From IoT to Data Center: SiFive Unveils Cost-Effective Intelligence XM AI Cores

Developer SiFive, known for its RISC-V processor cores, has decided to tap into the AI ​​boom with the announcement of Intelligence XM clusters, the industry’s first RISC-V solutions equipped with a scalable matrix calculation engine to handle AI workloads.

As SiFive notes, the new design should help developers of RISC-V-based chips in creating custom AI systems, including for autonomous transport, robotics, UAVs, IoT, edge computing, etc., where the role of such loads has recently has grown significantly, but the requirement for energy efficiency has not gone away. But if desired, server accelerators can also be created, the company says.

Each matrix block within one XM cluster is complemented by four X Cores, each of which contains two vector calculation blocks and one scalar calculation block. Together they share a common L2 cache. The XM cluster has a bus with a bandwidth of 1 TB/s and supports connections to two types of memory – coherent via a common CHI bus, to which external DDR/HBM memory is connected, or a high-speed port for SRAM. The performance of one XM cluster is 8 Tflops in BF16 mode and 16 Tops in INT8 mode per GHz frequency.

Source here and below: SiFive

The type of host core is not important, it could be RISC-V, Arm or even x86. However, host cores may be absent altogether. XM-based chips are expected to have an average of four to eight clusters, giving them up to 8 TB/s of memory bandwidth and up to 64 teraflops of performance in BF16 mode, all at just 1 GHz with low power consumption. But it is also possible to scale up to 512 XM blocks, which will give 4 Pflops of BF16. NVIDIA Blackwell, for example, has 5 Pflops performance in the same mode.

In order to further popularize the RISC-V architecture, the company also plans to open source the reference implementation of the SiFive Kernel Library (SKL). SKL includes an implementation of various popular algorithms optimized for RISC-V SiFive cores, including those for working with neural networks, signal processing, linear algebra, etc.

SiFive appears to be doing well, and as CEO Patrick Little noted, new core designs will help it maintain its growth momentum and keep up with the evolution of AI, while remaining a provider of unique open-source processor solutions. architecture. At the moment, SiFive already supplies its solutions to such giants as Alphabet, Amazon, Apple, Meta✴, Microsoft, NVIDIA and Tesla.