Comment by contrarian1234 | Hacker News Reader

Comment by contrarian1234 | Hacker News Reader

Not an industry expert.. just a curious onlooker, but this is not surprising in the slightest - though I am surprised the VC firehose has been turned off this soon. I thought they'd float along on stupid money for longer.

From the start they seemed overfunded and with no real good long term business model. They have no lock-in because of the whole open nature of RISC-V.

- They're not designing for any specific in-house purpose like Alibaba/Western-Digital

- The low end will always be dominated by Allwinner and company and they can't win

- The high end is competing with x64 and huge companies like Intel.. also can't win unless you're Apple

They make non-pricecompetitive middle of the road chips that nobody really needs other than RISC-V enthusiasts.. And at the end of the day effectively anyone can come in and do what they're doing at any point.

The latest announcement seems to show they're finally shifting to the strategy their Chinese competitors have been doing for years. Open RISC-V cores and then get lock-in with custom NPUs. I doubt they'll be able to compete with Chinese firms there though

Cool idealistic company, but at the end of the day making cool stuff isn't in itself a business model - and you gotta make money to pay the bills

RIP

Sounds like they were basically a victim of venture capital. First the pressure to grow (too) rapidly, then the gutting for failure to deliver. Growing a company with wild ambitions sustainably is hard. I feel sorry for the employees that tried.

And of course 9 women can't deliver a baby in one month.

I think it's an example of a company suited for venture capital. As far as I can see, there was no option other than to grow rapidly. They'd need to have matched Intel/Qualcomm/etc. to have any chance of success.

VC gave them money on the initial RISC-V hype but then the hyped died down. There was like a good year or two with no real affordable products on the market and the software/toolchains were all half-baked. It also became clear that an open instruction set doesn't really bring you any huge "win". As far as I can see, the necessary subsequent waves of VC with deeper pockets never materialized (like MagicLeap had managed)

The Chinese competitors were also very quick and managed to catch up and in essence beat them to market - so they've lost any first-mover advantage

yup this is those cases where VC money makes sense, or at least tactic should complete different without VC-money.

> I think it's an example of a company suited for venture capital.

I guess it makes sense from the investment angle. Ie. need to grow, and need to grow fast, therefore need for lots of cash, therefore need for risk takers.

But at the same time not so much so for the angle of is growth realistically possible for this type of company? It took ARM over 30 years to organically grow with ups and downs to what it is today, in an age where the foundry landscape was much less complex wand wild chip design innovations were much more common. (We had SPARC, PowerPC, MIPS, PA-RISC, Alpha AXP, i960, 68k, Itanium, Transmeta in the nineties.)

From what I read, SiFive had broad ambitions and wasn't very focused. Not sure if it is viable, but expecting it to turn $$$ into competitive chip designs in a couple of years was a bit of a stretch.

68k is 70s

sparc, mips, pa-risc, i960 are 80s

arm is 80s too

transmeta didn't ship shit until 02000

i'll give you alpha, powerpc, and itanic tho

You're absolutely right of course and I stand corrected.

I was thinking of the cool, companies (Silicon Graphics, Apple, NeXT, BeOS) and high end stuff (SUN, Hewlett Packard, DEC), of the pre-WinTel era. When CPUs had one core, integrated memory and periphery controllers and even floating point co-processors weren't ubiquitous and "graphics cards" did not include a "GPU".

Acorn worked with a single company (VLSI) to produce the first ARM chips. When ARM came of age, there wasn't a complex "foundry ecosystem" where dozens of companies specialize in different stages of what results in a SoC.

i feel like the complex foundry ecosystem makes wild stuff a lot more doable now

from my perspective i see a lot of totally wild stuff, not all of it successful: yosys, esp32, fram, reram, optane, tensilica, cortex-m0 socs, padauk's fppas, gigadevice, apple's m1 and m2, graviton, tpus, ambiq's subthreshold utter insanity, stm32g, the unbelievable explosion of photovoltaic, wch's ch32v003, gallium nitride, mram, wlcsps, jlcpcb's smd assembly service, mass-market lidar chips with picosecond timing, chalcogenide pram, greenarrays, silicon carbide, modern silicon mosfets (not to mention igbts), led streetlights, petabit interconnects in data centers, nvme ssds, the zillion variations of risc-v, bitcoin mining asics, oled displays (though those aren't chips), gpgpu, submillimeter phased-array mass-market products like starlink, indium phosphide amplifiers in oscilloscopes, amorphous-silicon-on-glass products from sharp hitting the mass market (look, ma, no cog!), and skywater's open-source pdk and the associated shuttle program

so i have a hard time agreeing that "wild chip design innovations were much more common" in the 01990s than now; from my perspective that could hardly be farther from the truth

"victim of venture capital", who do you expect to fund such a risky business?

RISC-V has a lot of potential to eat ARM's lunch and adoption by companies like Qualcomm signal a future industry shift to it. The problem for SiFive is none of that requires them to be around to see it happen. The success of the startup and ISA are not inextricably linked. They did a lot of important work early on to cheerlead for RISC-V and to build much needed support and infrastructure for it. But this feels to me like the end of the beginning for RISC-V and a transition to a stage where the industry will drive it from here.

The other day Qualcomm announced a RSIC-V based Android wearable SoC. Expect to see full smartphone SoCs in shipping devices in a few years. ARM's attempts to extract more value after going public and lawsuit with Qualcomm is souring their relationships and driving customers to RISC-V.

But this doesn't answer the question as to where firms go if they want to license a core rather than design in-house. Arm grew in large part because firms decided that designing CPUs cores wasn't a central competency and Arm would do a better job. And for a long time even a firm with the resources of Qualcomm couldn't outperform Arm.

Maybe (say) Tenstorrent will take up the mantle but just saying that 'the industry will drive it' doesn't really describe how Arm gets replaced.

> And for a long time even a firm with the resources of Qualcomm couldn't outperform Arm.

For a long time (around 2010 to 2016) the internally developed Qualcomm ARM core did outperform the cores you could license from ARM. Qualcomm's internal team fell behind and then Qualcomm started licensing the standard ARM designed cores. Now Qualcomm acquired Nuvia and wants to develop their own custom cores again. We will see how it goes.

Most of the old Qualcomm ARM CPU team went to Microsoft for a few years, got laid off, and now actually work for ARM.

Thanks for the correction / clarification. I agree 2016-2021 isn't really a long time!

Of course the Qualcomm's use of Nuvia cores is subject to the ongoing legal case so it's difficult to see how that will work out.

Qualcomm's Znew proposal is basically an attempt to skirt ARM royalties while keeping a design that is very close to ARM64 (basically, throw out 16-bit instructions then add some new modes and complex 32-bit instructions ARM uses to make up for it).

This smells of them trying to convert Nuvia from ARM to RISC-V so the entire Nuvia case basically goes away before they get forced to pay out a lot of money.

Arm's case isn't about the ISA, it's about use of IP developed by Nuvia with help from Arm under Nuvia's Arm license being used in a way that isn't compatible with the terms of Nuvia's license.

IANAL but I'd be astonished if they get away with that IP making its way into RISC-V designs, if Arm win.

Plus, they won't have anything RISC-V based available in anywhere near the required timescales. They're still arguing about the ISA after all!

ARM's case is that a uarch is necessarily tied to its ISA and that their license isn't transferrable because the contract plainly says so.

If Qualcomm can show up and say "here's our uarch running a different ISA", it disproves that point and at most leaves dispute about some patents where Qualcomm can probably get a quick settlement for far less than the royalties would cost them.

The whole point of Znew is to transform RISC-V into something so similar to ARM64 that they can swap out the decoder and be good to go.

Qualcomm was still developing a 32-bit ARM core and then Apple came out with their internally developed 64-bit ARM CPU and surprised everyone.

This Qualcomm exec got shoved aside after making this statement.

https://www.pcworld.com/article/447916/apples-64bit-a7-chip-...

Qualcomm quickly switched to licensing a 64-bit core from ARM which was the Snapdragon 810 that had overheating problems. After that the next Snapdragon 820 with Qualcomm's internally developed 64-bit CPU was fine.

That same team was also developing their ARM server CPU but that whole project got cancelled and the team got let go around 2018.

But then every generation after the 820 used a Cortex core so their in-house core only lasted one year.

Yeah, that team switched exclusively to the server chip which was cancelled a few years later and they laid off 500 engineers in that team.

Decoupling the ownership of the ISA from ownership of the silicon means chip makers can compete more openly. Anyone with chip design capabilities can design and market their own premade RISC-V cores or offer custom silicon design services.

With ARM they acted as a gatekeeper through their licenses. Anyone who wanted to design ARM based chips needed a license and their licensing structure has gotten more complicated and expensive over the years.

Sorry, this completely misses the point that many firms don't want to design their own CPUs, which is why Arm became so successful in the first place.

No I'm not missing that point. My point is Qualcomm, Samsung, Intel, whoever doesn't have to pay a gatekeeper a license fee to design RISC-V cpus. They are also free to design and sell premade cores similar to how ARM does it today, Sans ISA license structure.

> Free to sell pre-made cores

You do know that Arm makes a tiny amount on each core licensed and that none of the firms you mention will have any interest in that business model. If they’ve spent a lot of money on a competitive design they will not be handing to competitors for peanuts.

Are you saying Qualcomm doesn't have interest in selling cpus to customers without paying a third party a licensing fee? That doesn't make any sense. Companies are always interested in lowering their overheads.

> But this doesn't answer the question as to where firms go if they want to license a core rather than design in-house

My original comment. Qualcomm won't be doing this. Or at least not for a fee that is in any way comparable to what Arm charges.

I mentioned it elsewhere but a big draw is FPGA soft cores or others forms of embedded CPU stuff. ARM licensing terms are bonkers so riscv has a major advantage if you need something better than a NIOS II or Microblaze (soft cores that Intel and Xilinx provide respectively).

with so mych of xilinx' lineup having hard arm cores built in, what is the driver for soft cores?

Flying radiation hardened FPGAs is a big one. Just because a chip exists doesn't mean it's usable for any given application.

Ease of use. The area of a small soft core is trivial. And having a small CPU core all for yourself and dedicated to manage one piece of hardware without having to deal with an operating system and other processes interfering is incredibly convenient.

See my other comment.

I worked at a company that designed whole systems from chips on up. I thought it was odd that they chose MIPS as a base architecture, until someone explained that the alternatives all had far higher licensing costs. IP licensing for other parts of the chip was already their second biggest cost after payroll, so I guess that made sense.

I feel like it will converge to just having RV + FPGA chips rather than people wasting precious FPGA space on softcores. Zynq's are plenty popular already

Easier said than done. Lots of Zynqs on orbit and elsewhere right now but if you want to do anything at all custom you're stuck with soft cores because regardless of who you are, I know for sure you don't have the money to do a custom chip.

It's common to have high performance CPU cores for the main tasks yet still have small microcontroller cores sprinkled all over the place for specific sub-block management, even in FPGAs.

For example, a small 5-stage pipelined VexRiscv CPU take around 2000 logic elements. That's nothing in today's large FPGAs. Add, say, 4KB of RAM and put it right next to the complex HW core that you want to manage and you have the equivalent of a complex FSM that's programmable without any worries about having to share cycles with other processes and missing out events with hard real-time requirements.

It's done all the time, and it's great.

The latest microblaze is already RISC-V under the hood, but of course other vendors can be used too.

Reply on news.ycombinator.com