RISC-V chip designer Rivos has raised $250 million in series-A funding to bankroll production of its first accelerator for generative AI and data analytics workloads.

Founded in 2021, the California-based startup aims to produce RISC-V-compatible chips capable of running the kinds of large language models powering AI services like ChatGPT.

Rivos hasn't said much about the chip's architecture, beyond that it uses a combination of high-performance RISC-V CPUs and a data parallel accelerator (general-purpose computing on graphics processing unit, or GPGPU) that share a common memory domain across both DDR and HBM memory, and will be able to support workloads requiring terabytes of memory.

It's currently not clear what role those RISC-V cores will play in the design. However, we suspect they may function similarly to the SiFive cores used in at least some of Google's Tensor Processing Units to manage the hardware and keep its matrix multiplication units busy with numbers to crunch.

Rivos's chip will apparently be manufactured using TSMC's 3nm process tech - the same used by Apple's latest silicon - and are designed to slot into a multi-chip Open Compute Project reference server. This tells us the biz thinks its silicon will end up in the datacenter.

In an interview with Bloomberg, Rivos CEO Puneet Kumar revealed his outfit was targeting "potentially smaller installations where Nvidia might seem like an overkill from a cost perspective." We've asked the upstart for further comment and details regarding the silicon and its production timeline.

Additionally, Rivos is working on an open software stack to support development of its RISC-V based compute platform. As we've seen with the ramp of Intel and AMD's own accelerators, developers won't move until they see a mature ecosystem to work with. Nvidia's proprietary CUDA runtime is already mature and widely deployed. Others? Not so much.

Rivos appears to be targeting AI applications and databases that can run atop frameworks like PyTorch, JAX, Spark, and PostgreSQL, which allow applications to be recompiled to run on a variety of hardware.

"The rapid changes in LLMs and the merger with data analytics stack makes it vital that accelerators be easy to program and debut, and that data can seamlessly move between CPU and accelerator," Kumar explained in a statement.

Rivos's early efforts have not always gone well. In early 2022, Apple sued the RISC-V backer, alleging the upstart had poached employees working on its A- and M-series system-on-chip (SoC) designs. Last year, Rivos counter-sued Apple, arguing that Cook & Co had used the law as an instrument to thwart startups by "illegally restricting employee mobility."

Earlier this year, Apple agreed to settle the case, making it far easier for investors to consider Rivos.

Rivos's series-A funding round was led by Matrix Capital Management, with that outfit's Romit Shah joining the board. The chip startup also received support from several familiar names including MediaTek, Intel Capital, and Koch Disruptive Technologies.

The funding round will enable the startup to tape out its first chip - produce a finalized design for a fab to manufacture - though a timeline for production wasn't disclosed. ®

https://www.theregister.com//2024/04/17/riscv_ai_startup_rivos_funded/