RISC-V Gaining Traction

Except, this isn't really how things have worked in this industry. The primary driver of most innovation has been proprietary technologies and licensed technologies.

Most internet servers (virtually all supercomputers) run on linux, just to mention one counterexample.
 
Last edited:
But an architecture that doesn't specify Vectors in it's default configuration sounds like you're going to needs layers of cruft just to make a serviceable desktop processor :rolleyes:

https://sourceware.org/pipermail/binutils/2020-July/112570.html

It will take design wins in embedded, but once you step-up to desktop-level core complexity, the "free" advantage will be lost over already established ARM...not to mention the standardized Scalable Vector Extensions they defined.

Don't specifying Vectors in the base ISA is the reason why RISC-V is chosen for many embbebbed applications where vector insutrctions are uneeded.

No problem with making a desktop processor with the base ISA more extensions. It is still simpler than making a processor around a bloated ISA as x86.

RISC-V would end up excluding all but the largest tech companies from creating product. There is no RISC-V design team consisting of a couple hundred high level chip designers sitting around all day building RISC-V cores.

The ISA is new and part of it are still under development.

RISV-V good on paper... a controller chip / iot player at most.

You recall those that said me "ARM is only for low power phones". The fastest supercomputer in the world is ARM now. FYI there are lots of interesting high performance projects around RISC-V. Do you know that RISC-V is the default choice for EPI?

https://www.european-processor-initiative.eu/?p=unsubscribe

If say Nvidia decided they where going to take RISC-V and build their own open cores. If they managed to make them worth while to licence (which means they would have to be very custom... its a OPEN source ISA they can't copy write the ISA just potential extensions). Once the software ecosystem was in place. Nothing stops say Google from saying... ok good work NV. Here is our version... who wants it ?

Nvidia has already designed its own RISC-V cores. Nvidia Falcon Next core comes to the mind. They evaluated ARM, MIPS, Tensilica, and others options, and finally they chose RISC-V for their project. Some extra bits

Captura de pantalla de 2020-08-01 11:58:22.png
 
Don't specifying Vectors in the base ISA is the reason why RISC-V is chosen for many embbebbed applications where vector insutrctions are uneeded.

No problem with making a desktop processor with the base ISA more extensions. It is still simpler than making a processor around a bloated ISA as x86.



The ISA is new and part of it are still under development.



You recall those that said me "ARM is only for low power phones". The fastest supercomputer in the world is ARM now. FYI there are lots of interesting high performance projects around RISC-V. Do you know that RISC-V is the default choice for EPI?




Nvidia has already designed its own RISC-V cores. Nvidia Falcon Next core comes to the mind. They evaluated ARM, MIPS, Tensilica, and others options, and finally they chose RISC-V for their project. Some extra bi

One, everyone can do their own custom chips on RISC-V , two yes, correct Juangra, that is two of the first use cases for it, and yes, if everyone makes their own standards it will, mayhbe be a shit fest, but I see this leaching into the x86 as not a replacement, but a power saver for a lot of on board chip things. Could it be great ? Don't know, thinking about procuring one through my company just to know about it.

The arguments around x86 should stop, I personally love it and intel does also, could their stupid pre thing have worked? yea, but Alpha was already better but hp killed it to support that
 
If Nvidia bought ARM, the real result would be both boring datacenter stuff, but also: ~~~New Console Producer~~~!

Xbox vs PlayStation vs Nintendo vs Nvidia.

Would PC still be the top cat?

No, Nvidia has no internal game makers , and know how hard it would be to carve one out. They only launched the original Shield Portable to give themselves something to do with their mostly-abandoned Tegra parts. Then they launched the most advanced Android TV box ever assembled, with the latest Tegra, as a example to win car sales...they just continued the same Steam games ports they did with the original model.

I could still see them getting the design win for the next PS and Xbox, by finally attack that little Scaled Vector Extension monster ARM Inc has been too timid to tackle in their shipping cores.That's really the only thing holding back current ARM cores from being vector-powerhouses for consoles.
 
The ISA is new and part of it are still under development.

You recall those that said me "ARM is only for low power phones". The fastest supercomputer in the world is ARM now. FYI there are lots of interesting high performance projects around RISC-V. Do you know that RISC-V is the default choice for EPI?

Nvidia has already designed its own RISC-V cores. Nvidia Falcon Next core comes to the mind. They evaluated ARM, MIPS, Tensilica, and others options, and finally they chose RISC-V for their project. Some extra bits

View attachment 266296

Falcon is a controller. Its not really a general compute core. Its not accessible by users... or their drivers really. Its a controller chip nothing more. ARM made no sense cause why pay for a licence for a controller chip if it doesn't need to be compatible with anything else, and end users won't have to run code on it ect. The same is true of Western Digital who is using RISC-V for their controller chips. If they needed to run userland software RISC-V makes little sense... and going with something with billions of dollars of software investment makes a licence fee vs hiring software engineers a logical choice.

As to my point on no one is designing cores. I am referring to ARM cortex cores. There is no one, or any team of paid engineers working at the RISC-V foundation or anywhere else designing working high end core designs. ARM has 1000s of people that do nothing all day but design and build working ARM cores. If you wanted to start a ARM desktop startup... you could if you wanted hire 10 people, call up arm and pay for a stock Cortex A75 octo core design with some mali GPU complete with blueprints and everything required to basically send to a fab. At that point you call up Sammsung or TMSC negotiate how many you want to make... provide to them the 7nm plans that direct from ARMs office... and 6months to a year later you would be receiving working product.

If you wanted to do the same thing with RISC-V.... well there is no existing designs. There are no plans to buy. You will have to hire yourself 100-500 people... pay them their 75-200k yearly salaries + benefits to entice them to not go work for ARM/Intel/AMD/Samsung ect ect. Decide what FAB your going to shot for.... cause without a pre designed core you are probably 2-3 years away from a working design to send to a fab company. And you best choose wisely... this will be your first chip... start at 7nm now and you will be releasing a chip while everyone elsse has 3nm parts shipping. Design for 5nm and your likely safe... perhaps the competition has 3nm parts when you launch perhaps not. Shot for 3nm... and cross your fingers that the fab companies don't run into issues at both 5nm and 3nm that delay your ability to produce product at all. So you take your gabble... and I guess you find investors willing to sink multiple billions of dollars on a bet that may or may not pay off... and even if it does pay will have to recoup that 2-3 years of investment with NO income at all.

That is the uphill battle RISC-V faces.... there are companies that COULD take that sort of bet. But they are the Apples, Intels or the world. Even AMD doesn't have the cash to take that bet. On a ISA with really next to ZERO user land infrastructure at this point. Which would also have to be built out for any product to have even half a shot of being a viable consumer product.

ARM didn't get to where it is over night. It got there by licencing to everyone and anyone and having THEM fund development. ARM has 1500 employees that do NOTHING but R&D. They spend around 1 billion US a year on R&D. Without that yes a company like Apple would say thanks for the ISA... and a few product cycles down the road they would be so far ahead of any other ARM licence holder it would be come their ISA. As I said before the advantage of ARM... and one that isn't possible to replicate in a Open source ISA. Is the main office (arm) can afford thanks to licence fees to spend 1/3 of their resources keeping modern off the shelf designs up to date. Sure Apple may still have the fastest chip this cycle or that... or Qualcomm may spin off a core that is a bit better or sips more power. But a year or two down the line all that ARM R&D means the next stock Cortex core design is on par. So the off the shelf companies like say Mediatek may never take the performance crown from Apple or Qualcom who are heavily tweaking those designs. However they can put out good product and for sure = mid range product with a fraction of the investment. RISC-V will require massive investment from anyone wanting to product a performance RISC-V chip... and on going massive investments to stay relevant. Once you get one chip out you have to have the next one well into development already. Its why companies like Intel have 4 or 5 chips at various stages of development at all times. There designing chips for 5 years from now right now... there in late stage development of stuff for 3 years from now... and they have engineering samples of chips for 2 years from now already on a bench somewhere.
 
Last edited:
Don't specifying Vectors in the base ISA is the reason why RISC-V is chosen for many embbebbed applications where vector insutrctions are uneeded.

No problem with making a desktop processor with the base ISA more extensions. It is still simpler than making a processor around a bloated ISA as x86.


Not if you end-up with fifteen different revisions of 3D-Now!. Just because they support a base like FP64 plus some basic vector extensions doesn't mean vendors won't fragment\t that to shit with incompatible vector enhancements.

Because we all know just how many competing Open-Source software components do the exact same BASIC thing, but wildly differ on details. Who is preventing that level of fragmentation from happening in a RISCV-world?

These more advanced cores are going to be used for Linux workstations, so suddenly this could massively multiply the problems open source software already has (a dozen different versions of every little software thing, now ported to a dozen different versions of incompatible hardware revisions).

Reinvention of the Wheel is still a thing. As long as you let folks do witchcraft on the entire architecture, it leaves messy implementations open.
 
Last edited:
Not if you end-up with fifteen different revisions of 3D-Now!. Just because they support a base like FP64 plus some basic vector extensions doesn't mean vendors won't fragment\t that to shit with incompatible vector enhancements.

Because we all know just how many competing Open-Source software components do the exact same BASIC thing, but wildly differ on details. Who is preventing that level of fragmentation from happening in a RISCV-world?

These more advanced cores are going to be used for Linux workstations, so suddenly this could massively multiply the problems open source software already has (a dozen different versions of every little software thing, now ported to a dozen different versions of incompatible hardware revisions).

Reinvention of the Wheel is still a thing.

Well to be fair they do have a RISC-V foundation... and are trying to model the Linux foundation. (as I said in an earlier post though the Linux foundation is only responsible for the kernel... not an OS or software stacks ect ect. I don't think hardware can be treated the same way imo anyway)

So yes they control the base ISA... and some extensions. They can ask all the players to play nice and upstream their extensions. However they don't have to... and companies producing product don't HAVE to include those extensions at all. Or worse they could all start picking and choosing. And doing market segmentation... where standard desktop chips would have none of the high performance extensions. Think Intel AVX-512 only 10x worse. In theory that might be good in a way where chips could be designed for very specific uses... on the other hand, for consumer desktops it would be a complete cluster fuck, where everyone may be able to run the same software... but one chip may be stupid fast on X and Y application, and run like a TI calculator on another. It would be a cluster. Also to be fair Apple may be taking ARM that way anyway... good chance they are going to bolt a bunch of accelerators onto their macbook A14/A15 chips. In their case no one else is providing hardware... but for the PC world to replicate that might be a shit show even with ARM.
 
Well, there is no longer pure CISC chips in the x86 world. Any modern x86 core is a hybrid with a CISC front-end and a RISC engine. x86 instructions are translated into uops, which are stored (uop cache), renamed, reordered, issued, executed, and retired.

Also x86 is in clear decline

View attachment 266295

RISC finally won!

Dude obscure chart from some weird site from 2017. (I found it here )

Please source your stuff!!!

Obviously includes mobile which really isn't an overlapping segment.

So shenanigans.

I've been trying to run an ARM or risc-V desktop / workstation IMO it can't be done reasonably and I doubt that will change anytime soon. I guess some Jetson boards look good on paper, but not nearly good enough in reality.
 
So on an open source design who is responsible for finding / fixing vulnerabilities/flaws? IE - Intel's ongoing nightmare the last couple years. Intel has to address/fix it. But if no one "owns" the design, who would step up and fix stuff? Especially once a product makes it into production phase.
 
Falcon is a controller. Its not really a general compute core. Its not accessible by users... or their drivers really. Its a controller chip nothing more. ARM made no sense cause why pay for a licence for a controller chip if it doesn't need to be compatible with anything else, and end users won't have to run code on it ect.

If you pay attention to the table I posted, you can check that Nvidia chose RISC-V for technical reasons as well: performance, power, memory...

As to my point on no one is designing cores. I am referring to ARM cortex cores. There is no one, or any team of paid engineers working at the RISC-V foundation or anywhere else designing working high end core designs. ARM has 1000s of people that do nothing all day but design and build working ARM cores. If you wanted to start a ARM desktop startup... you could if you wanted hire 10 people, call up arm and pay for a stock Cortex A75 octo core design with some mali GPU complete with blueprints and everything required to basically send to a fab. At that point you call up Sammsung or TMSC negotiate how many you want to make... provide to them the 7nm plans that direct from ARMs office... and 6months to a year later you would be receiving working product.

If you wanted to do the same thing with RISC-V.... well there is no existing designs. There are no plans to buy. You will have to hire yourself 100-500 people... pay them their 75-200k yearly salaries + benefits to entice them to not go work for ARM/Intel/AMD/Samsung ect ect. Decide what FAB your going to shot for.... cause without a pre designed core you are probably 2-3 years away from a working design to send to a fab company. And you best choose wisely... this will be your first chip... start at 7nm now and you will be releasing a chip while everyone elsse has 3nm parts shipping. Design for 5nm and your likely safe... perhaps the competition has 3nm parts when you launch perhaps not. Shot for 3nm... and cross your fingers that the fab companies don't run into issues at both 5nm and 3nm that delay your ability to produce product at all. So you take your gabble... and I guess you find investors willing to sink multiple billions of dollars on a bet that may or may not pay off... and even if it does pay will have to recoup that 2-3 years of investment with NO income at all.

I already answered this. RISC-V just began. The foundation is currently focusing in developing the ISAs, although companies have begun using RISC-V for low power applications. No one is going to design a 4GHz 8-wide OoO core now, when there is no software no anything. As of 2019, RISC-V International freely published the documents defining RISC-V and allowing unrestricted use of the ISA for design of software and hardware.

Said that, some companies are already designing its own OoO cores as replacement for ARM Cortex cores.

Not if you end-up with fifteen different revisions of 3D-Now!. Just because they support a base like FP64 plus some basic vector extensions doesn't mean vendors won't fragment\t that to shit with incompatible vector enhancements.

Because we all know just how many competing Open-Source software components do the exact same BASIC thing, but wildly differ on details. Who is preventing that level of fragmentation from happening in a RISCV-world?

These more advanced cores are going to be used for Linux workstations, so suddenly this could massively multiply the problems open source software already has (a dozen different versions of every little software thing, now ported to a dozen different versions of incompatible hardware revisions).

Reinvention of the Wheel is still a thing. As long as you let folks do witchcraft on the entire architecture, it leaves messy implementations open.

The x86 world of extensions: 3DNow!, ABM, ADX, AES, AMX, AVX, AVX2, AVX-512, BMI1, BMI2, CLMUL, E3DNow!, EMMX, F16C, FMA3, FMA4, FPU, MKTME, MMX, MPX, PMEM, PREFETCH, RdRAND, SEV, SGX, SHA, SME, SMM, SMX, SSE, SSE2, SSE3, SSE4.1, SSE4.2, SSE4a, SSE5, SSSE3, TBM, TME, TSME, TSX, XOP....
 
So on an open source design who is responsible for finding / fixing vulnerabilities/flaws? IE - Intel's ongoing nightmare the last couple years. Intel has to address/fix it. But if no one "owns" the design, who would step up and fix stuff? Especially once a product makes it into production phase.

It depends. A flaw in the base ISA or extensions would be fixed by RISC-V International. A flaw in a custom commercial design would be fixed by the company behind it.

https://www.tomshardware.com/news/risc-v-not-vulnerable-meltdown-spectre-cpu-bugs,36231.html
 
The point is with all the x86 extensions... is Intel and AMD both can implement any of them as they have cross licence access to all and any x86 extensions forever. They agreed to that. No such agreements would exist in the RISCV world. Companies could be good citizens and submit them to RISCV to be official. For the best software support that would be logical. However nothing forces that... and once one player got to a point of market dominance WHICH would happen for sure as core development on a consumer product level time frame of a new core every 2 or so years would ensure only the most well funded players would be able to really lead, they would loose all incentive to continue to play along with the RISCV foundation. Once a Intel or another company of that size got to a 60%+ market share they would make it hard for anyone to catch them by simply adding new extensions and not share. The software market would jump to support the most popular chip... and before you know it we would have one company powering everything again. Relying on multi billion dollar publicly traded tech companies to do what is right is laughable.

As for NVs new RISCV falcon controller. Yes its RISCV based... but lets not kid ourselves. It is a 0.25 watt design that routes traffic from main memory to GPU memory and divides work loads between CCX style GPU core arrangements. Its not a processor its a controller chip. ARM didn't make sense cause for what it does even a IOT type ARM licence was money they didn't need to pay.... and having a super simple ISA to work with made sense. It would be like AMD saying the controller chip on their Ryzen 2s was a new ISA of their own design and naming it. lol They do the same job... AMD just made their own ISA for that bit of hardware.

I agree its a start... and I support RISCV development. There is a need for a free ISA people like NV and western digital can use for inexpensive relatively high power efficient hardware controllers. I am just suggesting people look at RISCV with a practical eye. We really don't want the industry to try and make it a general compute replacement for x86 or ARM. It would be very bad for consumers long term. x86 frankly was bad for consumers as well. The initial explosion of inter changeable designs was good... and then what happened was the same thing that would happen to RISCV One player would gain market dominance... and go to work making everyone else look like a knock off. Though marketing, back room tricks... and extensions and additions that make their market dominate solution the one software developers optimize for and ultimately the only chip that matters. If AMD and Cyrix hadn't beaten Intel in court a few times (or got Intel to a point where they figured they may or have the Gov step in and break them up) x86 would have one player. I mean AMD almost did go out ... and we are really only a few Intel screw ups the last 5 or 6 years away from only having one x86 player.

The best solution for consumers imo is for ARM to issue its IPO in the next year or two as planned.... and for it to hopefully grow to a point where its purchase by a single tech company becomes very implausible.
 
The x86 world of extensions: 3DNow!, ABM, ADX, AES, AMX, AVX, AVX2, AVX-512, BMI1, BMI2, CLMUL, E3DNow!, EMMX, F16C, FMA3, FMA4, FPU, MKTME, MMX, MPX, PMEM, PREFETCH, RdRAND, SEV, SGX, SHA, SME, SMM, SMX, SSE, SSE2, SSE3, SSE4.1, SSE4.2, SSE4a, SSE5, SSSE3, TBM, TME, TSME, TSX, XOP....

And that's my whole fucking point: if you think the mess of abandoned or proprietary extensions is bad on x85, prepare yourself for the mess that RISCv will bring.

That's only from THREE x86 COMPANIES. When you get dozen of major players, every little tweak is going to be another incompatible standard


ARM is the opposite, as it's been controlled by one company...that's the way to handle your architectural design.
 
That's only from THREE x86 COMPANIES. When you get dozen of major players, every little tweak is going to be another incompatible standard

Let's imagine that both Microsoft and Google add support for RISC-V to Windows and Android (and sure, ChromeOS, too, if you want.) Do you think that either of them will bother using the single-vendor extensions in the OS? Of course they won't. So the extensions won't matter.

But Riscy-Visuness may release a driver that supports their special snowflake extension under Windows, or maybe their CPUs will only ever be used in a line of IoT products. or something.
 
Let's imagine that both Microsoft and Google add support for RISC-V to Windows and Android (and sure, ChromeOS, too, if you want.) Do you think that either of them will bother using the single-vendor extensions in the OS? Of course they won't. So the extensions won't matter.

But Riscy-Visuness may release a driver that supports their special snowflake extension under Windows, or maybe their CPUs will only ever be used in a line of IoT products. or something.

So windows doesn't take advantage of MMX or SSE or any of the rest of the extensions Intel has added ?

Of course they will use the extensions that make sense when the market leader is pushing them. That is the trap of RISC-V. Sure they may all play nice for the first few years... but you can't trust any company gets a super majority of the market. AND that will happen.
Within a few years a market leader will emerge... and at some point as every other chip company ever has done they will take over on the spec. They will fire up marketing claiming they are the one true wearer of the ring. Once they have the ring of power they will use it to destroy the rest. Intel did exactly that with x86.... and all a open ISA does is make it even easier. If we go RISC-V and in 10 years X company has 80% of the market.... who stops them form introducing RISCV-MMX-ULTRA. No one. Who else makes a CPU using their extension... again no one. Cause they have no legal right to that IP. Does the software world ignore 80% of the market of course not. Then there you have it the other 20% are no longer the official RISCV standard.

The OS doesn't really have to use something... anyway its the software developers that decide what they want to use. We are going to get a taste of that in less then a year ... when Apple puts out a $800 mac book with a A14x in it. (according to rumors) Chances are its going to have AI bits bolted on to accelerate things like Photoshop, and Adobe and black magic and all the rest will code the bits to make it do its thing. Be ready.... we are going to be seeing legit benchmarks of low ball Macbooks destroying 10 thousand dollar x86 workstations in terms of things like Photoshop filters, and real time video effects. That will be but a taste of what would happen if RISC-V was to replace x86.... as it is we are probably heading there in the PC world anyway with ARM. Intel is going to try to take x86 there as well with big.Little chips with potential accelerators on the same SOC.

We are heading for a world of very software specific hardware... the worst thing that could happen in that world... is a open ISA, that excludes all but players with the bank roll to do full chip design. This is why frankly we need to go ARM... by using ARM the basic ARM designed cortex parts will be just fine. The interesting bits will be the accelerators tacked on. Whats cool about that is a very small team at a fairly small design house can build very cool accelerators and just slap them on licenced Cortex core designs. Intel knows this is the future which is why they have been working on their Foveros 3D chip stacking tech.... they are trying to get those folks to contract Intel instead of using ARM and doing it themselves. Having said that with everything going on at Intel who knows if that is still on the table. But they where spending a lot of money on it... their plan is to build a big.little chip with the same flexibility to add accelerators.

So ARMs plan is...
Small company... designs kick ass AI accelerator. Designs it to take up 10% of a 7nm die. Then licences a octo core Cortex design and a Mali GPU. Spins it up and sells it.
Intels counter...
Small company... designs kick ass AI accelerator. Designs it to take up X amount of space on a foveros package. Then Intel manufacturers them a x86 big.Little design with their accelerator built in.

IMO Arms plan is far superior. Intels fabs can't keep up with Intel demand. ARM has ready made blueprints for every fab you could licence production from. It gives small companies actual choice.. instead of locking them to being a Intel coat tail company.
 
Let's imagine that both Microsoft and Google add support for RISC-V to Windows and Android (and sure, ChromeOS, too, if you want.) Do you think that either of them will bother using the single-vendor extensions in the OS? Of course they won't. So the extensions won't matter.

But Riscy-Visuness may release a driver that supports their special snowflake extension under Windows, or maybe their CPUs will only ever be used in a line of IoT products. or something.


They're not going to not support advanced vector extensions, or the software developers wouldn't bother porting their software over :rolleyes: Just because embedded processor designs don' t need SIMD doesn't mean bigger platforms will not.

Also, there's no way in hell that paid-for software platforms will ever support the dozen-plus vector standards (which is why Windows quickly standardized on SSE). Even Microsoft put their foot own, and wouldn't support another Intel proprietary 64-bit extension for x86, after IA64 flopped.

But they do not delay in supporting unified standards:

https://www.drdobbs.com/parallel/windows-7-and-windows-server-2008-r2-ser/229300330
 
Well, there is no longer pure CISC chips in the x86 world. Any modern x86 core is a hybrid with a CISC front-end and a RISC engine. x86 instructions are translated into uops, which are stored (uop cache), renamed, reordered, issued, executed, and retired.

Also x86 is in clear decline

View attachment 266295

RISC finally won!



The "CISC baggage" matters for larger chips as well. It is the reason why Intel and AMD have included a uop cache in Sandy and Zen.

The x86 tax has two parts: a fixed and a variable. The fixed part affects more to small cores. The x86 tax not only affected Intel mobile desings, it is also part of the fiasco behing Larrabe/Phi or stuff as Quark.


So essentially a hardware x86 emulator sitting on top of a RISC processor :p
 
I think the important thing to take from all of this is that RISC-V gives management technocrats a new term they can use to make themselves look tech savvy.

It used to be that they would talk about plain old RISC, which was a technology that was heavily backed by IBM, but eventually everyone caught on to the fact that RISC (while certainly capable) was JAIS (Just Another Instruction Set), and RISC fell out of favor. So the technoflabs switched to talking about XML, and everyone was happy with that, especially since no one knew what that was either. If you invested a lot of time in learning the core of XML you might realize XML is the emperor's new clothes, it's just a data-transport markup language and a stupid one at that. But that knowledge allowed you to throw softball questions at the technolizards, which is fun (or, if you're like me, every time someone brings up XML you could make shitty comments like, "XML is great for teaching elementary grade children about data, but otherwise it's shit.")

But now we have RISC-V. I can smell it now. We'll be talking about mainframe-class data server purchases, and some boob who couldn't manage a yarn store will ask, "Do we know which companies are exploring RISC-V architecture?"
 
Softbank is hurting, yet another loss write off this one like 20 Billion from their tech holdings. SoftBank is looking to offload things to make money back and they are seeing who’s biting. NVidia May be the poster boy on this but I would bet Huawei would be interested, they can’t get any more Arm chips and they have started working on their own RISC-V stuff, they are touting it as the best thing since the communist party invented sliced bread. I’m less worried about NVidia buying it, they would have so many regulators up their behind that it would force them honest out of the vast teams of lawyers that would be watching their every bathroom break. A China owned one would just be bad news.

There's less than zero chance that Huawei, or any other Chinese company would be allowed to buy ARM.
 
There's less than zero chance that Huawei, or any other Chinese company would be allowed to buy ARM.
Legally it’s already been displayed that Arm is a European company and not an American one. This was done in 2019 so they could continue to license their designs to China. Europe as a whole doesn’t have the same public beef with China that the US does, European regulators may allow it. It’s not likely but I would not say zero, Huawei is going big on RISC-V now as a result so they and many other Chinese companies are likely to go that way just to avoid US entanglements. So it probably won’t happen for that reason alone they are likely more than happy to take the tech they need from ARM to incorporate into RISC-V and sell that in China and call it a win.
 
They're not going to not support advanced vector extensions, or the software developers wouldn't bother porting their software over :rolleyes: Just because embedded processor designs don' t need SIMD doesn't mean bigger platforms will not.

Also, there's no way in hell that paid-for software platforms will ever support the dozen-plus vector standards (which is why Windows quickly standardized on SSE). Even Microsoft put their foot own, and wouldn't support another Intel proprietary 64-bit extension for x86, after IA64 flopped.

But they do not delay in supporting unified standards:

https://www.drdobbs.com/parallel/windows-7-and-windows-server-2008-r2-ser/229300330

Well, that's my point. Further, using x86 and stuff like 3dnow isn't a good comparison because Intel owned x86, and wasn't just one of a dozen companies making the chips.
 
ARM didn't get to where it is over night. It got there by licencing to everyone and anyone and having THEM fund development. ARM has 1500 employees that do NOTHING but R&D. They spend around 1 billion US a year on R&D. Without that yes a company like Apple would say thanks for the ISA... and a few product cycles down the road they would be so far ahead of any other ARM licence holder it would be come their ISA. As I said before the advantage of ARM... and one that isn't possible to replicate in a Open source ISA. Is the main office (arm) can afford thanks to licence fees to spend 1/3 of their resources keeping modern off the shelf designs up to date. Sure Apple may still have the fastest chip this cycle or that... or Qualcomm may spin off a core that is a bit better or sips more power. But a year or two down the line all that ARM R&D means the next stock Cortex core design is on par. So the off the shelf companies like say Mediatek may never take the performance crown from Apple or Qualcom who are heavily tweaking those designs. However they can put out good product and for sure = mid range product with a fraction of the investment. RISC-V will require massive investment from anyone wanting to product a performance RISC-V chip... and on going massive investments to stay relevant. Once you get one chip out you have to have the next one well into development already. Its why companies like Intel have 4 or 5 chips at various stages of development at all times. There designing chips for 5 years from now right now... there in late stage development of stuff for 3 years from now... and they have engineering samples of chips for 2 years from now already on a bench somewhere.

You can have a similar model with an open source ISA.

The RISC-V ISA is being developed by volunteers and universities, so there are no costs associated with it that a license would be needed to pay for.

Just because the ISA is open source, doesn't mean that a player like ARM Holdings couldn't develop and license their own architecture based on that ISA, like ARM does with their architectural and "built on ARM Coretex" licenses.

No one is saying the entire chip has to be open source. Just the ISA itself, so that no one can be excluded from developing their own chip on that ISA, should they want to.

Agreed that RISC-V faces an uphill battle. It is very difficult to get critical mass of both manufacturers of chips and users, but if an industry where anyone who wants to make a chip can use a common ISA without anyone trying to stop them could result in a good competitive market place with many players across different architectural niche's. It also would have the benefit of vendors using the CPU products not having to deal with being painted into a corner where they only have one option. If your entire software ecosystem is built ontop of RISC-V, and the chip vendor making your CPU decides they don't want to sell it to you anymore for whatever reason, there could be another vendor also making a RISC-V chip you could use without breaking software compatibility.

It would hugely de-risk (pun intended) the system integrator role, and would likely be very appealing to them.

This is almost what ARM has been like for the last 20 years, and in large part why it has been so popular. ARM has been neutral, and sold licenses to anyone who wants them (as long as it has been legal). All of a sudden, the news that Nvidia is in advanced talks to buy ARM Holdings has to have a lot of these system integrator types rather concerned. It could upend their entire businesses. With an open source ISA this would not be the case. If it is open source, the ISA will never become off limits. You could very well lose a a chip architecture vendor, but then you can just go to another (presuming it is successful and takes off)

As it stands right now, proprietary instruction sets are just a huge "lock-in" feature that can be used to harm customers.

It is unclear to me what parts of ARM Holdings Nvidia are really interested in, but because of the above competition concerns it is not unlikely the deal if it becomes final will face regulatory scrutiny. If they do something like spin off the ISA portion of ARM Holdings to a neutral and independent entity, and buy everything else, that could go a long way to calm a lot of nerves.
 
Well, that's my point. Further, using x86 and stuff like 3dnow isn't a good comparison because Intel owned x86, and wasn't just one of a dozen companies making the chips.

Intel was just one of a dozen companies making x86 designs.

Currently their are 3 x86 manufacturers. Intel, AMD, and VIA (Zhaoxin).
In the past there have been chips from... Transmeta, Cyrix, Nexgen, C&T, IBM, NEC, UMC, Centaur, and RISE. Texas Interments and SGS also manufactured and sold chips with their names on them designed by Cyrix for awhile. There is also a small handful of companies that made embedded x86 chips, as well as at least 2 that made x86 SOC.

What happned to x86 the last 50 or so years... is Intel licensed production to increase market share, then they licensed design to stave off all competition. Then when they got a dominant market share position. They sued or bought out anyone they sold a licence to... and those that they lost to or couldn't buy out. They spent 100s of millions on marketing to destroy. Making consumers see them as knock offs. Pentium was the most obvious example of this. Where Intel trademarked its new name breaking from the x86 naming convention. So a Cryix 586 was clearly not a pentium class chip. They also extended the ISA... with extensions they did not include in their original licence deals.

Yes open ISA would lead to all the same tricks. Extensions are one of the most obnoxious ways to knock off your competition when your in a dominant market position. Of course 3d Now went no where it was created by the #2 manufacturer and the #1 didn't support it.

EDIT... and perhaps the real issue in the future more the extensions will be co processors. ARM doesn't solve that one either. Chip tech now allows any one to slap extra bits outside the core on the package. So open ISA or licenced it might not matter much to be fair. Nothing stopping Apple or Samsung from adding AI or vector co processors unique to their parts.
 
Last edited:
Yes open ISA would lead to all the same tricks. Extensions are one of the most obnoxious ways to knock off your competition when your in a dominant market position. Of course 3d Now went no where it was created by the #2 manufacturer and the #1 didn't support it.

I don't know what open source license RISC-V is made available under, but if they use something like the GPL license where derivative work must be redistributed freely under the same license, part of that issue goes away. The RISC-V developers are working on their own extensions, as I posted earlier, which are also distributed freely, and if they keep up with it, there will be no outstanding need for additional extensions, because there are already ones provided as part of RISC-V.

riscv-base-extensions.png


The point here is to create a n open ISA so no big player can do what Intel did, using licensing tricks to force competitors out of the marketplace.

You could have many competing proprietary and licensed architectures utilizing the same open source ISA.

The trick is just to get the snowball rolling with users and manufacturers, and eventually it will be the ideal solution. It looks like right now it is being embraced for embedded designs. Thats a first step. Hopefully it will continue.
 
  • Like
Reactions: ChadD
like this
You can have a similar model with an open source ISA.

The RISC-V ISA is being developed by volunteers and universities, so there are no costs associated with it that a license would be needed to pay for.

Just because the ISA is open source, doesn't mean that a player like ARM Holdings couldn't develop and license their own architecture based on that ISA, like ARM does with their architectural and "built on ARM Coretex" licenses.

No one is saying the entire chip has to be open source. Just the ISA itself, so that no one can be excluded from developing their own chip on that ISA, should they want to.

Agreed that RISC-V faces an uphill battle. It is very difficult to get critical mass of both manufacturers of chips and users, but if an industry where anyone who wants to make a chip can use a common ISA without anyone trying to stop them could result in a good competitive market place with many players across different architectural niche's. It also would have the benefit of vendors using the CPU products not having to deal with being painted into a corner where they only have one option. If your entire software ecosystem is built ontop of RISC-V, and the chip vendor making your CPU decides they don't want to sell it to you anymore for whatever reason, there could be another vendor also making a RISC-V chip you could use without breaking software compatibility.

It would hugely de-risk (pun intended) the system integrator role, and would likely be very appealing to them.

This is almost what ARM has been like for the last 20 years, and in large part why it has been so popular. ARM has been neutral, and sold licenses to anyone who wants them (as long as it has been legal). All of a sudden, the news that Nvidia is in advanced talks to buy ARM Holdings has to have a lot of these system integrator types rather concerned. It could upend their entire businesses. With an open source ISA this would not be the case. If it is open source, the ISA will never become off limits. You could very well lose a a chip architecture vendor, but then you can just go to another (presuming it is successful and takes off)

As it stands right now, proprietary instruction sets are just a huge "lock-in" feature that can be used to harm customers.

It is unclear to me what parts of ARM Holdings Nvidia are really interested in, but because of the above competition concerns it is not unlikely the deal if it becomes final will face regulatory scrutiny. If they do something like spin off the ISA portion of ARM Holdings to a neutral and independent entity, and buy everything else, that could go a long way to calm a lot of nerves.

I can understand the concern about ARM being sold to a company like NV... that could end their in house core designs. That would be terrible for everyone I agree completely. That is exactly however why I can't imagine NV would win regulatory approval even if they could swing a deal for ARM. There is also the issue of price. There is no way the early rumors that NV was looking at 35 billion are right. Bloomberg estimates Nvidia would have to pay more like 55 billion to swing a deal. NV does not have that sort of cash... in fact they have exactly 16 billion in cash right now. The rest would have to come from a massive stock deal... and if the rumors are to be believed Softbank is looking for cash not more ownership of tech companies. :)

My main issue with RISC-V replacing x86 or ARM.. is as you say the only people working on common cores for everyone are university students. I'm sorry but no university project in the world nor all of them combined can turn out a top performing chip of the day every year forever. ARM has been doing that with cortex. t costs ARM a billion a year and 1,500 full time chip engineers to do it. (I think they have just under 7000 employees in total). For ARM to continue to keep up to Apple when they start designing Desktop Mac Pro chips... is probably going to cost them 1.5-2 billion a year and 2-2.5k employees. Conjecture on my part.. but with Apple and Google (rumored) to be looking at powering laptops and desktops with ARM. ARM is going to have to spend more to keep up on all fronts. (although they do have multiple cortex cores... you don't really find the highest end out of order cortex cores in mobile parts)

Anyway ya I really hope the NV stuff is just a rumor, and even if they are looking into it I can't imagine they blow all their cash on hand and convince softbank to accept 30 billion in NV stock... what is the point of that trade for them really. We all know Nvidia would do nothing good for ARM. We all just need to cross our fingers and hope softbank sticks to their original plan for ARM... which is a IPO in the next year or two. IMO they would be insane to sell to NV right now even if NV could come up with 50+ billion. ARM is going to grow in value, and if it goes properly public probably by a lot. Softbank is probably looking at 70-100 billion in profit if they take ARM public again in 2022/23.
 
The trick is just to get the snowball rolling with users and manufacturers, and eventually it will be the ideal solution. It looks like right now it is being embraced for embedded designs. Thats a first step. Hopefully it will continue.

Agreed. I do want to see how far it can go. I think people just need to tap the breaks a bit if they honestly believe in 5 years we will be using RISC-V desktops or even phones. :)
 
The point is with all the x86 extensions... is Intel and AMD both can implement any of them as they have cross licence access to all and any x86 extensions forever. They agreed to that.

Currently some extensions are supported only by AMD hardware and other supported only by Intel hardware. There are even extensions only suported by some specific hardweare (e.g. all the extensions exclusive to the Xeon Phi line, now discontinued).

ARM is the opposite, as it's been controlled by one company...that's the way to handle your architectural design.

And that is the reason for the rapid expansion of RISC-V. Whereas ARM pursues the one-fits-all philosophy, RISC-V allows customers to optimize for the problem thanks to a modular design.
 
Currently some extensions are supported only by AMD hardware and other supported only by Intel hardware. There are even extensions only suported by some specific hardweare (e.g. all the extensions exclusive to the Xeon Phi line, now discontinued).

And that is the reason for the rapid expansion of RISC-V. Whereas ARM pursues the one-fits-all philosophy, RISC-V allows customers to optimize for the problem thanks to a modular design.

The differences between AMD and Intel are choices however. They both have 100% right to use any extension either of them includes in a x86 chip. AMD may not have AVX512 but that was a choice... once that even Intel makes as only a few of their products have it.

I understand that RISC-V allows its users to do more custom stuff with it... that is sort of the problem. lol ARM does allow companies to licence just the core ARM 8 ISA and build their own cores with whatever else in them they like...they simply bring the huge consumer advantage of having ready designed COMPETITIVE cores ready to licence by anyone. RISC-V may allow a Apple or a Samsung or a Western Digital to do whatever they like with the ISA... but that is no benefit to you as a consumer. At some point someone comes along and designs the superior chip, makes a ton of money cause everyone buys it... then most importantly there product is the ecosystem. So when they say this is canon and this is now RISC-V for desktop... it is. When they sue their competition out of re creating that canon they are are now future Intel. WTH is the point of that... from a consumer perspective. At least with x86 there are enough portfolios of old x86 licences... and the open we cross own each others additions deals forever that both AMD and VIA own that there is at least SOME competition. ARM is better of course cause anyone can be competition with min investment... as long as ARM stays independent and up to date with their cortex designs.

The only thing that would make ARM better... would be if they go public.... and every major chip player in the industry bought a good chunk of stock. Ideally Apple, Samsung, Intel, AMD all would be invested to the tune or 5% or so each with no one owning controlling shares. lol
 
The differences between AMD and Intel are choices however. They both have 100% right to use any extension either of them includes in a x86 chip. AMD may not have AVX512 but that was a choice... once that even Intel makes as only a few of their products have it.

It is not only AVX512.

FM4, TSX, XOP, LWP, AES, TBM, MPX, TME, SME, TSME, MKTME, AMX,... are some extensions only supported by AMD or by Intel.
 
RISC finally won!

Nah, modern "RISC" chips aren't RISC at all, just like modern CISC chips aren't really CISC. They are all hybrids. RISC is about reduced number and complexity of instructions. All instructions do only one thing, and you have as few instructions as possible. Makes it nice and simple to write a compiler to optimize for, among other benefits. However here's the thing: That doesn't work to get high performance. We can't just increase clock speed, add more pipelines, do more in parallel. We have hit limits on all those things... So we have to have more complex instructions that do more things, and ones that are specialized to accelerate certain common tasks. Things like vector units, encryption instructions, network instructions, etc. None of this is RISC. Having a chip with DSP instructions, SIMD instructions, Vector floating point instructions, AES instructions, etc as modern high speed ARM chips do is not following the strict RISC model. Instruction complexity and execution time vary, and many of them do far more than a single task.

Basically the best parts of different designs just kind of fuse together on modern CPUs. There isn't this strict differentiation.
 
It is not only AVX512.

FM4, TSX, XOP, LWP, AES, TBM, MPX, TME, SME, TSME, MKTME, AMX,... are some extensions only supported by AMD or by Intel.

By choice. They both have the legal right to use any of it. If any of those became popular they would find their way into both companies chips... and potentially chips made by those holding Via patents as well.
 
When it takes 145 pages just to describe the RISC V instruction set, I can't call that reduced...

By way of comparison, it took only 60 pages to descrive ARMv7.
 
Last edited:
When it takes 145 pages just to describe the RISC V instruction set, I can't call that reduced...

By way of comparison, it took only 60 pages to descrive ARMv7.


And that ARM spec also includes NEON, an SSE4 competitor.

The fact that RISCV hasn't fully defined this Tells us just how freewheeling the implementations are going to be. You think the vector spec battle between AMD and Intel was bad - this is going to be a Civil War full of hardheaded competitors.
 
Nah, modern "RISC" chips aren't RISC at all, just like modern CISC chips aren't really CISC. They are all hybrids. RISC is about reduced number and complexity of instructions. All instructions do only one thing, and you have as few instructions as possible. Makes it nice and simple to write a compiler to optimize for, among other benefits. However here's the thing: That doesn't work to get high performance. We can't just increase clock speed, add more pipelines, do more in parallel. We have hit limits on all those things... So we have to have more complex instructions that do more things, and ones that are specialized to accelerate certain common tasks. Things like vector units, encryption instructions, network instructions, etc. None of this is RISC. Having a chip with DSP instructions, SIMD instructions, Vector floating point instructions, AES instructions, etc as modern high speed ARM chips do is not following the strict RISC model. Instruction complexity and execution time vary, and many of them do far more than a single task.

Basically the best parts of different designs just kind of fuse together on modern CPUs. There isn't this strict differentiation.

A modern high performance server-class ARM core produces an average of 1.1 uops per instruction. So it is 90% RISC.
 
A modern high performance server-class ARM core produces an average of 1.1 uops per instruction. So it is 90% RISC.

That's not what makes something RISC. Like I said, the original design was about simplicity. A small number of instructions that each had a single function. Neat idea back in the day, but has serious limitations now. Hence we see lines blurred.
 
A modern high performance server-class ARM core produces an average of 1.1 uops per instruction. So it is 90% RISC.
Hey do you have a source for this? I would be interested to read it. I wonder if there is a comparison to x86 for the same software compiled to x86.
 
Back
Top