Semidynamics has released Cervell, a customizable AI accelerator (NPU) available to license in configurations from 8 to 128 INT8 TOPS and targeting industrial, consumer, and other chip designs. Cervell combines the Semidynamics all-in-one AI-processing element, memory, and Aliado AI software development kit (SDK). Although it’s based on a RISC-V core, Cervell is best thought of as a self-contained NPU—an AI unit incorporating a RISC-V core rather than a CPU with AI extensions.
RISC-V Eliminates Lock In
Nonetheless, Cervell’s RISC-V compatibility is an important differentiator. Many NPUs are black boxes. A developer can take an off-the-shelf model or a proprietary one built with PyTorch or TensorFlow, convert it to the ONNX format, and use the NPU vendor’s dev tools to translate, load, and execute it. If the model doesn’t run well or requires an ONNX operator that the NPU doesn’t yet support, the developer depends on the vendor for updates.
Semidynamics, by contrast, translates the model to RISC-V code, which the developer can modify or supplement with additional operators. The only thing that keeps the code from being bog-standard RISC-V is the matrix operation Semidynamics implements. It, and other parties, such as CPU licensor SiFive, have proposed various matrix-math RISC-V extensions, but the community has yet to ratify a standard.
Selling Flexibility
Several other companies employ RISC-V technology in their NPUs. For example, alongside other hardware, each Tensix core in the Tenstorrent Wormhole employs five RISC-V cores, offering computing and data-movement flexibility infeasible with state machines.
Other companies take a different approach to avoiding the black-box trap. The Imagination E-series GPU, for example, has matrix units, which are similar to Nvidia GPUs’ tensor cores. Developers, therefore, can employ a GPU-style programming model like they would for data-center AI processing.
Semidynamics describes a few Cervell configurations, but the company has staked its business on customization. Licensors can customize the RTL, adding instructions, defining memory interfaces, and requesting proprietary features. It’s a service-heavy approach in an industry that usually sells many copies of the same intangible good.
Bottom Line
Semidynamics operates in a marketplace crowded with dozens of RISC-V and NPU suppliers. Among RISC-V companies, the company has sought to stand out through customizable CPUs and hardware add-ons. Among NPU companies, Semidynamics replicates this customizability advantage but focuses on reducing customers’ lock-in risk by offering the option of low-level programming. Chipmakers on their second or third AI-enabled SoC will be the first customers to value this if they have struggled to adapt past designs to new AI models.