How SiFive is Driving AI and Datacenter Innovation

Ian Ferguson, VP Vertical Market and Business Development

The advent of generative AI has brought an increased focus on datacenter design and optimization. Datacenter chips need to be high performance with low latency to power the latest AI workloads. At the same time, sustainability is a growing concern for datacenters, so datacenter chips need to be extremely efficient. Additionally, chips need to be designed with flexibility in mind so companies can scale the number of clusters as needed.

One of the extremely interesting and potentially disruptive announcements in the last couple of months has centered on Deepseek. Deepseek claims that its model uses significantly less computing resources than competitive options, and therefore uses less energy. Up until the end of 2024, the clear focus from customers we spoke to was time to deployment. Items like energy efficiency and cost took a back seat. I remember being in meetings in the middle of 2024 when our chief architect articulated a belief that the industry would enter a phase where sustainability is a much higher priority. Whether Deepseek ends up being a winner or not, the key takeaway is that new possibilities open up when technologists focus on different metrics. This will be an inflection point that causes companies to reevaluate system architectures as they scale out deployments more aggressively.

Many companies are turning to RISC-V to meet these performance and efficiency challenges. One of the reasons that RISC-V is gaining popularity is its tremendous flexibility. If you are designing a system from scratch, it makes less sense to build on an existing ISA with historical baggage since you will have to use multiple types of software across the device or datacenter. The RISC-V architecture offers a clean slate in a flexible, modular design.

The SiFive Performance P870-D takes advantage of the flexibility of RISC-V to bring high compute density and scalability to datacenters, vehicles, and embedded systems. Higher compute density enables improved performance per watt, which offers advantages in total cost of ownership. The P870-D is specifically tailored to meet customer requirements for highly parallelizable infrastructure workloads including video streaming, storage, and web appliances. Find out more about the P870-D here.

When the P870-D is paired with products from the SiFive Intelligence product family, datacenter architects can also build an extremely high-performance, energy efficient compute subsystem for AI-powered applications. SiFive’s Intelligence XM Series stands out with its super scalable and efficient AI compute engine. By bringing together scalar, vector, and matrix engines together, the XM Series achieves much more efficient memory bandwidth while offering extremely high performance per watt. In addition to targeting datacenter applications, the XM Series is ideal for AI applications for edge IoT devices, consumer devices, next generation electric and/or autonomous vehicles, and beyond. Read more about the XM Series here.

To help companies speed up development time, SiFive is open sourcing a reference implementation of the SiFive Kernel Library (SKL). SKL is a fully optimized C/C++ library with a set of tuned routines that maximize algorithm throughput on SiFive processors. Additionally, SiFive offers a number of resources to help optimize large language models (LLMs). Take a look at our recent blog post on SiFive’s AI/ML software stack for additional details. You can also check out our overview of RISC-V software milestones over the past year.