Is the SiFive FE310 a Toy?

[Sal Ammoniac]

Over the Christmas break I was sent a Sparkfun Red-V board, which has a SiFive FE310 RISC-V MCU.

I downloaded the docs for this MCU from the SiFive website, and what a MAJOR disappointment! I started with the usual blinky project to blink an LED on the board, so I needed to know how the GPIO on the MCU worked. Well, the GPIO section in the User's Manual is barely five pages in length and while it has a table listing the GPIO registers, it doesn't bother to explain the function of each of them, nor does it show the bit encodings of the registers. I literally had to guess and use trial-and-error to figure out how to blink the LED.

The entire user manual is only 116 pages long. This is ridiculous! How does SiFive expect to play in the professional MCU market with piss-poor documentation like this? It reminds me of the crappy documentation that comes with a lot of open source software. 

[westfw]

Use "user manual" in conjunction with the chip datasheet https://static.dev.sifive.com/SiFive-FE310-G000-datasheet-v1.0.4.pdf to see which bits go to which pins, I guess...
I remember when the ESP8266 docs were about like this.
I get more than a vague feeling that the FE310 is more a "demonstration" product, rather than something meant for mass use.
Wikipedia says "SiFive's business model is based on designing custom computer chips for other businesses."

[ataradov]

Yes, it is a toy. It is a test chip that they decided to sell for fun.

Also yes, FE310 is not a player in this market. You can't even buy them in quantity.

[donotdespisethesnake]

Is it possible to buy any amount of FE310 chips anywhere?

I guess despite all the publicity, there will be some people who don't realize SiFive is not just another company selling chips.

[Sal Ammoniac]

I get more than a vague feeling that the FE310 is more a "demonstration" product, rather than something meant for mass use.
Wikipedia says "SiFive's business model is based on designing custom computer chips for other businesses."

Ah, that would explain things. Still doesn't excuse the crappy docs, though--I've seen other "demo" chips with much more complete and polished documentation.

I really hate developing for poorly documented parts, so I'll toss this board on my pile of development boards that didn't make the cut and wait for a real RISC-V MCU.

[ataradov]

There is a real RISC-V MCU from GigaDevice.

[tmadness]

From what I remember SiFive is more into selling their IP and expertise to companies that want to integrate RISCV and a few other concepts with ICs. The MCU is more of a dev board/ POC as to what RISC V is capable of, somting to give to the dev team while the ASIC guys build the actual chips.
And as far as any RISC V processor being ready fro a commercial product, that's going to take time ARM Cortex M took a decade to get to this point with support, documentation and wide spread availability. RISC V needs time to mature.   

[brucehoult]

How does SiFive expect to play in the professional MCU market

It doesn't.

SiFive is in the business of licensing cores to to other companies, who might in turn make and sell stand-alone MCUs, or embed them in some other product.

Exactly like ARM.

Early versions of the FE-310 are a demo of the E31 core, not a mass market general purpose MCU. The FE310-G003 (with 64 KB SRAM and other features) has been designed for a specific high volume product, the "SiFive Learn Inventor" and might also see general availability. But it's a sideline.

Hundreds of millions of SiFive cores will ship in 2020 inside Samsung's flagship phones (controlling the camera and 5G), in Qualcomm 5G SoCs, in Huami wearables, in Allianz microwave ovens (>50% world market share, OEMing for many brands such as GE) and other whiteware, and many others that may or may not ever be publicly announced.

It's up to the MicroChips, NXPs, STMs etc of the world to produce retail RISC-V MCUs with their usual assortment of peripherals if they see a market for them.

[brucehoult]

From what I remember SiFive is more into selling their IP and expertise to companies that want to integrate RISCV and a few other concepts with ICs. The MCU is more of a dev board/ POC as to what RISC V is capable of, somting to give to the dev team while the ASIC guys build the actual chips.

Exactly.

And as far as any RISC V processor being ready fro a commercial product, that's going to take time ARM Cortex M took a decade to get to this point with support, documentation and wide spread availability. RISC V needs time to mature.

There are millions of products already shipped with RISC-V processors in them, and hundreds of millions in 2020 -- possibly near a billion when you count all vendors (Andes is very big in China).

I'm not sure whether Western Digital will be shipping in 2020, but they'll add a billion plus CPU cores a year all by themselves.

[TK]

If it is similar to ARM (licensing) what is the added value?  Is it offering a cheaper licensing per core?  In my opinion Samsung is involved because one of the co-designers of RISC-V is South Korean.  Lots of Chinese companies are developing RISC-V based SoC's, I guess without paying any royalty to SiFive... Maybe SiFive has an advantage today (RISC-V knowledge) because it was founded by the designers, but once the knowledge is widely spread, then they will be at par with other chip designers... unless they keep strict control on the ISA and become sort of proprietary like ARM.

[Sal Ammoniac]

SiFive is in the business of licensing cores to to other companies, who might in turn make and sell stand-alone MCUs, or embed them in some other product.

Exactly like ARM.

I thought the whole point of RISC-V was to get around ARM royalties? If that's the case, wouldn't licensing a RISC-V core from SiFive be just paying royalties to a different company? When I first heard about RISC-V, I assumed that companies, like Samsung, who needed an MCU core would take the RISC-V API and implement their own core, hence avoiding the payment of royalties to anyone.

The only way I can see the SiFive business model working is for them to offer some value over using an ARM core, either a technical advantage (more speed, features, etc.) or a financial advantage (lower royalties).

[ataradov]

If that's the case, wouldn't licensing a RISC-V core from SiFive be just paying royalties to a different company?

No, because you have an option to roll your own. Plus there is an actual competition, so pricing is way better. SiFive is not the only company offering RISC-V core IP.

I assumed that companies, like Samsung

And Samsung will roll their own without paying anything to anyone. Except in engineering costs, of course.

And competition with RISC-V also forces ARM to lower their prices and improve offerings, so it is a win-win situation for the customers.

[Sal Ammoniac]

And Samsung will roll their own without paying anything to anyone.

Ah, but brucehoult says Samsung will be shipping hundreds of millions of SiFive cores in their products.

[ataradov]

Ah, but brucehoult says Samsung will be shipping hundreds of millions of SiFive cores in their products.

Well, they can make their own. By "Samsung" I meant more "big corporation" in general. Ultimately it is a business decision. They may not want to invest money into R&D at first. And they may want to roll their own for future products. They have this flexibility with RISC-V, but not with ARM.

Also, if Samsung does not like licensing terms or performance of the cores, they can find another supplier. There are at least 3 or 4 big ones already. Again, no such luck with ARM.

[hamster_nz]

There are plenty of sides to this:

- The technical side - Is it actually a better designed or more capable ISA than others?
- The legal side - Licensing, patents, ownership
- The money-go-round - Who pays money to who for what. How money is made.
- Is having a common, open ISA platform that anybody can implement actually good for the tech industry?

On the technical side, I think it is a better core ISA than x86. It still isn't fully formed (with some extensions in flux), but what is there is arguably technically better - no "real mode", no super-complex instructions to gum everything up. Also the ISA is a lot more approachable than other ISAs.

On the legal side, having an unencumbered ISA is great. If you win the lottery and want to set up "GaAsSix", which sells high performance solutions can competes with SiFive you can. And if your IP vendor is yanking your chain you have options. It then focuses on who can provide the best value to their customers, not who has owns what IP.

Is having a common, open ISA good? I think so. Even if it is just so we don't have just x86 for big systems and ARM for smaller systems.

The money-go-round is the most confused story at the moment. I really don't think you deliver ARM-like levels of products, support and documentation without an ARM-level licensing revenue stream to pay for your efforts. The question is how much lower can you deliver and still have a viable, sustainable  product to the market.

[brucehoult]

Note that Samsung (like Apple, Qualcomm and a handful of others) holds an ARM architecture license and is allowed to design their own micro-architecture (but not vary the instruction set in any way [1]) and has done so in the Exynos SoCs in the global versions of the Galaxy S7, S8, S9, S10. However they recently announced that they will revert to using standard ARM cores in Exynos SoCs and laid off their ARM core designers in Austin.

Samsung certainly could, if the need arises, design their own RISC-V cores. At the moment they seem to be moving in the opposite direction, with the cancellation of their custom ARM-compatible cores as noted above, and also the cancellation of the in-house GPU project I worked on for a couple of years in favor of (so I've heard on the grapevine) licensing GPU technology from AMD.

[1] ARM will later this year start to allow licencees to add custom instructions on the Cortex M33 *only*

[brucehoult]

The only way I can see the SiFive business model working is for them to offer some value over using an ARM core, either a technical advantage (more speed, features, etc.) or a financial advantage (lower royalties).

Support and guarantees are worth something.

Note that RedHat, who basically just package up open source software that anyone could do for themselves, were acquired by IBM for $34 billion, about the same as the $32 billion Softbank paid for ARM.

RISC-V microcontroller cores are pretty much open source now, with excellent free cores equivalent to Cortex M0 to M4 range (e.g. PULP project Zero-riscy (now Ibex) and RI5CY) and M7 (Western Digital SweRV).

SiFive seems to be the main organization developing 64 bit Linux-capable cores so far, with the U54 having been available for nearly two years now (in the HiFive Unleashed), dual issue A55-class U74 announced 14 months ago, and out-of-order A72-class U84 announced two months ago. Those are serious cores with serious teams and serious development money and highly experienced CPU designers ex ARM and MIPS and Intel and AMD and Qualcomm and others behind them -- it's not just a couple of grad students any more. It would be a very rare (and large) company that found it worthwhile to develop this class of core in-house instead of licensing it.

[SiliconWizard]

To add to this, it may show that products like the Sparkfun Red-V board, although it gives exposure to RISC-V and to SiFive, may actually slightly harm SiFive in some way, as we can see from reactions in this thread. Yes the FE310 was meant as a demo device and has value as such, but this isn't necessarily obvious to the uninformed lambda person, who will see the dev board as just one with a new MCU to play with (and then expecting what they'd get with the same using mature devices from the big players).

Of course SiFive had no say in this as you can't direct what Sparkfun does with your products, but I think this is certainly something you guys should consider and take a look at before it could get out of hand. Maybe discuss things with Sparkfun as well. (As I got from your answer to my post talking about this board, it even looks like you didn't know what FE310 version they used until you zoomed at the board picture!!) It's all your business of course, but I just think giving access to your  technology to the masses like this may have consequences, and should be handled with some care. Just a thought. (Even if it's just a question that's easily answered, the topic of this thread alone is not that great for SiFive IMO, and don't forget EEVBlog is very well referenced in Google...)

[GeorgeOfTheJungle]

What I don't get is their prices, when a +600MHz Teensy 4 is less than $20 and dual core esp32s (240MHz) are $5... Sparkfun: HiFive1 (320MHz): $67, RedV RedBoard (150MHz): $39.95, RedV Thing Plus (150MHz): $29.95

[brucehoult]

As I got from your answer to my post talking about this board, it even looks like you didn't know what FE310 version they used until you zoomed at the board picture!!

I'm just one minion working on specific things in a company with 600+ very busy people. I certainly don't know about every deal -- and if they're buying the chips from someone like Mouser then it's possible no one knew.

I think the HiFive1 has probably done a lot to kickstart RISC-V. Definitely it's what got me interested enough to go out and join the company that made it. As volume production boards come on stream (and we're seeing a number from the Chinese) it may be time to retire it, or at least only give it to pre-qualified customers for cores or custom SOCs.

But we believe you can never know who might come up with a bright idea in a garage, get $10k or $20k to prototype an idea, and then get the couple of hundred $k to get into volume production of a custom 180nm SoC for their product.

Maybe just price the dev board higher, like the HiFive Unleashed, I don't know. ARM's A53 / A72 dev board is $10k.

[brucehoult]

What I don't get is their prices, when a +600MHz Teensy 4 is less than $20 and dual core esp32s (240MHz) are $5... Sparkfun: HiFive1 (320MHz): $67, RedV RedBoard (150MHz): $39.95, RedV Thing Plus (150MHz): $29.95

You can't compare volume production retail products with short-run things using MPW ("shuttle run") production.

Paul makes great products, and I suspect very few are aware of the staggering volumes of them he sells! He doesn't have much NRE cost in each board.

The Teensy 4 isn't made by ARM, and the NXP iMXRT1062 chip isn't made for the Teensy 4, and it isn't made by ARM either. It's made in very high volume by a 3rd party.

[Sal Ammoniac]

Yeah, those prices are out of line with other boards, but I suppose that may be a consequence of the FE310 being a limited production part. Maybe, maybe not--I have no insight into how Sparkfun set the price for this board.

I'm still firmly of the opinion that if a company is going to release a "demonstration" part, they should at least document it well. The FE310 is not a complex MCU--it doesn't have USB, Ethernet, and other complex peripherals, just a few basic ones like UART, SPI, and PWM--so it shouldn't be too hard to thoroughly document those peripherals. The current state of the documentation leaves a bad taste in my mouth. It's very reminiscent of the inadequate documentation that accompanies a lot of open-source software.

Question for brucehoult: does SiFive have tech writers, or was the existing documentation written by the engineer(s) who implemented this chip? I can certainly believe the latter--engineers are often pretty poor writers and often assume others already know how the part works.

[brucehoult]

I'm not sure why you think it's not thoroughly documented. Others say it's good and I found it pretty clear in January 2017 as a n00b to microcontrollers other than on AVR Arduinos. (At which time, by the way, SiFive had about a dozen people -- there are 600 now and, yes, they include tech writers)

The main thing is to make sure you look at the *right* document, depending on whether you want to know about the CPU core, the SoC / peripherals, or the board. They are all pointed to from here https://www.sifive.com/documentation

[Sal Ammoniac]

I'm not sure why you think it's not thoroughly documented. Others say it's good and I found it pretty clear in January 2017 as a n00b to microcontrollers other than on AVR Arduinos. (At which time, by the way, SiFive had about a dozen people -- there are 600 now and, yes, they include tech writers)

The main thing is to make sure you look at the *right* document, depending on whether you want to know about the CPU core, the SoC / peripherals, or the board. They are all pointed to from here https://www.sifive.com/documentation

The core is reasonably well documented. I'm talking about the peripherals here. Take the GPIO as an example. The relevant document (SiFive FE310-G002 Manual, v19p05) has a GPIO chapter. That chapter is five pages in length. It has a table listing the GPIO registers, but the function of each register, and the bit encodings in each, is not explicitly defined. One has to make assumptions, and then test those assumptions to verify correct operation. I'm of the opinion, which I think I share with other embedded developers, that MCU hardware reference manuals should be complete, explicit, and leave nothing up to the read's imagination. One shouldn't need to extrapolate information out of a hardware manual--it should be right there in black and white.

As a comparison, another MCU that I've used recently has a 46 page GPIO chapter in its hardware reference manual. It goes into explicit detail for each and every register and explicitly describes the operation of the peripheral.

[SiliconWizard]

I personally find it reasonably well documented. For the GPIOs, sure it's not very detailed and kind of assumes that you already know all the concepts usual with GPIOs, but it looks enough IMO. Seems like for each register, one bit is to control the corresponding IO, you don't need each register detailed for this?