本文由半导体产业纵横(id:eQviews)编译自TechCrunch
More and more startups are focusing on RISC-V.
Rivos made headlines in 2022 after Apple filed a commercial lawsuit against it, which accused Rivos of hiring dozens of Apple engineers and using confidential information to develop chips.
The company has denied the allegations and countersued Apple on the grounds of unfair competition. Apple finally settled the lawsuit in February of this year. Around the same time, it ended a separate lawsuit with several Apple engineers hired by Rivos.
Now, with the court farce over, Rivos is doubling down on its efforts to bring its chipset technology to market, CEO Puneet Kumar told TechCrunch.
"Rivos was founded with a mission to build industry-leading, energy-efficient, high-performance chips," said Kumar, "and we're excited to target customers who are building data-driven solutions." ”
A significant new amount of funding will help to assist in these efforts.
On Tuesday, Rivos announced it had raised more than $250 million in an oversubscribed extended Series A funding round led by Matrix Capital Management with participation from chip giants, including Intel (through its corporate venture capital arm) and MediaTek. Other investors include Cambium Capital, Hotung Venture Group, Walden Catalyst, Dell Technologies Capital and Koch Disruptive Technologies.
It's a pretty big shift for Rivos, founded in 2021, about a year ago, in the shadow of Apple's lawsuit, Rivos was struggling to raise money from investors and hire employees. In August, Rivos laid off nearly 20 employees, or 6% of its workforce at the time, and was forced to postpone a planned $400 billion Series A funding round, according to The Information.
Custom server chips
According to Kumar, Rivos' long-term goal is to build chips for servers that can handle intensive data analytics and AI workloads, including generative AI workloads.
"Our goal is to attract customers who are developing data-driven solutions (for example, those who leverage generative AI and data analytics to drive decision-making)," Kumar says. ”
Rivos' first chipset is built on the open standard instruction set architecture (ISA) RISC-V.
The ISA is the technical specification on which each chip is based, describing how the software controls the chip's hardware. For general-purpose computing, chip design teams typically license existing ISAs from existing companies, such as Arm or Intel. But RISC-V offers an open, royalty-free alternative.
Rivos' chip has what Kumar describes as a "data parallel accelerator" to accelerate computing related to AI and big data, essentially a GPU designed for purposes other than graphics processing. It is manufactured using TSMC's 3nm process process.
Companies such as Qualcomm, MediaTek, Nvidia, and AMD are expected to use TSMC's 3nm process in their upcoming chip series, but Apple is the only company to use the process in its M3 chipset family in 2024.
In addition to building chips, Rivos is also working on stand-alone data center hardware based on the Open Compute Project's modular standard, which will effectively be used as a plug-and-play chip. Kumar said it is creating a "firmware-to-app" software stack for programming the chip.
Kumar adds, "Our customers' workloads can be easily deployed on our more efficient hardware, but still use their existing models and databases, giving them immediate benefits.
Currently, Rivos is profitable, and it plans to make money by charging customers, mainly large data center operators, for hardware and complementary software solutions. Early investor David Goel said Rivos' "low-friction" adoption pipeline was a key differentiator in the highly competitive chip market.
"The Rivos team has skillfully integrated a groundbreaking new RISC-V architecture with a creative accelerator to effectively bring this vision to life," Goel told TechCrunch. ”
Rat race
As the generative AI boom continues, big tech companies (one of Rivos' potential customer base) are racing to develop their own in-house chips for AI and big data analytics.
On its fifth-generation TPUs, Google recently released Axion, its first dedicated chip for running models. Amazon has several lines of custom chips. Microsoft launched the Azure Maia AI accelerator and Azure Cobalt 100 CPUs last year. Meta is also steadily advancing its own design.
At the same time, dozens of startups are vying for a share of the custom data center chip market, which could reach $10 billion this year and double by 2025.
Groq, a company that develops chips that run faster than traditional hardware, recently launched a new business unit for enterprise applications and use cases. Tenstorrent, an artificial intelligence hardware startup helmed by engineering personality Jim Keller, is looking to build its chipsets into data centers. South Korean fabless AI chip company Rebellions has raised hundreds of millions of dollars to boost production of its data center-focused chip, Atom.
Currently, Nvidia is the dominant force in the field of AI chips, and it has proven to be difficult to overthrow.
Nvidia briefly became a $2 trillion company this year, fueled by demand for its GPUs for AI training. Wells Fargo Equity Research estimates that Nvidia has a 98 percent market share in the data center GPU space, and the company's data center business grew by more than 400 percent in the fourth quarter of 2023 as Nvidia builds a new division to design custom chips for cloud computing and other companies.
Given the intensity of the competition, and the chilling effect of Nvidia's supremacy on potential competitors' funding, it's tough for some custom server chip upstarts.
A few months ago, Graphcore's valuation was reportedly slashed by $1 billion following a failed deal with Microsoft, and the company said it was planning to lay off employees due to an "extremely challenging" macroeconomic environment. Habana Labs, Intel's artificial intelligence chip company, laid off about 10 percent of its workforce last year. Also last year, SiFive, like RISC-V startup Rivos, laid off 20% of its workforce and halted its core product line.
So will Rivos perform better?
Kumar is reluctant to talk about customers, and Rivos' chips are not expected to reach mass production until sometime next year. But with 375 employees and hundreds of millions of dollars in bank deposits, Rivos is well positioned to scale up manufacturing and double down on platform and software engineering, Kumar said.
"The rapid change of generative AI and the merging with the data analytics stack makes it critical that accelerators are easy to program and debug, and that data can move seamlessly between CPUs and accelerators," Kumar said. Rivos meets this need with our 'recompile, not redesign' approach. ”
*Disclaimer: This article was created by the original author. The content of the article is his personal point of view, and our reprint is only for sharing and discussion, and does not mean that we agree or agree, if you have any objections, please contact the background.