In brief Samples of Qualcomm's Cloud AI 100, teased more than a year ago, are now shipping to select customers, the tech giant said this week. The hardware is due to officially launch in the first half of next year.
The Cloud AI 100 is a system-on-chip designed to accelerate machine-learning workloads at the edge of networks, in data centers, and in 5G infrastructure – the idea being running neural networks right where decisions need to be made rather than piping data back and forth between remote systems. It can handle INT8, INT16, FP16, and FP32 vector math in hardware; has up to 144MB of on-die SRAM; has up to 16 AI-focused processor cores; and supports LPDDR4x RAM and PCIe 3 and 4.
The hardware can be obtained in various form factors: a dual M.2 format for edge that has a 15W TDP and up to 70 TOPS performance; a dual M.2 format that's 25W and up to 200 TOPS; and a HHHL PCIe card that's 75W and up to 400 TOPS. Each card can have up to 32GB of RAM.
Accompanying this is a Cloud AI 100 Edge developer kit, shipping next month, for those who want to experiment with the idea of machine-learning applications at the edge of networks before deploying silicon in anger. The dev kit features a Snapdragon 865 system-on-chip, 12GB of LPDDR5 RAM, a Cloud AI 100 DM.2 board, a Snapdragon X55 5G modem, Ethernet, Wi-Fi, and NVMe SSD storage.
IBM spins out AI toolkit collection as open source
IBM has open sourced its PowerAI software under the moniker Open-CE, or the Open Cognitive Environment. It is a suite of open-source deep-learning tools, such as TensorFlow and PyTorch, all packaged up to make it a bit easier and less of a faff to train and use models.
Big Blue also this week followed up the release of its A2I POWER processor core blueprints with its out-of-order cousin, the A2O – that's a capital 'o' at the end, not a zero. This 64-bit CPU core is written in Verilog and prioritizes performance-per-thread over getting as many threads as possible into a chip.
Tesla driver accused of sleeping at the wheel as Autopilot went nuts
A 20-year-old Canadian bloke was caught napping at the wheel of his speeding Autopilot-controlled Tesla Model S, Alberta Mounties said this week.
The driver and his passenger were both asleep with their chairs reclined as the flash motor sped along a highway between Edmonton and Calgary at more than 140 KMPH (87 MPH). When a cop car gave chase and turned its flashing lights on, other cars moved out of the way. But the Tesla's Autopilot seemingly interpreted this as a clear road, and increased its speed to precisely 150 KMPH, the police claim. The limit is 110 KMPH (68 MPH).
The British Columbia man was charged with speeding and dangerous driving. "Nobody was looking out the windshield to see where the car was going," RCMP Sgt Darrin Turnbull told CBC News. "I've been in policing for over 23 years and the majority of that in traffic law enforcement, and I'm speechless.
"We believe the vehicle was operating on the autopilot system, which is really just an advanced driver safety system, a driver assist program. You still need to be driving the vehicle." Amen to that.
Stock exchange launches money-laundering detector AI
Nasdaq said this week it will deploy AI to help banks investigate money laundering scams. The software will sift through alerts of suspicious activities – as many as 300,000 a month – and pick out the ones that look most likely like wrongdoing.
“We have been thinking long and hard on how we want to go beyond trade surveillance,” said Nasdaq senior veep Valerie Bannert-Thurner. “It’s a product launch but strategically it’s a launch beyond trade surveillance. We have great ambitions in the space.” ®