Nvidia joins chip design startup's $400m round

SiFive, a chip design startup founded by UC Berkeley engineers and built around the open RISC-V architecture, has raised US$400 million at a US$3.65 billion valuation, with Nvidia joining the round led by Atreides Management.

The company licenses CPU designs rather than manufacturing chips, following a model similar to Arm.

The funding marks a step up from its previous round in 2022 and comes as SiFive expands beyond embedded systems into AI data center processors.

Its latest designs are being positioned to work with Nvidia’s software and infrastructure, as demand grows for alternative chip architectures in AI computing.

🔗 Source: TechCrunch

🧠 Food for thought

Implications, context, and why it matters.

SiFive goes after an AI bottleneck that older chips struggle to fix

SiFive’s data center plan goes after the “memory wall,” an AI slowdown where processors sit idle while memory catches up 1.
In some large language model steps, GPUs can drop to 10% utilization or less because data transfers take too long 1.
Its design uses selective cache bypassing. Some vector loads skip the Level 1 (L1) cache so L1 can focus on control-flow data rather than being swamped by model weights 1.
The aim is steadier throughput for hyperscalers, large cloud companies, running agentic AI workloads. Older designs often need a full redesign to reach the same consistency 1.

Nvidia’s investment pulls open RISC-V closer to Nvidia’s NVLink Fusion ecosystem

Nvidia’s funding links SiFive’s open RISC-V central processing unit intellectual property (CPU IP) with Nvidia’s AI infrastructure stack 2.
NVLink Fusion is a rack-scale AI infrastructure platform 3. It lets partners pair custom CPUs and XPUs, a broad term for accelerator chips, with Nvidia’s NVLink scale-up interconnect 3. It also bundles ConnectX SuperNICs, BlueField data processing units (DPUs), and Mission Control software 3.
Nvidia can still earn money from surrounding NVLink Fusion hardware when third-party silicon runs the compute layer. Some analysts call this a “tax on custom application-specific integrated circuits (ASICs)” 4.
The contest shifts toward ecosystem choice. One path runs through Nvidia’s platform, while another uses the UALink open standard backed by rivals like AMD and Intel 4.