Alibaba Crafts A 16-Core RISC-V Chip @ 2.5GHz

[tuxd3v]

I'm not sure why you seem to think the only notable CPU-bound workloads are common video codecs and crypto.

If I'm sitting around and waiting for my computer to do stuff, there is a very non-negligible probability that it's spending lots of time in vectorizable inner loops, and I can't exactly petition intel to add a hardware block for some numerical library I'm developing.

Yes for some type of workloads, SIMD is very nice,
And when there are no hardware extensions, to accelerate workloads, its the best/only option..

I am not saying that, you should ignore SIMD, on corner cases, nor on some workloads..
What I was saying is that,
To measure core efficiency, you should discard mechanisms that doesn't make part of the core..

But AES, H.264, H.265, VP9, MPEG2, and such,
Should be included in the core has a hardware extension in Consumer Electronics CPUs, they are massively used..
And people should vote with their wallets on that too,
Because even tough that simd, is a general purpose compute mechanism( that helps to accelerate demanding workloads ), it should be used only when there are no hardware extension for that feature you need( assuming that what you need is a math library, a compression algo, and so on.. )

That been said, I would love to see lz4 has a CPU extension( of course that could be a treat to hardware vendors so sell RAM mem.. )

[audir8]

No, that's not what I'm saying.

I don't know how to phrase it better than I already did. I think you just kind of "shut down", when the topic of patents arises. It seems you can't shake your association between them and the big, evil empires. But they're really just a tool - not so different from copyright (which is also used by plenty of big, evil empires).

A method of operation can't be copyrighted (ie. an API or ISA). Logic in software and hardware is math and not patent-able. Copyleft ensures restrictions to user freedoms are not added by downstream redistributors. Disagree with your wording and premises? Yes I do.

Mickey Mouse might finally become public domain, and OIN exists for patents. GPL hardware hasn't found a great footing yet, but it still might. The trends in open source and in hardware and software in general have been and are still towards copyleft and patent-left for companies and projects of almost any size apart from the persistent trolls.

[coder]

Logic in software and hardware is math and not patent-able.

So, why do companies like Apple have to buy a license to implement the ARM ISA? I'm not an expert on the subject, but I believe the answer is that ARM holds patents needed to efficiently implement various aspects of that ISA.

Copyleft ensures restrictions to user freedoms are not added by downstream redistributors. Disagree with your wording and premises? Yes I do.

Copyright is a legal construct. Copyleft is branding of a specific license. When I say copyright, I mean copyright, even when it's used as a foundational component of Copyleft. Without copyright, there can be no Copyleft. If nobody owns the original IP, then you cannot restrict what downstream parties do with it. It all just collapses into a pile of public domain.

GPL hardware hasn't found a great footing yet, but it still might.

You cannot have something like an ISA which restricts all implementations to terms like those dictated by GPL, if you don't have any claim of legal ownership to the ISA and its derivative works. The only way to do that is with patents. If you're unwilling to consider such a use of patents, then you're foreclosing this avenue. That's all I'm saying.

Without this, GPL hardware is limited to whatever implementation someone creates, in which others see enough value to use as a starting point for their own efforts. The problem with this is that, unlike software, the one-off costs of producing a chip are so large that the amount you save by taking someone else's design as a starting point are much less. And the benefits are less, because end-users won't have the means to fab their own chips, whereas anyone can compile an OS kernel, etc.

[coder]

did I? I thought I was talking of this specific usecase, let's see what I said:

hardware design is extremely expensive, and Alibaba isn't a hardware maufacturer, but a service-oriented company, similar to Google.

This is one of the few situations where going opensource can actually makes economic sense.

You never made this economic argument.

AI hardware is very different from general-purpose CPUs so I don't need them to state their usecase with a lot of detail as it's already obvious.
You won't do shit in AI with a processor using 16 high-Ghz cores, you need thousands of small/weak cores. Look again at the custom AI hardware you cited. It's more similar to a GPU.

I guess the closest parallel would actually be Apple's phone SoCs. But my point is that of Alibaba's peers, you can find both edge, server, general-purpose, and domain-specific chips that not one of them has ever opensourced. If the economics would favor Alibaba doing that, then why haven't their peers ever tread that path? You can find (and have done) specific differences with each of those examples, but they've basically bounded Alibaba's effort, on all sides.

I said "big and powerful designs", they are not. They are for embedded, not for computing.

As the article points out, Alibaba said their chip is for edge applications.

And that does contradict what I said in what way?

You said it was tied to their specific algorithms and needs.

custom AI/neural-network accelerator hardware because that's usually very specific to their own AI software designs, that are not shared,

Their hardware is tailored to their own AI software, then they offset the costs and make money on the side by also allowing others to run other AI software in there that will of course not run as well as their own stuff.

See, this is just blind speculation. You don't know this. You cannot possibly make this case. You're just arguing this point to be contrarian--not because you actually know it to be true.

When they launched their TPUs, there was nothing on the market that could deliver anything like the same level of energy-efficiency. If you'd read even the first thing about their TPUs, you'd know that's where it came from. There were other things that fit their usage model & software, but nothing so efficient that fit anybody's usage model or software.

Well, Intel and AMD and IBM aren't rebuilding their designs from scratch for every new process node.
Yes you need to port stuff over but it's still evolving a base design over time.

You have to go a long ways to get from "opportunities for design reuse" to saying it's economically advantageous to cede the competitive advantages of keeping their IP proprietary. You have not covered that gap.

I already cited my examples, you are just rejecting them because they aren't from the same field or purposely misreading stuff.

Okay, then your examples are irrelevant, because they don't speak to the economic peculiarities of chip development and manufacturing. Honestly, even my worst example is more relevant than any of those. That's pathetic.

Really, it's one thing to idly speculate. We all do it, to some extent. But when somebody calls you on your shit, either put up or shut up. It seems to me you just like to argue the point, to see if you can.

[audir8]

So, why do companies like Apple have to buy a license to implement the ARM ISA? I'm not an expert on the subject, but I believe the answer is that ARM holds patents needed to efficiently implement various aspects of that ISA.


Copyright is a legal construct. Copyleft is branding of a specific license. When I say copyright, I mean copyright, even when it's used as a foundational component of Copyleft. Without copyright, there can be no Copyleft. If nobody owns the original IP, then you cannot restrict what downstream parties do with it. It all just collapses into a pile of public domain.


You cannot have something like an ISA which restricts all implementations to terms like those dictated by GPL, if you don't have any claim of legal ownership to the ISA and its derivative works. The only way to do that is with patents. If you're unwilling to consider such a use of patents, then you're foreclosing this avenue. That's all I'm saying.

Without this, GPL hardware is limited to whatever implementation someone creates, in which others see enough value to use as a starting point for their own efforts. The problem with this is that, unlike software, the one-off costs of producing a chip are so large that the amount you save by taking someone else's design as a starting point are much less. And the benefits are less, because end-users won't have the means to fab their own chips, whereas anyone can compile an OS kernel, etc.

You're really just furthering my point that patents are only used in hardware to strengthen monopolies. If RISC-V did try to get patents, I'm pretty sure Qualcomm and Intel wouldn't be funding it or letting it grow as much. I've said it before and I will say it again: patents are a scourge on open source, and the less of them the better. They hold back innovation, and provably so:
1) We have instances of Google or Facebook giving patents to some open source communities, but that is already going too far if you consider that this causes friction with other projects and at worse leads to people re-inventing the wheel.
2) OIN is the answer to patent-trolls, which nobody like an Intel or Qualcomm or ARM will join since they have monopolies to protect.
Patenting anything at all seems like more work for an open source project without any real payoff and the added risk of being sued by a major corporation. Why bother?

I don't think Google will lose it's appeal and copyright APIs or ISAs will never happen. I never said copyleft isn't based on copyright. I just gave a better understanding of copyleft.

The alternative to having no copyright or patents on an ISA would be to have an ISA with low/mid/high end designs+peripherials be GPL with exceptions as needed. You could make a GPL exception for adding instructions, and even for ASIC implementations which run on a generic BUS. As long as there is a whole ecosystem of GPL CPU/GPU/SoC/Peripheral implementations, everyone will use whatever is cheaper and more convenient, and contribute to the GPL'd parts like they do with Linux and GCC. No patents required.

The cost of making a working toolchain with kernel support clearly took RISC-V some time. I wouldn't call it negligible compared to fabbing a chip. Fabbing a chip at older process nodes has only gotten cheaper, and this will continue to happen with the tools making it easier to do a tapeout as well. There will always be some startup-costs, but there are more and more hardware startups which can afford to do this with the backing of the open source community and corporations which rely on open source.

[coder]

You're really just furthering my point that patents are only used in hardware to strengthen monopolies. If RISC-V did try to get patents, I'm pretty sure Qualcomm and Intel wouldn't be funding it or letting it grow as much. I've said it before and I will say it again: patents are a scourge on open source, and the less of them the better. They hold back innovation, and provably so:
1) We have instances of Google or Facebook giving patents to some open source communities, but that is already going too far if you consider that this causes friction with other projects and at worse leads to people re-inventing the wheel.
2) OIN is the answer to patent-trolls, which nobody like an Intel or Qualcomm or ARM will join since they have monopolies to protect.
Patenting anything at all seems like more work for an open source project without any real payoff and the added risk of being sued by a major corporation. Why bother?

I don't think that proves patents can't be used to further a GPL agenda, when Google and the backers of OIN aren't exclusively pro-GPL.

The alternative to having no copyright or patents on an ISA would be to have an ISA with low/mid/high end designs+peripherials be GPL with exceptions as needed. You could make a GPL exception for adding instructions, and even for ASIC implementations which run on a generic BUS. As long as there is a whole ecosystem of GPL CPU/GPU/SoC/Peripheral implementations, everyone will use whatever is cheaper and more convenient, and contribute to the GPL'd parts like they do with Linux and GCC. No patents required.

The cost of making a working toolchain with kernel support clearly took RISC-V some time. I wouldn't call it negligible compared to fabbing a chip. Fabbing a chip at older process nodes has only gotten cheaper, and this will continue to happen with the tools making it easier to do a tapeout as well. There will always be some startup-costs, but there are more and more hardware startups which can afford to do this with the backing of the open source community and corporations which rely on open source.

Okay, you've made your position clear. We'll just see how it plays out.

According to what I've read, Alibaba did say they would open source some parts of their design. Let's see exactly what that is. SiFive has stated it wasn't practical for them to make a fully open source design, due to the need to license some IP for things like memory controllers. Hopefully, Alibaba will give away something like that, which could be a real benefit to other small hardware design projects (RISC-V and otherwise).

[SystemCrasher]

The fact that this design is proprietary is a perfect example of why permissive licenses are bad. If RISC-V was copylefted, people would be going through the publicly available design right now, and could consider grouping together to order these chips from a fab.

Hardware world now could be compared to software world of 80s-90s. Right now HW manufacturers are mostly greedy proprietary nuts.

Btw there're OpenRISC cores (and set of peripherals) that are copylefted. One can also put copylefted "extensions" on top of BSDish things to make whole thing copylefted. Sure, if someone would be unhappy with it, they can remove copylefted bits afterwards, to be able to close source. However, should copyleft improvement be good, it would be loss (think moving to BSD from Linux to get idea). Or just tightly intermingle copylefted design with rest of code - so it could be hard to remove without breaking things apart or copyleft parts sneaking in. However right now HW ecosystem is toxic proprietary minded swamp, and it would take quite some years to change that I guess.

Btw, less expensive ways to get sense of some CPU arch could be at least:
- Running that design in FPGA. Unfortunately FPGA world isn't what I would call opensource-friendly either. However, one way or another you can end up running cpu core in FPGA. This doesn't fully solves puzzle, merely shifts paradigms: instead of vendor lock on CPU manufacturer you get vendor lock on FPGA manufacturer. Also FPGA CPUs usually have relatively low performance compared to "dedicated" SoC of comparable size/power consumption. This said, FPGA can offload some heavy data movement tasks and so on, offsetting this problem - so some real world systems actually using this approach. But it suboptimal approach for e.g. desktops. However, it can be ok to run prototype and then create real SoC, that would enjoy much better tradeoffs (HW isnt like SW - starting production of bugged IC is very costly mistake).
- Another option to run VM. It doesn't fully solves underlying HW woes though - low level always wins, so if underlying HW isnt trustworthy, VMs are in jeopardy as well. So it mostly like convenient way to get sense of arch and fiddle with software built for this arch, even if you lack actual HW. It can also be somewhat slow.

[Guest]

Hardware world now could be compared to software world of 80s-90s. Right now HW manufacturers are mostly greedy proprietary nuts.

My thoughts exactly. The reason why I push heavy copyleft is to avoid repeating mistakes that happened to software back in the day.

[SystemCrasher]

My thoughts exactly. The reason why I push heavy copyleft is to avoid repeating mistakes that happened to software back in the day.

There is rather good saying: freedom of one ends where freedom of others starts. Somehow BSD-licensing fans fail to consider that - only caring about maximizing their own freedoms, without caring if it turns out badly for others or project itself. The result is obvious: if that really works out, one or very few strongest entities prevail, dominate markets and eventually get idea they are cool enough to no longer share, to thwart competitors and so on. That's how, say, "proprietary BSDs" looked at the dawn of opensource age. Either you get some abstract source you can't really use for pragmatic purposes or you have to deal with some greedy proprietary nuts. Then Linux appeared and has shown there is great demand for something that is both really open and usable - at once. So, say, I can grab kernel source, patch it the way it fits me - and use result where I see it fits. Even if I have to share changes eventually (I can even take it as benefit, proving dev had no ill intentions to customers is good example). That's what makes Linux powerful and empowers me when I use it. Actually I used these powers several times - and I hope it would keep going like that.

Most ironic part? Isn't it funny how some mediocre chinese company with crappy R&D but long bucks going to (ab)use efforts of US university and plenty of (all around globe) contributors - without being oblidged to contribute anything back at all? Oh, in past US companies also been like that, in software world. And still mostiy like that in hardware world, even though paradigms finally starting to shift in HW world, likely due to software world showcasing what opensource could really be and HW more and more turns to "special kind of SW" (likely because SW development patterns proven to be very efficient). This said, HW world right now is quite grim: backdoored proprietary crap dominates markets, companies like Intel, ARM and so on abuse their powers exactly same way it happened in software world of 80s-90s, feeding us with "Management Engines", "boot guards" and "secure boots". Where no feature would serve to owner and customer, only to manufacturers and their questionable practices.

Does someone wants "trial version of CPU"? How about "online activation" of hardware? Or maybe perma-blocking your PC remotely, because CPU manufacturer decided you don't deserve to use it for whatever reason? Actually e.g. Apple can do exactly that to iPhones. How about some nag screens? I can readily propose how GPU can show really-hard-to-remove nag screen on top of everything. "Pay us $5 to remove this advertisement for a month!". Sounds like a business plan?