This article is more than 1 year old

Qualcomm serves up trio of new chips garnished with lavish ladles of AI gravy

First AI chip for cloud computing

AI Day Qualcomm has dropped the veil on a trio of new chips today, including two Snapdragon processors for smartphones and one AI accelerator for cloud services.

The chips were confirmed at Qualcomm's AI Day keynote speech in San Francisco. While the new Snapdragons are logical evolutions of the processor line, what's more interesting, perhaps, is its AI accelerator chip, which has been dubbed the Qualcomm Cloud AI 100.

It's the first time the semiconductor biz has ventured to build an AI chip for cloud computing and it claims the chip will max out at more than 350 tera operations per second (TOPS) at peak performance. That said, full performance benchmarks are still under review.

"There's a paradigm and architectural shift in infrastructure. AI training was originally done on CPUs, then FPGAs and GPUs,” said Keith Kressin, senior veep of Qualcomm. But now the market has gotten to the point that power is so important that the industry can build purpose built AI silicon."

Qualcomm was quick to show off the firms using Cloud AI 100 smarts - wheeling out businesses like Google, Facebook, Microsoft and Tencent who've been using it for applications such as self-driving cars, smart cameras, and virtual and augmented reality.

Not much detail was shed about the accelerator's specs, however, except that it's a 7nm architecture. When The Register asked for more information, Keith Kessen, SVP of product management at Qualcomm, told us the announcement was more of a "soft launch".

Arm's object detector core at work

Within Arm's reach: Chip brains that'll make your 'smart' TV a bit smarter

READ MORE

"There is a lot of noise about accelerators at the moment, but we aren't releasing the details just yet," Kessen said.

He did reveal that the chips would be sampled later this year and should be in production next year.

So, let's turn to the two mobile phone chips: the Snapdragon 665 and Snapdragon 730 are aimed at optimising AI, camera and gaming applications on Android handsets. Each system-on-chip comes packed with eight components: a main CPU core, a GPU, a visual subsystem unit, Wi-Fi connectivity, an audio and security unit, and a modem.

Here's a breakdown of the mobile platform specs, starting with the mid-range 665.

  • Qualcomm's Kryo 260 eight-core CPU system is an 11nm chip based on Arm's Cortex 64-bit A73 model running up to 2.0 GHz.
  • It also comes with an Adreno 610 GPU and a Hexagon 686 digital signal processor (DSP) to accelerate vector math calculations to run neural network models.
  • The AI engine includes Qualcomm's Neural Processing SDK to support models run in Caffe, Caffe 2 and TensorFlow across the CPU, GPU, or DSP to do things like voice recognition for digital assistants or image processing
  • The Adreno GPU also supports low-power graphics with Vulkan 1.1, OpenGL ES 3.2 and OpenCL 2.0
  • For the camera, there’s a separate Qualcomm Spectra 165 image processing unit that can capture 22MP at 30FPS for 4K video or up to 48MP for the single camera.
  • On the connectivity side of things, it uses Snapdragon X12 LTE for the modem, dual band Wi-Fi at 24GHz and 5GHz and a Bluetooth 5.0 stack.
  • Its Processor Security element handles the applications for any biometric sensing such as facial recognition or fingerprint detectors. It’s also where mobile payments are processed.
  • To top it all off, the SoC communicates with LPDDR3/LPDDR4x with up to 1866MHz at 8GB RAM.

Now, the high-range 730.

  • It has a slightly better eight-core 8nm CPU system. The Kryo 470 is based on Arm's Cortex 64-bit A76 blueprints at a clock speed of 2.2 GHz.
  • The AI engine here is mostly the same as the 655, packed with an Adreno 618 GPU, the Hexagon 686 DSP and Qualcomm's Neural Processing SDK, albeit with the better Kryo 470 CPU.
  • The camera is also superior compared to the 655 as it has Qualcomm Spectra 350 for image processing instead of the Spectra 165. Here, the dual camera can take 22MP 30FPS video at 4K HDR and the single camera can snap 48MP with multi frame noise reduction or up to 192MP.
  • It's slightly faster too with a X15 LTE modem, similar Wi-Fi at 24 GHz and 5GHz and Bluetooth 5.0 included.
  • Again, on the graphics rendering side, it supports an Adreno 618 GPU, with OpenGL ES 3.2, OpenCL™ 2.0 FP and Vulkan 1.1.
  • It also has the same processor security hardware and software to read irises, fingerprints and fight malware.
  • The general specs are slightly different, it talks to a 2x16-bit, LPDDR4 also clocking it at 1.866GHz with 8GB of RAM.

And finally, there's the 730G, a variant of the Snapdragon 730 designed for gaming. The hardware specs between both SoCs are almost identical, except for the fact that the 730G has better HD display support for devices with larger screens.

All these specs have been focused to support a wide range of computer vision algorithms for more camera functions. For example, you can take a single photo and process it in three modes: a close up telephoto portrait, a wide-angle pose, and a super-wide angle one with its 5X zoom.

There's also the bother of handling 3D facial recognition to unlock phones as well as the usual detecting faces and objects in photos with bounding boxes, some style transfer modes, augmented reality, and augmented reality. ®

More about

TIP US OFF

Send us news


Other stories you might like