Hailo's Latest Hailo-15 Chips Bring Up to 20 TOPS of Compute to Bear on Edge AI Image Workloads
With up to 20 tera-operations per second (TOPS) of compute, these tiny chips pack a punch well above their weight class.
Hailo, specialist in artificial intelligence at the edge, has announced the launch of its next-generation vision family: Hailo-15, able to run the ResNet-50 model on-device at an impressive 700 frames per second (FPS).
"Hailo-15 represents a significant step forward in making AI at the edge more scalable and affordable," claims Orr Danon, Hailo's chief executive officer, of the company's latest designs. "With this launch, we are leveraging our leadership in edge solutions, which are already deployed by hundreds of customers worldwide; the maturity of our AI technology; and our comprehensive software suite, to enable high performance AI in a camera form-factor."
To target a broader range of use-cases, the company has split Hailo-15 into three variants: Hailo-15H, Hailo-15M, and Hailo-15L. All support multiple input streams at up to 4k resolution, but compute performance from the AI core ranges from 7 tera-operations per second (TOPS) in the low-power Hailo-15L up to 20 TOPS in the Hailo-15H — five times better than the competition, Hailo claims, at the same price point.
Designed for smart camera systems, the flagship Hailo-15H is claimed to be able to run the ResNet-50 benchmark at 700 frames per second (FPS) entirely on-device — with Hailo claiming support for the YoloY5M6 object detection model in real time over a 1280×1280 input feed. The same compute capacity can also be turned, the company points out, on image enhancement tasks — with the vision subsystem offering three-exposure merging, wide dynamic range (WDR) support, noise reduction, and a 600 megapixel per second image processing rate at up to 12 megapixel resolution.
Elsewhere on the chip is a quad-core Arm Cortex-A53 CPU system, four lanes of PCI Express (PCIe) Gen. 3, gigabit Ethernet, and USB 3.1 Gen. 2 with Host and Device support plus USB 2.0 Host-only connectivity. Memory interfaces include LPDDR4/4X, SDIO, and quad-SPI (QSPI), with MIPI Camera Serial Interface (CSI) and Display Serial Interface (CSI), I2C, and I2S connectivity too.
"With Hailo-15, we’re offering a unique, complete and scalable suite of edge AI solutions," Danon adds. "With a single software stack for all our product families, camera designers, application developers, and integrators can now benefit from an easy and cost-effective deployment supporting more AI, more video analytics, higher accuracy, and faster inference time, exactly where they’re needed."
More information on the Hailo-15 family, along with a link to order a camera development kit based around the chip, is available on the Hailo website.