NextSilicon¶

The NextSilicon Maverick-2 accelerator is one of the first commercially available devices that utilizes dataflow processing instead of classical parallelization paradigms. More information can be found at NextSilicon Website.

Hardware Overview¶

NextSilicon develops hardware based on the Maverick‑2 Intelligent Compute Architecture (ICA), which is described as a form of software‑defined hardware capable of adapting to software behavior during execution. According to the manufacturer, the system is intended to adjust its configuration in real time, with the goal of improving performance across a range of workloads without requiring modifications to existing application code.

The Maverick‑2 architecture is positioned for use in high‑performance computing, AI, and vector database workloads. It is claimed the hardware can be reconfigured dynamically to suit workload requirements and that it aims to deliver improved efficiency and performance at scale.

In terms of software integration, NextSilicon emphasizes compatibility with commonly used programming languages, including C/C++, FORTRAN, OpenMP, and Kokkos. The company also notes planned support for CUDA, HIP/ROCm, and various AI frameworks. This approach is presented as a way to reduce the need for specialized languages or extensive code porting.

Maverick-2 is installed as a PCIe add-in card in a standard x86 server and can therefore be accessed via the x86 login node.

Access¶

Access to the Maverick-2 nodes is restricted. Users must be added manually to the Maverick-2 access group. To request access, open a support ticket and briefly describe your intended use case.

Using Maverick-2¶

To use the Maverick-2 server, you need to request it via Slurm. The partition is named maverick-2.

ssh username@ftp-x86-login.scc.kit.edu

Use sbatch or salloc to get access to the maverick-2 partition:

sbatch -p maverick-2 script.sh

or, for interactive use

salloc -p maverick-2 -t 30

Software for Maverick-2¶

Hello World Example¶