SyncAI.news, a Varaisys broadcasting
NextSilicon Offers A YACA (Yet Another Computing Architecture) For AI
KF

Karl Freund, Contributor

· 1 min read

World NewsForbes: Innovation

NextSilicon Offers A YACA (Yet Another Computing Architecture) For AI

There are many approaches to building a computational accelerator, but nearly all are based on programmable cores in GPUs or fixed ASIC logic. NextSilicon’s Maverick-2 Intelligent Compute Architecture (ICA) represents a dramatically different approach, using reconfigurable software-defined hardware acceleration that delivers real-time adaptability and high performance. Think of it as an FPGA on steroids, but a lot easier to program. Let the software do the work. Let’s take a look.

RISC-V CPU

First, NextSilicon built a fast 2 core CPU, using the open source RISC-V architecture, augmented with out-of-order execution and 256-bit vector processing. This could be ideal for HPC clusters if and as the RISC-V Ecosystem matures. And the cores lead into Maverick-2, where they supply scalar operations and control, offloading the host CPU.

Maverick 2: A Reconfigurable Accelerator for HPC and AI

Maverick-2 represents a significant shift in computing with intelligent software-defined hardware acceleration that delivers real-time adaptability and excellent performance. The technology is designed to accelerate applications initially for HPC ((including national laboratories, hedge funds, and geoscience companies), with AI and vector databases next in line for exploitation. Given the support for 64-bit math, their booth at Supercomputing was always full of curious engineers and scientists looking at both the Mav2 and the RISC-V demos. You could almost envision people pulling out their credit cards to buy samples.

The addition of the RISC-V cores to Maverick-2 is design to resolve some of these issues. Rather than requiring developers to rewrite software in specialized dataflow languages, ICA compiles standard C++, Fortran, Python, OpenMP, Kokkos and CUDA into an intermediate representation and automatically extracts dataflow graphs. In addition, ROCm and HIP are in development to work with AMD codes.

Original source

This story was published by Forbes: Innovation and written by Karl Freund, Contributor. SyncAI.news shows a preview; the complete article is on the publisher's site.

Read the full story on forbes.com

Similar News