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

[audir8]

You could say the same about copyright, but it's an essential component of GPL. You cannot license source code that's not copyright, because the copyright is what establishes ownership and gives the owner the right to assert the conditions of use.

Assuming you cannot copyright an ISA (at least, in a way that constrains implementations), then you need some other legal tool for establishing ownership and exercising control. AFAIK, patents are the only alternative.

BTW, the contradiction of GPL is that it preserves some freedom for all, by restricting total freedom for some. This requires control, which you can't exercise without some claim of ownership.


The way it would work is that you grant free use of the patents by any who follow the terms of your license.


Yes, of course. I don't know about proprietary, though.

https://en.wikipedia.org/wiki/RISC-V#Packed_SIMD

Sorry, but comparing copy-left with patents and saying copy-left doesn't foster innovation is BS. A third party can take MIT/GPL licensed code, add to it and do something new that's faster/better/cheaper. With patents, you're waiting on Intel/Qualcomm as the monopoly holders, or hoping to sue or not get sued by Oracle/Facebook. Copy-left has enabled innovation, and things like AV1, XZ, Zstd and even Linux are proof where almost nobody has to worry about patents anymore where they once did.

Considering SiFive has been funded in part by Qualcomm and Intel's VC arms (https://www.crunchbase.com/organizat...tion-investors), it makes sense for the RISC-V ISA to be non-GPL (what better way to compete with licensed ARM?). Linux and GCC have been successful GPL projects because they have convinced everyone the trade-off you mention is worth making, though I wouldn't call it a contradiction at all. It's just a trade-off for what you think should be a higher priority: getting contributions back vs letting people do what they want. People do share modifications in MIT/Apache licensed projects, and if they didn't some projects would die off. People also interact with GPL code using shims. You are not any less free to use a GPL project, you'll just have to share code back, and possibly forego some profit.

Google has appealed the ruling to the supreme court, so hopefully APIs won't be copyright-able. Ultimately, I think some combination of GPL with exceptions could work in hardware, OpenSPARC was/is GPL (https://www.oracle.com/technetwork/s...nload_for_Chip) and if there are patents involved, I think they an be worked around. The biggest roadblock to GPL hardware is probably that the EDA companies would lose their shit if I could just get a USB3 or PCIe IP for free from a GPL project.

[tuxd3v]

I'm running a stock threaded CoreMark build. So any difference would be tuning by Alibaba. But seeing the Xeon 10x faster than a RISC-V chip wouldn't be surprising considering it's a new(er) architecture and that the software support and compiler around RISC-V are far from mature, etc.

I think the confusion came from the fact that you are showing results in Iterations/s,
And we don't have then a way to translate to CoreMark/Mhz( Because we don't know how much threads you are running, and also the frequency.. )

Better to normalise to CoreMark/Mhz,
In this way, it doesn't mater, if you are running at 50Ghz, or only at 1 Mhz( the result already reflect it.. )

Iterations/s to CoreMark/Mhz
In rpi 4 example:
19513 Iterations* 4 /s @ 1.5Ghz

19513/(4 threads * 1500 Mhz) = 3.25216666667 CoreMark/Mhz

This value, seems to be too low,
I was expecting something between [ 4 - 6 ] CM/Mhz( I tough that you run 4 threads@1500Mhz.. ? )
Cortex a15 gives 3.6 - 5

Its not so easy to measure, because the OS, is also playing tricks..

One way, is bootargs= " isolcpus=x ..."

Then reboot, fix the frequency, and launch a thread in the x core..
Then you only have to divide by frequency

One table I found about Cortex a15( 2 different implementations.. ),
Aortex a72 is a improvement, but more power efficient.. and maybe a bit slower( no free launch's.. )

[coder]

Do I need to remind you the usecase?

No, because it's not relevant. You were saying that it would make sense for cloud operators to open source their chips and I pointed out that we have other examples of cloud operators making chips but none are open sourcing them.

And before you claim that I'm off-topic by talking about AI chips, let me refer you to the fact that Alibaba didn't even describe the use case for this chip as cloud. Quoting from the (very short) article:

Alibaba will be using this CPU for edge computing purposes and other domestic uses.

So, how's that relevant to what I cited? Microsoft made a custom HPU for their Hololens, that they haven't opensourced, and Google made a custom TPU for phones, that they haven't open sourced. Both are for edge, though. And at least Google's AI chip is somewhat general, in that they open it up for apps to use.

But, let's not get distracted by my examples. You're the one who needs to defend his assertion that this chip makes sense to open source. If that were so, then others building similar chips or chips targeted at similar markets would be doing it. Either show me examples or drop your case.

Other cited examples are custom silicon for AI/neural-network accelerator hardware. There is no standard implementation for that, no big and powerful designs you can buy from someone.

Another statement made in ignorance. A lot of folks are using Tensilica cores in their SoCs, for AI. Apple, for one. CEVA is another big licensor.

It's hard to make a similar argument for custom AI/neural-network accelerator hardware because that's usually very specific to their own AI software designs, that are not shared,

More ignorance! Anyone can rent time on Google's TPU cloud instances. It uses TensorFlow, which is an open-ended framework for building whatever kind of network you want.

https://cloud.google.com/tpu/pricing

These AI chips support popular AI frameworks, because that's what everyone in AI is using.

Remember how it worked for IBM compatibles, and how it works for Android? Or how it works for Linux.

What does that have to do with open-sourcing a chip design? A lot of the work in chip designs is one-off effort that's tied to the specific, proprietary libraries for a given process node at a given fab. Unless it's a soft core, the design isn't terribly useful, since it's not like you can just take the design and send it off to another fab or make changes to it without a lot of follow-on work to prepare it for manufacturing.

Again, examples if you got 'em. If not, we're done.

[coder]

Considering you thought Chromebooks (and a Chrome tablet) were relevant to the discussion, no, you really weren't following it.

sigh.

It doesn't fit with your conjecture that what's motivating all of these brands is really a subterfuge by MS to make the market seem bigger than it is. I don't buy it. I think some of those guys also have an eye on the Chromebook market. You don't have to agree, but please don't try to tell me what I was and wasn't following.

Now seems like an apt time to ask whether you picked your username out of awareness of how you can sometimes be perceived.

[coder]

That's why, you see,
At least on ARM cpus, AES, H.264, H.265, VP8, VP9, MPEG2, and so on, implemented in hardware, they are all established standards..

Heh, their implementations are very rarely as hard as you seem to think. Most of these video compression accelerators are fairly programmable, from what I know.

[coder]

Sorry, but comparing copy-left with patents and saying copy-left doesn't foster innovation is BS. A third party can take MIT/GPL licensed code, add to it and do something new that's faster/better/cheaper. With patents, you're waiting on Intel/Qualcomm as the monopoly holders, or hoping to sue or not get sued by Oracle/Facebook. Copy-left has enabled innovation, and things like AV1, XZ, Zstd and even Linux are proof where almost nobody has to worry about patents anymore where they once did.

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).

[starshipeleven]

You were saying that it would make sense for cloud operators to open source their chips

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.

And before you claim that I'm off-topic by talking about AI chips, let me refer you to the fact that Alibaba didn't even describe the use case for this chip as cloud.

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.

You're the one who needs to defend his assertion that this chip makes sense to open source.

Defining what I actually said is an important part of it. I never said "it would make sense for cloud operators to open source their chips". I said that it would make sense for this specific case which is for general-purpose computing.

If that were so, then others building similar chips or chips targeted at similar markets would be doing it.

There are more factors at play than simple economics, I already said that it would be a risk with potential gains, but that's the case also for most opensource software too. Also opensource isn't in the mindset of hardware designing companies.

Tensilica cores in their SoCs ... CEVA is another big licensor

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

More ignorance! Anyone can rent time on Google's TPU cloud instances. .. These AI chips support popular AI frameworks, because that's what everyone in AI is using.

And that does contradict what I said in what way?

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.

A lot of the work in chip designs is one-off effort that's tied to the specific, proprietary libraries for a given process node at a given fab.

Yes it is tied I know.

Unless it's a soft core, the design isn't terribly useful, since it's not like you can just take the design and send it off to another fab or make changes to it without a lot of follow-on work to prepare it for manufacturing.

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.

Again, examples if you got 'em.

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

[schmidtbag]

It doesn't fit with your conjecture that what's motivating all of these brands is really a subterfuge by MS to make the market seem bigger than it is. I don't buy it. I think some of those guys also have an eye on the Chromebook market. You don't have to agree, but please don't try to tell me what I was and wasn't following.

Both Dungeon and I were well aware of Chromebooks before you brought them up, but we didn't mention them because the discussion at hand was about Windows 10 laptops running on ARM, which if you were following you would have understood.
I actually do agree that some people looking into Windows ARM laptops might have an eye on Chromebooks; I think MS is trying to steal some of the Chromebook market (because it's largely un-tapped, and leaves room for improvement). But, the same people shopping for those devices are also keeping an eye on the Macbook Air, iPad, and Android tablets. Chromebooks are not a direct competitor to these Windows ARM laptops because they target people with a different set of needs; different enough that for some, the choice is obvious, and therefore, the markets don't overlap as much as you think. The people who buy the Windows laptops have good reasons to not get a Chromebook. The people who buy a Chromebook have good reasons to not buy a Windows laptop.

[tuxd3v]

Heh, their implementations are very rarely as hard as you seem to think. Most of these video compression accelerators are fairly programmable, from what I know.

Agree, SIMD instructions are usually used in corner cases, were the functionality has less hardware implementation, or even lack it..
Or to glue together some hardware/software parts of it..

[blargh4]

Agree, SIMD instructions are usually used in corner cases, were the functionality has less hardware implementation, or even lack it..
Or to glue together some hardware/software parts of it..

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.