Intel Reportedly Cancels Thunder Bay Hybrid SoC

Intel has quietly canceled its hybrid Thunder Bay system-on-chip (SoC) that integrates general-purpose CPU cores and computer vision-oriented Movidius hardware. The chipmaker does not disclose the reasons behind its decisions, but it looks like Intel's CPUs and vision processing units (VPUs) will remain separated for now. 

"Remove Thunder Bay specific code as the product got canceled and there are no end customers or users," a Linux patch discovered by Phoronix reads.

Intel kept details about its Thunder Bay SoC under wraps. Based on Linux patches uncovered by Phoronix, the Thunder Bay SoC was meant to be a low-power design packing Arm Cortex-A53 CPU cores and Movidius VPU hardware (which Intel acquired by taking over Movidius in 2016). Still, the exact configuration of the product remained unknown. 

Intel's Thunder Bay SoC was intended for commercial and Internet-of-Things applications requiring computer vision acceleration and general-purpose processing capabilities. Such edge-computing applications are expected to get increasingly common in smart cities. 

Meanwhile, it looks like users of applications that need CPUs and VPUs are perhaps satisfied with their edge servers running Xeon and Movidius silicon, such as the Keem Bay accelerator card introduced in 2019.

Furthermore, as machine learning acceleration gets ubiquitous, many applications may adopt different hardware, including Intel's own Habana Gaudi, Nvidia's GPUs or Jetson SoCs (with integrated GPU cores). As a result, it remains to be seen whether Intel decides to offer a Thunder Bay-like SoC in the future and how this potential product will be configured.

While Movidius VPUs are not mentioned regularly, they have their benefits. The Movidius vision processing unit packs general-purpose MIPS cores with programmable 128-bit vector processing (called SHAVE cores), various hardware accelerators, and image signal processing capabilities. Therefore, VPUs are somewhat more tailored for edge-computing applications from power consumption and footprint points of view than high-performance AI/ML accelerators.