanks to ashutosh28 on Twitter for the tip!)
Meanwhile, the February 2024 GPSU is slowly rolling out for some Pixel users.
Neither the January nor the February 2024 GPSU has any really interesting changes, but the March 2024 GPSU will have a pretty significant change. More on that soon 👀
(Thanks to ashutosh28 on Twitter for the tip!)
👍72🥰7🫡5🔥1
Google is planning to push an update to Android phones that could dramatically improve AV1 video playback on budget devices. The March 2024 Play System Update may swap Google's libgav1 codec with VideoLAN's libdav1d codec.
You can read more about why this matters over on Android Authority.
Edit: A VLC for Android developer told me that the version of VLC I used back in November 2023 during testing actually shipped about a year old version of libvlc at the time of release. Plus, as the VideoLAN Twitter account points out, libdav1d 1.4.1 is coming soon which boosts AV1 decode performance even further!
So we're about to see an even bigger AV1 decode performance gain on many devices than I previously thought once this change rolls out!
You can read more about why this matters over on Android Authority.
Edit: A VLC for Android developer told me that the version of VLC I used back in November 2023 during testing actually shipped about a year old version of libvlc at the time of release. Plus, as the VideoLAN Twitter account points out, libdav1d 1.4.1 is coming soon which boosts AV1 decode performance even further!
So we're about to see an even bigger AV1 decode performance gain on many devices than I previously thought once this change rolls out!
Android Authority
Upcoming Android update will give AV1 videos a big boost on budget devices
Android currently ships with the libgav1 AV1 decoder, but a future update will switch that to libdav1d, which offers much better performance.
🔥108👍18👀10👏3
If you develop a watch face for Wear OS, heads up: Upcoming releases of Wear OS may require you to use the Watch Face Format (WFF) in order to "show data from most data sources in watch face complications."
In future releases, watch faces built on the legacy Jetpack Watch Face library or the Wearable Support Library will only be able to show basic data like the battery level, unread notification count, time, date, and an app shortcut.
Google sent an email to developers yesterday (h/t 9to5Google) which leads to the above snippet in the developer docs.
Watch face developers on Reddit point out that some features unique to legacy watch faces, like certain in-app features or complex animations, don't translate to the new WFF.
If I had to guess why Google is doing this, my bet is that this has to do with supporting the new hybrid OS interface that Google announced a few days ago. Watch faces built on the WFF are rendered on the ultra low-power MCU instead of the AP. It makes sense that heavily customized watch faces can't easily be made to operate on the MCU which runs an entirely different OS. Google wants to standardize watch faces so users don't have to worry about whether a watch face will heavily limit their smartwatch's battery life.
In future releases, watch faces built on the legacy Jetpack Watch Face library or the Wearable Support Library will only be able to show basic data like the battery level, unread notification count, time, date, and an app shortcut.
Google sent an email to developers yesterday (h/t 9to5Google) which leads to the above snippet in the developer docs.
Watch face developers on Reddit point out that some features unique to legacy watch faces, like certain in-app features or complex animations, don't translate to the new WFF.
If I had to guess why Google is doing this, my bet is that this has to do with supporting the new hybrid OS interface that Google announced a few days ago. Watch faces built on the WFF are rendered on the ultra low-power MCU instead of the AP. It makes sense that heavily customized watch faces can't easily be made to operate on the MCU which runs an entirely different OS. Google wants to standardize watch faces so users don't have to worry about whether a watch face will heavily limit their smartwatch's battery life.
🔥29👍14👏2🤔2❤1🤮1