Hacker News Reader

Comment by Laaas | Hacker News Reader

No. 6: How can an ISA "be closed"? It seems to me like all ISAs are open by definition, since they are simply an interface between the CPU and human (essentially). See e.g. https://en.wikipedia.org/wiki/List_of_x86_manufacturers. Many of these haven't acquired any "license" from Intel to implement a processor that complies to the x86 interface. If things are as I assume, I find it disingenuous to portray RISC-V as an "open" ISA since it implies people can't build on top of other ISAs. If there's something more to it, it would be helpful if someone could explain.

I still think RISC-V is superior to x86(-64) FWIW.

View on HN

You cannot sell a ARM processor that implements custom instructions.

Or more properly to sell* ARM processors (or devices with ARM processors) you need a license from ARM where they give you permissions and (unless you are Apple) ARM is not going to give you a license where custom instructions are not forbidden.

RISC-V is open in the sense that you do not need permission from the IP owner to use it.

* I believe that you probably can manufacture custom ARM processors without a if you never plan to sell them.

I think its possible that you could sell this processor but that you can't sell it as an Arm processor - in the recent Arm / Qualcomm case Arm seemed to rely heavily on trademark infringement rather than Qualcomm infringing their IP wrt the ISA.

Happy to be corrected though if I've got this wrong.

> You cannot sell a ARM processor that implements custom instructions.

That was the case if you are an ARM architecture licensee. And of late, that's not even the case anymore.

https://www.arm.com/technologies/custom-instructions

What do you think about the unusual choice for mask register semantics in RISC-V V which apparently forces chips to include silicon to support semantics nobody will actually want? https://www.computerenhance.com/p/the-problem-with-risc-v-v-...

It's not "unusual"; using data registers for mask is a valid tradeoff especially for low-end implementations, whereas higher-end architectures can easily use shadow registers. Discussed in depth at https://github.com/riscv/riscv-v-spec/issues/811

It's not "using data registers for mask" that is claimed to be unusual, it's requiring a the mask bit for the 64th byte of a comparison to be usable as as part of the 8th byte of a subsequent addition. I have read that, and it seems like the motivation is that someone who has written a PhD thesis thinks that people don't actually need filtering in their SIMD kernels, which is obviously ridiculous for general-purpose programs manually written in terms of vectors, but which may have been true for the particular programs under study.

It's complicated by patents for some of the low level implementations. Typically, you'd license x86 from Intel if you want to build an x86 chip, whereas there are no licenses/patents around RISC-V as far as I know.

The huge advantage of RISC-V is the lowered barrier to entry for chip manufacturers. It's now becoming possible for tech companies to make their own processors for their specific hardware and requirements.

In my mind, the even bigger advantage is that the door is now open for open source RISC-V implementations, so that people/companies can work together on high performance/low energy CPUs. It seems such a waste of talent for different companies all spending time and effort to implement similar features in their chips, when they could be building and extending open source chip designs.

So, what you're saying is that you no idea what's involved in actually manufacturing these chips?

I am not involved in chip manufacturer, so my domain knowledge is very limited.

What in particular do you think is incorrect in my comment and can you provide a correction please?

You can definitely license RISC-V implementations as in actual cores/silicon IP. And you can definitely patent specific hardware implementations of a RISC-V CPU.

For example whilst both AMD and Intel produce X86 CPUs and share the license they also have numerous patents for their specific implementations e.g. memory controller, prefetch unit, branch predictor etc.

The idea that the RISC-V ecosystem isn’t going to get rather quickly bogged down by patents is too optimistic unless I’ve missed something substantial in their licensing terms that prohibits RISC-V ISA users from issuing patents on their in silicon implementation.

Building a CPU today without having to work around patents is already rather hard and eventually there will be only so many actual hardware implementation of the RISC-V ISA that are both efficient in terms of performance and manufacturing.

So whilst the ISA might be open the secret sauce of how to actually make something useful out of it isn’t going to be open or free.

That last point is the point a lot of people miss. RISC-V is permissively licensed. The example of, "it is free like Ethernet" is not a great example. The best example is, "it is free like FreeBSD/NetBSd/OpenBSD." People can choose to use the open aspects of it. Or people/companies can choose to use that as a base then do whatever they please on top of it. Much like Sony can take FreeBSD, make an operating system for the console and not share any part of it except by giving a mention that they used FreeBSD somewhere in their stack.

The difference is that you need an x86 license to produce a processor that uses the same instruction set. IIRC there are just three companies that are licensed - Intel, AMD and possibly VIA if they're still around. It's not just the implementation that's patented.

Whether or not RISC-V implementations become encumbered by patents, at least it is now possible for companies to make their own processors without having to create the entire ecosystem too.

Because companies will want to use their own custom chips as a competitive edge. If google started making streamlined RISC-V CPUs, they could drown out a lot of RISC-V competition because Google has the capital to acquire some of the best to design their chips. Why give away what they could turn into a competitive edge away for free? As a lover of open source, this sucks and I wish it were different, but it is the reality of business. Especially since making hardware is so much more expensive.

Nowadays, I doubt that there is a huge advantage to having slightly more performant CPUs - companies typically would benefit more by incorporating custom requirements into the chip so that they can reduce chip count and size etc. Arguably Google has the resources to make their own performant CPU (with blackjack and hookers) without needing to use RISC-V anyway.

Intel will sue you if you don't have a patent portfolio to fight back with.

Reply on news.ycombinator.com