Arm is offering early-stage startups free access to its chip designs | Hacker News Reader

Arm is offering early-stage startups free access to its chip designs

techcrunch.com

77 pointslihaciudaniel35 comments

Hah... More importantly if you're a silicon startup and you need a CPU core for some brand new design, you are probably not tied to a specific instruction set.

The obvious one to choose in that case is RISCV because it already has a working compiler and Linux port, has a thriving community, and there are no paperwork delays to get a prototype built.

ARM is making its stuff free to try to compete with the free RISCV, in the hope that when that startup becomes a big company, migrating to a new architecture will prove too hard, and they'll be locked in.

Yea this method has worked for everyone else:

* Oracle with Databases, Servers and Tooling

* IBM with servers, COTS packages

* Microsoft with OS, ActiveDirectory, Skype/Teams etc.

* AWS / Google / DigitalOcean etc.

Is your argument that every tech business has a free dev/test tier?

I don't see an issue with Arm giving free stuff (I think) to startups, but if they started paying (bribing) startups to use ARM designs instead of something else, even RISCV, then there should be huge repercussions.

ARM will have to eventually support RISC-V before they succumb. If they innovators dilemma now, they might save themselves, but they would need to figure out what they are to become. The moat is no longer around the ISA, they should be doing design engineering and delivering systems, and selling systems and components.

If ARM doesn't transform, they will die like all the other Unix vendors did when Linux took their lunch.

I think there is far larger chance that RISC-V will fail to materialize than ARM will eventually support it.

It depends for what you need that core. If you want to build a mobile phone, a router, a NAS, a set top box, a server, you might be better served by ARM since it has much better software support and their cores are already optimized for some scenarios.

SPARC is also free like RISCV but the greatest SPARC manufacturer, Fujitsu, is moving over to ARM 64.

It doesn't matter if an ISA or a core is free if it will cost you very much to make optimizations so it's fast enough and efficient enough, to fit it for some fab like TSMC and have the software ported and bugs sorted.

RISC-V has a nice system for extending the architecture, either in public or with vendor extensions.

This is the opposite of what you want for a having nice software tools.

The way it works is your software tools let you write the basic RV64GC code, and you use a bit of inline assembler for your custom instructions. Usually the custom part of the code is small and hidden behind a library.

You are correct, right now Arm has better software support etc.

But how long will that last? The competition is not happy paying Arm large fees, and it may well happen in the long run that RISC-V will have a better eco-system. 5 years? 10 years? Arm is a small company (~6k employees) in comparison with some of its competitors. Arm giving free access is a sign that they are beginning to worry!

Moreover there are geo-strategic considerations for some countries: what if global tensions rise and Arm gets strong-armed into withdrawing from selling to e.g. Baikal or Huawei?

>ARM is making its stuff free

As I've read they only provide free access to some of it's cores but you will still be required to pay the licensing fees if you use them.

It's like being able to read the source code for free but pay if you use the software. Pretty much what Unreal does with it's game engine. You can read the code for free, toy with it how much you like, build anything you want but at the moment you gain money from it, they want their share, too.

RISC-V ISA is just a programming interface and non-physical IP.

The license fees for ARM Processor families are mostly for physical IP for microarchitecture.

ARM provides selection of 40+ highly optimized physical processor IP cores for different purposes. Maybe you get eventually similar selection RISC-V architectures but they are not free. It takes lots of full time work to test and optimize those cores. If you write processor core implementation in verilog and give it away free you have done just 10% of the work.

Arm's competitive edge will whither, since other companies are frantically working on RISC-V implementations!

Physical IP for up-to-date silicon fab nodes is never free. But SiFive is already providing it for RISC-V - and other vendors could easily get involved too, seeing as good high-level designs (that can be simulated in FPGA with relative ease) are available with permissive licensing.

Sure, but from the pov of a startup that needs a core, the choice is whether you pay SiFive or similar, or ARM.

> Physical IP for up-to-date silicon fab nodes is never free.

This is baffling. My understanding of how a startup that wants to make its own custom CPU would work is this:

- Write Verilog or HDL sources for your CPU

- Test your CPU in sims and FPGA's, fix any bugs in the Verilog

- Send the Verilog files off to a fab

- Write a check in the $1 - $10 million range for the fixed-cost component of production (i.e. the mask)

- Get a shipment of ASIC chips for like a $1 unit cost

- ??? (your specific business model)

- Profit

Are you saying nobody has open-sourced Verilog files for a RISC-V CPU? This seems like something that's easily within the resources of volunteers, and is probably actually less work than all of the kernel and toolchain support.

What the hell is "Physical IP". Last I heard you either had IP or you had just plain old physical property. Not a weird combination of both.

It sounds like some convoluted way of making an end run around first sale doctrine by saying you bought the atoms that make up that processor, but you didn't purchase the arrangement of them; which sounds absolutely absurd.

Or am I missing something?

It's an industry-specific term. "Physical IP", as in the opposite of "logical IP", does not refer to atoms, but to a low-level specification ("netlist") of the transistor layout of computer chip components. You could compare it to machine code that is optimized for a very specific processor.

That is juxtaposition of terms is unfortunate... but thank you for clarifying. It's still IP, just in the form of a specification for physical components.

Better explanation:

Physical IP is the CPU design at the transistor level.

Logical IP is a spec document saying how the processor should work, plus a bit of pseudo code and maybe a simulator or two.

I think RISCV will just end up occupying the same space that MIPS did before --- ultra low cost (and performance) devices.

Which is funny, really, since for a long while MIPS provides amazingly high-performance parts used in Silicon Graphics (SGI) workstations.

As someone who is far removed from the hardware and instruction set relationship, when would one need a new processor and instruction set? How do you know you're at that point?

You want to make a new ASIC for some application and you don't want the hassle and cost of having to negotiate a license with ARM. Recent RISC-V chips that I've seen have been used for machine learning for vision in cars, camera controllers, crypto co-processors, hard disk controllers, massively parallel co-processors (also for ML), GPU controller, "hard cores" embedded in FPGAs.

The real chips doing all those things in actual applications are not RISC-V

So far. I wouldn't bet against free and zero friction however. Arm themselves are clearly worried otherwise the announcement in this article wouldn't have happened.

Yeah – ARM is just making it easy to get access to their IP while you’re in the strapped-for-cash startup mode. Once you start producing and selling you’ll be on the hook for licenses like anybody else.

i actually just started a job in hardware in sv and i see most people reaching ARM, its a nice set

This normally costs $75k/year. It's the "Entry" level of the ARM Flexible Access program [1].

It's limited to producing one chip per year. This seems low for a startup which needs to iterate.

They do not provide a DRAM controller. Startups will have to license one. There is a BSD-licensed controller: LiteDRAM [2].

This ARM offering helps with the first stage of creating an SoC. It's probably about 20% of the effort. The remaining 80% requires a lot of expertise and expensive software and prototyping hardware. University of Washington's Bespoke Silicon Group is developing free tools for the rest of the process: http://bjump.org

[1] https://www.arm.com/-/media/global/products/flexible-access-...

[2] https://github.com/enjoy-digital/litedram

Slightly off topic - While I broadly understand cpu architecture and instruction sets, I've often wondered how advanced CPU / GPU chips are actually developed and manufactured? And why are there so few major foundries, like Intel and TSMC (the only 2 I know)?

Chip manufacturing has the problem of being

A) incredibly difficult

B) incredibly capital intensive

C) low margin

Theres smaller fabs out there, but if you're going to do a new start up, why would you choose to compete with these guys, unless you think youve come up with a new process somehow that will beat them? The other fabs out there are also part of huge companies that do lots of other stuff (Samsung) or were spun off from them (GlobalFoundries/AMD).

IMO any new major foundry would end up being a strategic geopolitical move with heavy subsidies rather than a pure business play.

Obviously this is the case wrt. the newest process nodes, but there's plenty of older technology that's perfectly usable for some applications. Even preferable in some cases e.g. for analog devices.

Oh totally. But I believe OP was mostly talking about the most obvious consumer facing stuff (phones, computers, GPUs etc)

I doubt the DIP packaged 555 timers I played with in college were using a 14 nm process :D

TSMC is reportedly investing $20bln in the plant for the next node (more advanced, smaller, faster, less-power-hungry technology). Few companies can invest that kind of many. And for the ones that could, it's often better to invest it elsewhere and use the services of a multiple-client but single-purpose, laser-focused fab like TSMC.

Latest fab costs around $20 billion. You need to manufacture in volume and take huge risks. The know-how inside fab companies goes far beyond patents and IP. Chinese are trying to get into the game with great effort, but they are still struggling at least 2 generations behind.

Until just recently there were 4 players in the hidh end game GF, Intel, TSMC and Samsung. GF (Used to be part of AMD) quit the game in the middle of their 7nm effort because they could not keep up and changed their focus. Intel stumbled and partially failed their 10nm effort.

Reducing transistor size yields a quadratic increase in transistor density, reducing cost per chip. [1]

Chips with smaller transistors use less power. [2]

Therefore foundries with the smallest transistors get more orders than competitors. They use the revenue to develop the next generation of tech. The revenues are huge. For example, TSMC earned USD 35,000,000,000 in revenue in 2019 [3, 4]. Starting a new company to compete with TSMC would require possibly the largest investment in history.

[1] https://www.icknowledge.com/news/Technology%20and%20Cost%20T...

[2] Section 2.5 Power - https://www.doc.ic.ac.uk/~wl/teachlocal/cuscomp/notes/chapte...

[3] TSMC Annual Report 2019 page 91 - https://www.tsmc.com/download/ir/annualReports/2019/english/...

[4] https://www.xe.com/currency/twd-taiwan-new-dollar

Reply on news.ycombinator.com