Plug-and-play AI acceleration with Thunderbolt

Published:

Topics: Open FPGA IP, Open ASICs

The flexible Thunderbolt standard is taking consumer devices by storm, and its availability beyond the x86 architecture is an interesting development that carries significant implications for industrial systems as well. Based on our experience in building various edge-to-cloud AI applications, we have devised a technological concept that helps us deliver extremely flexible and scalable AI-compute applications using Thunderbolt, with several industrial customer projects involving the design already being underway at Antmicro.

AI acceleration with Thunderbolt concept graph

Thunderbolt’s capabilities

Thunderbolt is a high-speed, versatile interconnect that provides a throughput of as much as 40Gbps over medium distances. It also lets you daisy-chain up to six devices to a host, while the included PCI Express interface enables adding external GPUs, AI accelerators or FPGA-based systems. Type-C USB, which is the most common connector/cable standard used by Thunderbolt devices, supports bi-directional power transfer, allowing the host to charge the peripheral device, and vice versa. Most importantly perhaps, once widely adopted, USB-C with Thunderbolt will enable plug-and-play connection of various peripheral devices to a system, increasing its modularity, scalability and portability.

Scaling on the fly

The plug-and-play functionality will enable users to quickly scale up their system by hot-swapping AI-accelerators or other components, such as hard drives, network cards, FPGAs, or SDI/HDMI grabbers. The ability to add a computing unit to a system without the need to power off the device can be crucial in many industrial environments, where the continuous operation of a system must be ensured.

Antmicro’s projects often involve emerging standards and new computing platforms alongside programmable logic and chip technologies. This makes us well-positioned to take advantage of the imminent proliferation of Thunderbolt, and build dedicated Thunderbolt-compatible accelerators - be it GPU-based ones or AI ASIC / FPGA accelerators - for your industrial device.

Eliminating trade-offs

Systems that integrate a CPU with a full-size GPU are typically not suitable for embedded applications due to their size, varying shapes, and different cooling requirements, which makes it hard to accommodate the elements in the machine without compromising its efficiency. Particularly, use cases in areas such as robotics could benefit from bigger freedom of arranging individual system components within or even outside the device, without adding latency to the communication between them. Our solution scratches that itch, eliminating the trade-off between performance and portability.

Lightning-fast

Thunderbolt can be used to build highly mobile, flexible, and modular devices.
The direct PCIe connection makes it possible to decouple the accelerator and the computing unit, add a cable between them, and arrange in the most suitable way. This way system architects can distribute the volume and weight of those components across the machine, making it portable, easy to enhance with AI capabilities, and effective in harsh industrial conditions.

Antmicro’s solutions making use of Thunderbolt achieve high flexibility and functionality that translates into better performance and results while operating in the field. Reach out to us at contact@antmicro.com to learn exactly how our services can help you build a next-generation system.

See Also: