NextSilicon announced the development of Maverick-2 devices – the so-called Intelligent Compute Accelerator (ICA). The products are said to provide high performance and efficiency when solving HPC and AI problems, as well as when serving vector databases.
NextSilicon is developing a new computing platform for demanding applications. Special software algorithms are used to dynamically reconfigure equipment based on data obtained directly during task execution. This allows you to optimize performance and power consumption.
Image source: NextSilicon
The Maverick-2 ICA, according to the company, is a software-defined hardware accelerator. According to NextSilicon, in terms of performance per watt of energy consumed, the product is more than four times superior to traditional GPUs, and in comparison with top-end CPUs, it achieves 20 times superiority. At the same time, it is said that operating costs will be reduced by more than half.
«Telemetry data collected while the application is running is used by NextSilicon’s intelligent algorithms for continuous self-optimization in real time. The result is efficiency and performance in HPC workloads while reducing power consumption by 50-80% compared to traditional GPUs,” the company says.
Maverick-2 solutions are available as a single-chip PCIe 5.0 x16 expansion card and a dual-chip OAM module. In the first case, the HBM3e memory capacity is 96 GB, power consumption is 300 W. For the second product, these figures are 192 GB and 600 W. The clock frequency in both versions is 1.5 GHz. TSMC 5nm technology is used in production. It talks about compatibility with popular programming languages and frameworks such as C/C++, Fortran, OpenMP and Kokkos. This allows many applications to run unchanged, making porting easier and eliminating the need for a proprietary software stack.