Rivos, the creator of the RISC-V processor, has secured Series-A financing to support the advancement of its inaugural accelerator tailored for tasks demanding sophisticated AI and data analytics processing.
Founded in 2021 and headquartered in California, the company’s objective is to develop RISC-V-compatible cards capable of accommodating large-scale language models such as ChatGPT.
While details regarding the chip’s architecture remain undisclosed, Rivos has revealed that it integrates high-performance RISC-V CPUs and a data parallel accelerator (referred to as General-Purpose Computing on Graphics Processing Unit, or GPGPU) that share a unified memory domain encompassing DDR and HBM memory. This innovative chip is poised to manage workloads necessitating terabytes of memory.
The specific function of the RISC-V cores within the design has not been explicitly outlined at this juncture. However, it is anticipated that these cores will oversee hardware operations and engage in intensive number crunching tasks, akin to the role played by SiFive components in certain Google Tensor Processing Units.
Reportedly, Rivos’ chip will be fabricated utilizing TSMC’s cutting-edge 3nm process technology, akin to Apple’s latest silicon iterations, and is projected to integrate seamlessly into a multi-chip Open Compute Project reference design, underscoring the company’s focus on data center applications.
In an interview with Bloomberg, Puneet Kumar, the CEO of Rivos, highlighted the company’s aim to target “potentially smaller installations where Nvidia may appear overpriced from a cost perspective.” Further inquiries have been made to obtain additional insights regarding the chip and its production timeline.
Furthermore, Rivos is actively working on an application suite to bolster the expansion of its RISC-V-based computing ecosystem. Analogous to the evolution of Intel and AMD’s personal computing platforms, engineers are inclined to engage fully once a well-established ecosystem is in place. Rivos seems to be concentrating on AI software and data adaptable to platforms like PyTorch, JAX, Spark, and PostgreSQL, facilitating software recompilation for diverse hardware compatibility.
According to Kumar, “The rapid advancements in LLMs and their integration with the data analytics stack make it imperative for accelerators to be easy to program and deploy, and for data to seamlessly flow between CPU and accelerator.”
Despite encountering initial hurdles, such as a legal dispute with Apple in early 2022 over alleged employee recruitment from Apple’s SoC design teams, Rivos has managed to resolve the matter through a settlement earlier this year, streamlining the path for potential investors evaluating Rivos.
The Series-A funding cycle for Rivos was spearheaded by Matrix Capital Management, with Romit Shah from the firm assuming a board position. Noteworthy contributions were also made by prominent entities like MediaTek, Intel Capital, and Koch Disruptive Technologies, underscoring the industry’s confidence in Rivos’ endeavors.
Although a specific production timeline has not been disclosed, the financial infusion from this funding round will empower the startup to fabricate its inaugural chip.
