Arm IPO to kick off today
theregister.com
theregister.com
The legendary Jim Keller is going all in on RISC-V, if you don't know Google him. His company has many core designs coming as well chiplets: https://tenstorrent.com/risc-v/
Because of Jim Keller and similar efforts I wouldn't be surprised for RISC-V to see both core count as well as per core performance meet ARM over the next few years. Maybe even exceed if Jim can push the chiplet approach faster than ARM can roll theirs out.
Hopefully this drives a lot of innovation and we all benefit as a result.
I think that using ARM is going to be viewed as being locked into ARM's ever increasing licensing fees, where as if you go RISC-V, you are free to switch CPU providers.
Idk, this really isn't as true as you'd think. Yes RISC-V is an open and free ISA which will cut out some fees, but you'd still have to license IP/chip designs (if you can't make your own) and they could only undercut ARM by a little bit. Further, the lack of mobility across chips/boards/whatever is not usually from the ISA, but from the BSP, SDK, etc. so you still would have substantial lockin unless we somehow standardize on that (lol)
Whereas once you have a free ISA which is actually in widespread use, you could get some designs out of universities or major corporations which intend to use them rather than sell them as their primary business and then release the design for the same reason they do for Linux code, potentially under a copyleft-style license.
Those designs aren't going to be competitive with the state of the art, at least in the beginning, but they don't have to be. All they have to be is low power enough to stick in an embedded device and fast enough to run the display on a refrigerator and the lack of a license fee would cause them to replace a zillion ARM chips that currently go into every IoS device and <$200 phone and consumer internet router. That's a huge chunk of ARM's market.
Then someone like Amazon evaluates this thing for something like the Kindle and finds that it's almost good enough, all they have to do is throw a couple of engineers at it for a short period of time, so they do and it gets released because what do they care of some non-competitor uses it in some cheap laptop. Commoditize your complement -- now laptops are cheaper and people buy more laptops on Amazon.
Meanwhile the "small project" uncompetitive boards start to have their own open source BSP and SDK under the BSD license, which means anybody else can fork it for their own thing, and that's a competitive advantage so it happens a lot. But now it's way easier to reverse engineer the code because 95% of it is unmodified, so you start getting community replacements for that stuff and with any luck the OEMs stop even producing it and point you to the open source repository.
It could be a long time before that takes over the high end, if that ever happens, but the low end? It's almost inevitable. And the high end is AMD64 and Apple, the latter of which has all the leverage in the world over ARM because they could always switch to RISC-V.
We've had open ISAs for decades. The problem isn't with ISA licensing, the problem is cost-of-entry.
Most people have a computer...it costs nothing to install Linux and GCC/rustc/Python/etc and start contributing to open source software. For open hardware, you need (at minimum) a decent FPGA for testing/development; but access to fabrication resources to be in anyway competitive.
Don't get me wrong, I'm generally pro on RISC-V; I just don't think it's the revolution everyone is making it out to be.
That wasn't even true when open source got started. Typical access was to expensive mainframes and VAXen available only in universities and major corporations. PCs were only starting to become a thing and most people didn't have one.
That model still works if you need something expensive. Get your university to buy one, or your company if it will save them $0.05/unit on the millions of devices they sell. Or you rent one in the cloud.
You also might not need your own. A project gets started by someone who does have access to a decent FPGA, you want to contribute to it, great. They give you remote access to the FPGA.
Meanwhile FPGA prices keep going down.
> but access to fabrication resources to be in anyway competitive.
Which is available to anyone, if you have customer demand. They've made millions of Raspberry Pis.
Which is why it took off when computers became accessible; and continued growing with even further ubiquity.
> You also might not need your own. A project gets started by someone who does have access to a decent FPGA, you want to contribute to it, great. They give you remote access to the FPGA.
Cool, so we're back to bottlenecked accessibility.
> Which is available to anyone, if you have customer demand.
Not hobbyists. To reiterate, open hardware (including ISAs) is as old as open software.
You're delusional if you think either are anywhere equivalent.
So things like FPGA prices going down over time or being available to hobbyists via cloud providers.
> Cool, so we're back to bottlenecked accessibility.
Anyone with a computer and an internet connection can contribute, all they have to do is send an email to the project maintainers. It's obviously not as good as everybody already having the hardware, but it could be good enough to actually make it happen.
> Not hobbyists.
The model would be that there is a public repository with a design in it, anyone can contribute, and then every year or two they do a freeze and a production run which anyone can buy including the hobbyists who made contributions.
It feels like your argument is that Debian can't exist because a hobbyist doesn't have the resources to build a whole operating system. One doesn't, but a million of them do, especially if they can start with a basic design released by a major corporation or university or government.
You probably won't ever know. Do you know the architecture of the microcontroller on your first computer's hard drive? I don't.
The stuff that RISC-V will replace first is the low-margin hardware that can be mass-produced by China et. al and exported without license violation. ARM's goose was already cooked in this sense, and even ARM China won't fix it. For IO controllers, ICs and network switches, it's hard to see why manufacturers would stick with higher-margin ARM hardware. If RISC-V cores are cheap, stable and available, you could replace them without the user ever noticing.
> By the time Apple is shipping a RISC-V phone I hope to be retired!
Apple is reportedly looking to replace some of their A-series ICs with RISC-V ones: https://www.techpowerup.com/298936/report-apple-to-move-a-pa...
Hopefully you don't intend to retire this soon :p
Sorry, the architecture license wasn’t a thing when Arm was founded and there is no way that Arm would hand out free licenses at a later date. It’s just a myth.
I assume they paid something for it, but who knows how much, ARM and Apple have not decided to tell us.
Source: https://www.sec.gov/Archives/edgar/data/1973239/000119312523...
The people dealing with assembly know who they are, but much of that is the sort of work that cares a lot about shaving off a few cents worth of license fee for an embedded device.
With open source you could just recompile for the new architecture. Linux has always run on just about every one under the sun and the user largely can't even tell the difference. But if you don't have the source code you can't do that.
Closed-source systems could avoid this by compiling to bytecode, which is essentially what Android does, so transitioning to RISC-V would be much less trouble there than it would be for e.g. Windows.
But previous open ISAs were in the wrong market segment. Okay, SPARC is open, but it's also Oracle and no one trusts them, and the existing SPARC ecosystem is the enterprise market which expects big, hot hundred-thread systems and has no use for a simple dual-core CPU at 800 milliwatts. But that's the thing you can produce with a low budget, so it never goes anywhere or no one even makes one.
People are starting to use RISC-V for the thing where that actually works.
They're all open, albeit with somewhat varying definitions of the term.
PowerPC basically the same, not opened until after it was irrelevant. OpenPOWER is in the same enterprise market as SPARC and IBM is not far from Oracle on the "do not engage litigious bureaucracy" chart.
Actually> "Apple (AAPL.O) has signed a new deal with Arm for chip technology that "extends beyond 2040," according to Arm's initial public offering documents filed on Tuesday."
Which means that if Apple decide to switch to RISC-V to save money, the deal is still in effect, but no money will be paid, so there might as well be no deal.
For Apple, they won't offer much because they could develop their own CPU architecture ISA pretty easily (they control all code and compilers for the platform after all, and already employ lots of compiler and CPU developers).
Why not? It's not like Apple will have anywhere else to go for the next 8 years. x86 cores are too slow and power-hungry for the foreseeable future, and building their own risc-V core will take... at least 8 years.
* Expand our System IP and SoC Offerings. To enable further improvements in performance and efficiency, we continue to develop a broader set of configurable systems IP offerings, including proven on-chip interconnect, security IP, memory controllers, and other design IP to be used with our processors, including the integration of multiple IP technologies into a subsystem and additional information to assist in fabrication. More recently, we have invested in a holistic, solution-focused approach to design, expanding beyond individual design IP elements to providing a more complete system. By delivering SoC solutions optimized for specific use cases, we can ensure that the entire system works together seamlessly to provide maximum performance and efficiency. At the same time, by designing an increasingly greater portion of the overall chip design, we are further reducing incremental development investment and risk borne by our customers while also enabling us to capture more value per device.*
ARM doesn't just give you the architecture, they give you a reference design that already works and then lets you customize it however you want.
Companies like Apple don't care about that, but MediaTek, NXP, etc. use it to speed the design process meaningfully.
Integrators don't typically go all in on ARM IP though for a couple major reasons. 1) They want to provide some value add beyond the standard offerings or else why pick them over a competitor? 3) They have massive collections of IP blocks already that they don't have pay ARM for. Maybe it takes some engineering to convert to the next chip, but that's probably cheaper than what ARM is asking for.
So RISC-V really only needs to focus on CPU cores to be an existential threat to ARM.
At this point their biggest most is probably the patents on AMBA specs, but if they tried hard to enforce those the industry would switch to something like TileLink in a relative heartbeat.
Maybe in 5-10 years, RISC-V gains share in initial designs, but the semiconductor design process is loooong.
And if you're going to talk about Microcontrollers and simpler chips, keep in mind we had a bunch of other RISC architectures a decade ago that all lost to Cortex-M for simplicity/cost reasons (RIP MIPS).
But all that will be left in a few years is the high end areas and where there is a lot of lock in to the ISA.
It isn't yet clear when RISC-V will be ready to compete at the highest end of things, but that is where Jim Keller comes in.
RISC-V is a classic market disruptor for ARM, just as ARM was a disruptor to Intel.
That's not even true within the ARM ecosystem itself. The chips from Infineon are not source-code compatible with STM, STM is not compatible with Microchip, Microchip is not compatible with TI...
The problem is that the ARM core is just a portion of the architecture. Everything on top of that - GPIO, memory interfaces, timing, etc - is vendor specific, and will stay that way for RISC-V. RISC-V is just an instruction set architecture (with some appendages), not a blueprint for a complete CPU / MCU / SoC.
Not to mention, the chips also won't be electrically-compatible. Your hardware architecture can be as daunting to redesign as the code, if not more so. There's a reason why we try to do as much as possible in software, after all...
Not as true for RISC-V, as there's efforts (some of them complete) to standardize interfaces to standard peripherals.
E.g. timers, gpios and watchdogs.
Like, if we ignore the chip design aspects - is it guaranteed software compiled on one RISC-V chip will run on another RISC-V chip? (in the same way it does on ARM and amd64/x86?
What about hardware interoperability? PCI? UEFI/Boot semantics? Power management?
Who enforces that stuff for arm and x86 platforms?
Afraid I have bad news about ARM right there.
Refer to the RISC-V platform spec (ongoing).
It depends on other specs, some of which (like those related to the boot process) have already been ratified.
I don't think arm has anything standard around the boot process and vendors have various implementations
Now here we are 30 years later and I'm typing on a laptop running on a ARM (née Acorn RISC Machines, then Advanced RISC Machines), I have an ARM in my pocket, another in a tablet in my bag. My wife is next to me on an ARM laptop, she's also got an ARM on her arm... we're listening to music on a wifi speaker running on an ARM. There are probably 30-40 ARMs in one way or another around our house. Amazing really.
Newb /s
I cut my teeth on an Acron Atom in 1982, and upgraded to a BBC Micro (a renamed Acorn Electron) the next year.
But, yes, I never would have thought that, 40 years later, most of the CPUs in my house would still be derived from that scrappy British's company's designs.
Actually renamed Acorn Proton. Electron came later.
I bought some "Armh" (ARM's old ticker symbol) which did really well before they got bought by softbank. Oddly hard to find any info about historical stocks on the interwebs.
Valero Energy. Profit: $10b. [valustox.com/VLO]
DR Horton home builders. Profit: $5b. [valustox.com/DHI]
General Motors (!). Profit: $10b. [valustox.com/GM]
Aflac Insurance. Profit: $5b. [valustox.com/AFL]
Nucor steel. This one is selling for only $40b. Profit: $5.6B. [valustox.com/NUE]
Microchip Technology Inc. Also $40b. Profit: $2.4b. [valustox.com/MCHP]
Arm has it's work cut out to raise their profits from the rumored $1b. I'm not sure why an investor would consider them without a strong plan to 10x-100x their profits when there are companies selling at a similar rate but making much more money in boring product categories. Hope they can do it while staying true to their mission.
They don't manufacturer anything, correct?
I don’t know anything about Arm’s business interals, but I think they have more of a low-revenue problem than a low-margin problem.
According to [0] they had 2.5b in revenue and 0.5b in profit. A healthy margin, and much lower revenue than the other companies I mentioned.
[0] https://www.barrons.com/articles/arm-stock-price-ipo-caf090b...
https://www.theguardian.com/business/2023/mar/03/uk-chip-des...
One basic rule for making money as a country is to ensure capital flows through your country (e.g., services, manufactured goods, FDI, etc). Then skim off some of the captial that flows.
If ARM were listed in the UK and say, $30b of shares is traded anually, of which say $10b is in the UK, then some of that will go to UK ltd.
Are the licensees using parts of the actual chip design? Are their own designs too far down the "derivative work" rabbit hole due to not being cleanroomed that they have no hope of ever not licensing?
What are the specifics?
The crux of modern chip design is the tradeoff in MHz, peculiarities of Tomasulo's Algorithm (out of order buffers, tuning sizes and number of pipelines, etc. etc.).
Lets take an example: should you have 200 64-bit words in the reorder buffer, or should you have 800 (Apple M2). What's the tradeoffs? How much slower does accessing the reorder buffer get when you need to go from 8-bits to 10-bits to address the various locations?
How many multiplication units should you have? I know Intel has 3 of them per core, is that enough? Or do you go IBM Power route and go for like 20 pipelines wide?
Etc. etc. etc.
ARM Neoverse N2 makes a lot of these decisions, packages them up into an easy moniker ("General Purpose"), and also has customizations towards V-cores (higher performance but bigger) vs E-cores (lower-performance but smaller and more power-efficient).
You then make decisions based off of the core as a whole, rather than designing a core. Ex: do you use 128kB of L1 cache? Or 64kB? Do you do L1 / L2 / L3 cache like Intel? Or do you do L1 / L2 cache like Apple?
You still need to make these "uncore" decisions, including the important MESI (core-to-core communications: Modified data vs Exclusive data vs Shared data vs Invalid data). Even _IF_ you buy an off-the-shelf core like Neoverse N2, you're no where close to finishing an actual chip yet cause the darn thing can't even talk to RAM yet.
You can clean room implement the trade secret part, but the patents would be an issue and ARM could still sue and drag things out.
You also could never legally call it ARM because it's trademarked. This makes it harder for semiconductor vendors to sell chips.
See the Qualcomm lawsuit shitshow which is now causing Qualcomm to invest big in RISCV.
Cyrix and AMD had licenses to do so, ultimately deriving from a time when Intel needed second sources of their CPUs in order to win defense contracts.
Cyrix has also sued Intel over their Pentium chips. In the end the results of all the lawsuits are cross-licensing deals.
What is the legal interpretation of IP law that says you need permission to design a chip that implements an ISA?
In order to get the contract for the IBM PC, Intel had to agree to have a second-source manufacturer for their CPUs. Intel already had a relationship with AMD so it made sense to use them.
Cyrix and AMD got (cross-)licenses from Intel:
* https://www.sec.gov/Archives/edgar/data/2488/000119312509236...
* https://www.kitguru.net/components/cpu/anton-shilov/amd-clar...
This has been true since the very beginning:
> Early 1980s--IBM chooses Intel's so-called x86 chip architecture and the DOS software operating system built by Microsoft. To avoid overdependence on Intel as its sole source of chips, IBM demands that Intel finds it a second supplier.
* https://www.cnet.com/tech/tech-industry/intel-and-amd-a-long...
* https://jolt.law.harvard.edu/digest/intel-and-the-x86-archit...
The reverse engineering happened with Compaq doing the BIOS:
* https://www.allaboutcircuits.com/news/how-compaqs-clone-comp...
This was a major plot point in AMC's (very good) show Halt and Catch Fire:
* https://www.internethistorypodcast.com/2014/05/the-incredibl...
* https://arstechnica.com/gaming/2014/05/review-halt-and-catch...
* https://en.wikipedia.org/wiki/Halt_and_Catch_Fire_(TV_series...
My interpretation of this is that their investors suspect that whatever boost ARM is getting from AI optimism will likely peak soon, so the timing has to be now.
I know that's not fully rational because they're mostly unrelated, but certainly the optimism because of AI has to be good for them. Curious to see if this "signal" plays out - investors tend to be pretty savvy about timing.
The "investors" are SoftBank; they've been pretty much the stupidest money in the world for years.
While it's possible that they're being more savvy about the timing of this sale than about their investment decisions, I don't know that it should be the default assumption.
It's not as cut-and-dry if Softbank just puts the shares on the open market, and NVIDIA pays market value for them. Are there reporting requirements around NVIDIA buying stock on the open market? Freeze-out periods? Regulatory delays? If there's collusion it might be a crime but even if you're cynical and refuse to believe that there might not be, you still have to prove it.
"At this price, the IPO is reportedly set to raise about $4.9 billion for Arm’s parent company, Softbank, which is less than the $8-$10 billion the Japanese investment outfit had previously said it hoped to generate. Softbank itself posted a record $39 billion loss earlier this year."
In so doing, Son agreed to pay Son twice what Son sold it to Son for six years earlier."
Others do seem to be backing off, but there is still definitely a lot of way overvalued tech out there.
If NVDA's forward projections are correct, and they can maintain that same growth, then there is still room for upside. Personally, I'm not playing this game of musical chairs because I don't think these valuations are sustainable, but as they say, the market can stay irrational longer than you can stay solvent, so shorting is a terrible idea.
Other sectors do not price in the way tech does.
https://www.ifrs.org/content/dam/ifrs/publications/pdf-stand...
> Fair value is a market-based measurement, not an entity-specific measurement. For some assets and liabilities, observable market transactions or market information might be available. For other assets and liabilities, observable market transactions and market information might not be available. However, the objective of a fair value measurement in both cases is the same—to estimate the price at which an orderly transaction to sell the asset or to transfer the liability would take place between market participants at the measurement date under current market conditions (ie an exit price at the measurement date from the perspective of a market participant that holds the asset or owes the liability).
https://www.arm.com/glossary/isa
https://www.theregister.com/2023/08/31/a_star_star_domains/
https://www.theregister.com/2023/09/12/arm_markstedter_domai...
You can't call a company that literally developed the IP they patented "trolls."
"Patent troll" has the connotation of an organization or individual who buys cheap patents and sue other organizations on mostly bogus claims related to these patents, hoping that the other side won't want to bother and just settle the matter. That's not Arm's business model.
ARM doesn't have a big enough moat to stay static and just troll. There are competing ISAs, some free, and there isn't as big of an lock-in as for e.g. x86.
The linked articles are about the brand name, Arm, used in a domain. You wouldn’t put “intel” or “nvidia” in your domain name without drawing attention. Maybe we need a moniker similar to x86 to define the arm architecture more generally.
Computing is a complex business and few companies do everything in house. ARM's business model of focusing solely on the IP seems to have worked well for them. From the engineering side, it's certainly nice to have the ability to pick and choose which IP blocks we'd like to use and then shop around the design ourselves versus having to battle with Intel to get what we need while trying to poke through the various layers of obfuscation they tend to put in place.