Overview
FuriosaAI develops AI accelerators that power next-generation applications by running the world’s most advanced models in a power-efficient manner. Furiosa believes this is a critical lever in making AI computing more sustainable for the next generation.
While Furiosa’s first-generation AI accelerator, WARBOY, targets Computer Vision applications, their second-generation solution, RNGD (pronounced “Renegade”) targets Large Language Model (LLMs) and multimodal applications.
Furiosa’s AI accelerators leverage their Tensor Contraction Processor (TCP) compute architecture and offer numerous benefits including:
- Energy Efficiency: Designed with power optimization in mind, Furiosa’s accelerators provide top-tier performance while minimizing energy consumption.
- Programmability: The accelerators are highly versatile, making them suitable for deployment across a wide range of applications, from edge servers to large-scale data centers.
- Ease of Integration: Furiosa also provides various tools to allow seamless integration of their products into existing systems, reducing deployment time and costs.
Roadblocks to High-Performance AI Solutions: Furiosa’s Development Challenges
Developing high-performance AI accelerators presented Furiosa with several challenges:
- Time-to-Market Pressure: In the fast-paced AI industry, getting to market quickly is crucial. Delays in chip development could have a significant impact on Furiosa’s competitive edge.
- First-Pass Silicon Success: A flawless first-time chip design was essential to avoid costly redesigns and further delays.
- Balancing Performance and Power: Striking the optimal balance between raw processing power and power consumption was critical for Furiosa’s success.
Revving Up Development: How Rambus PCIe, HBM and XpressAGENT IP Fueled Furiosa’s Success
Furiosa turned to Rambus for PCIe 5.0 and HBM3 Controller solutions. Since Furiosa had already successfully used the Rambus PCIe 4.0 Controller in their first product, WARBOY, leveraging Rambus’s extensive PCIe experience for Gen5 in RNGD was a natural choice.
XpressAGENT, a unique solution from Rambus for integrating PCIe subsystems with auxiliary logic, was especially helpful in reducing the effort required to develop control, monitoring, and debugging features. The Rambus HBM3 Controller was chosen because it was a good match with the third-party HBM3 PHY that supports the desired clock speeds.
Furthermore, Furiosa’s engineers received ongoing assistance from Rambus’s team of experts, ensuring a smooth integration of Rambus IP into Furiosa’s products by addressing any technical challenges that arose.
A Collaboration Rooted in Innovation
The collaboration between Furiosa and Rambus proved to be highly successful. By integrating the Rambus Digital Controller, Furiosa was able to overcome key development challenges and bring their products to market on time. The collaboration not only facilitated first-time functional success but also optimized accelerator performance, power efficiency, and scalability.
“The Controller IP from Rambus, along with exceptional technical support, were instrumental in the successful development of our accelerators,” said June Paik, CEO of Furiosa. “The bring-up process for our RNGD chip has been efficient and smooth, allowing us to adhere to our ambitious timeline for commercializing the product.”
For more detailed information on Furiosa and its products visit www.furiosa.ai
Leave a Reply