Accelerator card by Untether AI delivers up to two Peta OPS for AI inference

AI workloads for inference require increasing amounts of compute resources, far outstripping the gains available to traditional CPU and GPU architectures, says Untether AI. It has unveiled the tsunAImi accelerator cards which is powered by the runAI devices. Untether AI says its use of at-memory computation, breaks through the barriers of traditional von Neumann architectures, to offer industry-leading compute density with power and price efficiency.

AI accelerators are used in data centres and Untether AI says that it focuses on inference acceleration, transferring the weights and activations between external memory, on-chip caches to the computing element to maximise power efficiency. Untether AI is able to deliver two Peta operations per second (OPS) in a standard PCI-Express card form factor.

“For AI inference in cloud and datacenters, compute density is king, said Arun Iyengar, CEO of Untether AI. “Untether AI is ushering in the Peta OPS era to accelerate AI inference workloads at scale with unprecedented efficiency,” he added.

The tsunAImu accelerator card is based on the company’s runAI200 devices. They are tailored for inference acceleration and operate using integer data types and a batch mode of 1. The at-memory compute architecture features a memory bank of 385kbytes of SRAM with a 2D array of 512 processing elements. There are 511 banks per chip and each device offers 200Mbyte of memory and operate at up to 502 Tera OPS in “sport” mode. It may also be configured for maximum efficiency, offering eight TOPs per watt in “eco” mode.

The runAI200 devices are manufactured by Untethered AI, using a cost-effective, mainstream 16nm process.

The tsunAImi accelerator cards are powered by four runAI200 devices, providing more than two times any currently available PCIe cards, says Untethered AI. This compute power translates into over 80,000 frames per second of ResNet-50 v 1.5 throughput at batch=1. This is three times the throughput of its nearest competitor. For natural language processing, tsunAImi accelerator cards can process more than 12,000 queries per second (qps) of BERT-base, four times faster than any announced product.

The Untether AI imAIgineTM software development kit (SDK) provides an automated path to running networks with push-button quantisation, optimisation, physical allocation and multi-chip partitioning. The imAIgine SDK frees data scientists from having to perform low-level optimisation tasks and instead spend time in developing models. The imAIgine SDK also provides an extensive visualisation toolkit, cycle-accurate simulator, and an easily integrated runtime API.

The imAIgine SDK is currently in early access with select customers and partners. The tsunAImi accelerator card is sampling now and will be commercially available in Q1 2021.

http://www.untether.ai

> Read More

Pre-certified modules accelerate time to market, says Silicon Labs

Bluetooth, Zigbee, Thread and multi-protocol modules from Silicon Labs reduce time-to-market for secure smart home, building and industrial automation applications.

The pre-certified wireless modules are claimed to be the only modules in the industry with full stack support for multi-protocol solutions to enable commercial and consumer IoT applications. y offer package options and integrated device security.

“Silicon Labs has mastered the challenges that come with adding wireless connectivity to IoT devices,” said Matt Saunders, vice president of IoT marketing and applications at Silicon Labs. The modules allow IoT device makers to get pre-certified and secure wireless devices quickly to market.

The integrated modules are available in multiple package options including system in package (SiP) and traditional PCB versions. SiP modules contain miniaturised components that remove the need for complicated RF design and certification by opening space-constrained IoT designs to module-based solutions, says Silicon Labs. PCB modules enable flexible pin access and additional options to extend RF performance.

The xGM210PB features Secure Vault, the ARM PSA Level 2-certified IoT device security, as well as dynamic multi-protocol support for Bluetooth, Zigbee, and OpenThread and supports Wi-Fi co-existence. It is optimised for IoT applications, such as connected lighting, gateways, voice assistants and smart meter in-home displays, xGM210PB modules offer up to +20dBM output power and better than -104dBm sensitivity.

The BGM220 is one of the world’s smallest Bluetooth modules, says Silicon Labs, and the first to support Bluetooth Direction Finding. The integrated, pre-certified Bluetooth 5.2 module delivers up to 10 years’ operation on a coin cell battery, making it suitable for appliances, asset tags, beacons, portable medical, fitness, and Bluetooth mesh low-power nodes.

The MGM220 is a low-power, low-cost module for lighting controls, building and industrial automation sensors as well as Zigbee Green Power and energy harvesting applications.

Finally, the BGX220 Xpress includes an on-board Bluetooth stack, Xpress command interface and pre-programmed cable replacement firmware to deliver serial to Bluetooth LE with no need for firmware development. The modules are suitable for industrial applications including human machine interface (HMI) devices, addressing size, safety, and environmental challenges.

The modules are supported in the free-of-charge Silicon Labs Simplicity Studio 5 integrated development environment (IDE), which has software stacks, application demos and mobile apps, network analyser and patented energy profiler. Developers can also use simple APIs from Silicon Labs Xpress modules.

Samples, production quantities and development kits for all four modules are available now.

http://www.silabs.com

> Read More

Xilinx introduces Zynq RFSoC DFE for mass 5G radio deployments

Adaptive radio platforms are flexible for evolving 5G standards and with a hardened radio digital front end for performance, power, and cost effectiveness, says Xilinx.

The Zynq RFSoC DFE adaptive radio platform is designed to meet the evolving standards of 5G NR wireless applications. It combines hardened digital front end (DFE) blocks and adaptable logic to build low power, cost-effective 5G NR radio solutions for use cases ranging across 5G low-, mid-, and high- band spectrums. According to Xilinx, it offers the best balance of technologies between the cost economies of an ASIC which uses hardened blocks and the flexibility, scalability, and time-to-market benefits of a programmable and adaptive SoC.

5G radio must meet bandwidth, power, and cost challenges for widespread deployment, but must also adapt to the three key 5G use cases, i.e. enhanced mobile broadband, massive machine type communication, and ultra-reliable low-latency communication. They must also scale for evolving 5G standards such as OpenRAN (O-RAN) and new, disruptive 5G business models, Xilinx advises. Zynq RFSoC DFE integrates hardened DFE application-specific blocks for 5G NR performance and power savings yet also offers the flexibility to integrate programmable adaptive logic to enable a futureproof solution for evolving 5G 3GPP and O-RAN radio architectures.

“For the first time, Xilinx is providing a wireless radio platform with more hardened application-specific IP than adaptive logic to address low power and low cost 5G requirements,” said Liam Madden, executive vice president and general manager, Wired and Wireless Group at Xilinx. As the 5G landscape continues to evolve, it is imperative that integrated RF solutions are adaptable to address future standards. “Zynq RFSoC DFE provides the optimal balance between that adaptability and fixed function IP,” said Madden.

Zynq RFSoC DFE offers 2X performance-per-watt compared to the earlier generation of adaptive computing and scales from small cell to massive MIMO macrocells. It is claimed to be the industry’s only direct RF platform that enables carrier aggregation/sharing, multi-mode, multi-band 400MHz instantaneous bandwidth in all FR1 bands, and emerging bands up to 7.125GHz. When used as a mmWave intermediate frequency transceiver, it provides up to 1,600MHz of instantaneous bandwidth.

The Zynq RFSoC DFE architecture allows customers to either bypass or customise the hard IP blocks. For example, they can leverage Xilinx’s field-proven DPD that supports existing and emerging GaN power amplifiers or insert their own DPD IP.

Zynq RFSoC DFE design documentation and support is available to early access customers, with shipments expected during the first half of 2021.

http://www.xilinx.com  

> Read More

Integrated ToF module is multi-zone

Offering an increased camera field of view (FoV) coverage and spatial resolution, the VL53L5 FlightSense time of flight (ToF) sensors are believed to be the first 64-zone devices. The sensors from STMicroelectronics breaks a scene into separate areas to help an imaging system build the most detailed spatial understanding of a scene, the company says.

It has a 940nm vertical cavity surface emission laser (VCSEL) light source, an SoC sensor integrating a VCSEL driver, the receiving array of 40nm single photon avalanche diodes (SPADs), and a low-power 32-bit MCU core and accelerator running firmware.

The SPAD array can be set to favour spatial resolution, where it outputs all 64 zones at up to 15 frames per second, or to favour maximum ranging distance, where the sensor outputs 4×4/16 zones at a frame rate of 60 frames per second.

The miniature module contains optical elements in the receive aperture to create the 64 ranging zones. Range is 4m.

Eric Aussedat, general manager of ST’s imaging division, said that the VL53L5 delivers “64x more ranging zones than previously available . . . [and] performance improvement in laser auto focus, touch-to-focus, presence detection, and gesture interfaces while helping developers create even more innovative imaging applications.”

ST’s ToF technology includes human-presence detection to control the wake up and hibernation of laptops or monitors and laser autofocus in hybrid focusing algorithms for smartphone cameras. The auto focus feature is embedded in most of the highest-ranking smartphone cameras according to DXOMARK, an independent benchmark that assesses image quality.

Laser auto focus assures quick, accurate focusing in low-light scenes or when capturing low-contrast targets.

ST says that key smartphone and PC platform suppliers have pre-integrated the sensor onto platforms. Android and Windows device drivers are widely available for the FlightSense modules.

The VL53L5 is packaged in a 6.4 x 3.0 x 1.5mm module and integrates both transmit and receive lenses and has and expanded FoV of 61 degrees diagonal. This wide FoV is especially suited to detect off-centre objects and ensure perfect auto focus in the corners of the image.

For laser auto focus, the VL53L5 gathers ranging data from up to 64 zones across the full FoV.

ST’s architecture can automatically calibrate each ranging zone and direct ToF technology allows each zone to detect multiple targets and reject reflection from the cover-glass. FlightSense also gathers the raw data collected by the SPAD array and performs post processing, via a proprietary, embedded MCU and accelerator before transferring the ranging data to the system host over an I2C or a SPI bus. This removes the need for a specific camera interface and powerful receiver MCU.

The VL53L5 retains the Class 1 certification of all ST’s FlightSense sensors and is fully eye-safe for consumer products. It is in mass production with millions of units already shipped to leading wireless and computer manufacturers, says ST.

http://www.st.com

> Read More

About Smart Cities

This news story is brought to you by smartcitieselectronics.com, the specialist site dedicated to delivering information about what’s new in the Smart City Electronics industry, with daily news updates, new products and industry news. To stay up-to-date, register to receive our weekly newsletters and keep yourself informed on the latest technology news and new products from around the globe. Simply click this link to register here: Smart Cities Registration