SiFive: a RISC-V Fabless Semiconductor Company

A couple of weeks ago I talked to Krste Asanović and Jack Kang of SiFive. Jack is the VP of product and business development. Krste is the chief architect, and, of course, the leader of the team at UC Berkeley that defined the RISC-V ISA (although he is currently on leave-of-absence from Cal), and the chairman of the RISC-V foundation. Jack has a semiconductor background from NVIDIA and Marvell. Other co-founders of SiFive are Yunsup Lee and Andrew Waterman, who were also on the team defining RISC-V at Berkeley.

The company is funded by Sutter Hill Ventures (one of the investors in NVIDIA, LSI Logic, and Mentor Graphics, among others in the semiconductor ecosystem). They are not disclosing how much money they raised. They are growing fast and currently have around 20 employees.

One of their motivations is to bring the cost reduction that goes along with open-source software (and instruction sets!) to the hardware world. There is an increasing move in this direction as companies like Facebook, IBM, and Google put parts of their server infrastructure into the public domain. This is especially important in semiconductors since Moore's law has run into a wall, at least economically, even though we have a couple more process generations coming at us. This has a big implication: most companies cannot afford to design a leading-edge-node design, and probably couldn't recover their costs as profit even if they did since most markets are not big enough.

Semiconductor is a support-intensive business since the semiconductor company only makes any money when the customer goes into production after working through all their issues. There is an especially heavy burden with the software deliverables. Typically, the customer cannot work through issues on their own due to lack of documentation and knowledge. This means that chip companies need to be very picky about picking their customers, since it will take lots of people to support them. Often they can only take on three to four customers. Eacn new customer is, in a sense, competing with all the other customers. The result of this is that there is an entire market of underserved customers who don't really have access to silicon. Open source is a way to solve this problem: open-source software, open ISA, and open-source hardware specifications. A good example is IoT. Since nobody really knows which products will be successful, it is very fragmented and the industry is not set up to serve it well.

SiFive was announced in July. It is a fabless semiconductor company. They will build silicon and deliver chips to customers. Well, TSMC will actually build the silicon. But the business model is to sell chips.

In July they also announced two platforms:

Freedom Unleashed Platform

This is targeted at applications needing 64-bit cache-coherent processors and high-performance peripherals. It will be built in 28nm

Freedom Everywhere Platform

Aimed at IoT, embedded microcontroller. 32-bit, with on-chip RAM and ROM. It will be built in 180nm (aka 0.18um).

The specifications are open so everything can leverage the same software infrastructure. SiFive will build the base chips and then customize them for their customers. They already have them available as FPGA kits and you can just download the bitstreams and buy the boards.

The idea is that they will build custom silicon on top of these platforms, and there will be a well-supported open software ecosystem. For example, a security company wanting to build a specialized product today has to design the whole SoC before they can add their secret sauce, which doesn't really add much value.

The schedule for these first chips is in a little flux since they are trying to align tapeouts with their customers. But silicon this year.

Next: Cadence's History with MIPI

Previous: CDNLive Boston Overview