Question Why don't we have a massive push for ARM chips in desktop computers and laptops?

[ibex333]

Note: The following may sound like idiocy, and I admit, I have no clue what the hell I am talking about, but I think its an interesting discussion nevertheless...


Can someone please explain to a complete noob, in layman terms, WHY cant we start using ARM chips instead of Intel? They sure seem a hell of a lot faster and better than Atoms at least! And in many cases faster than Celerons.

Look, I get it, ARM does not support x86 apps, and forget about 64bit... The instruction sets are different. Yadda, yaddaa, yadda. I heard it all before. But they ported Half Life to Android! They emulated a Playstation 3 on PC! They even got Windows 95 to run on gaming handhelds and managed to run Linux on Android phones! Hell, even iOS can run Android now. An unheard of thing five years ago.

So whats the matter? MAKE x86 apps run naively on ARM CPUs! MAKE the same instructions sets for all CPUs regardless of devices they are used for... Just make it happen. And also, why is it, that ARM CPUs were not designed right from the start, with an ability to run x86 apps?

I am aware, that at least one manufacturer, Hewlett Packard, released a laptop that runs on a Snapdragon CPU. It basically "emulates" x86 apps, and I believe the platform is called "Windows on ARM" or something like that. Because code is not ran naively, but emulated, performance is abysmal, but battery life is fantastic. Whats really tempting to me, is that many times, the graphical capabilities of some high end ARM chips seem to far surpass what Intel offers. The gaming potential is immense. But because smartphones don't have full size keyboards and mice attached to them, and do not have PC like games on them, that potential is very limited. But it does not have to be, if we had something like a giant, dedicated tablet, with a ton of usb ports for peripherals and a huge screen. Something like this could potentially be much better than any console and better than many low end gaming PCs.

Am I missing something here? Am I talking complete nonsense? Lets discuss.

 

[ElFenix]

MAKE x86 apps run naively on ARM CPUs! MAKE the same instructions sets for all CPUs regardless of devices they are used for... Just make it happen. And also, why is it, that ARM CPUs were not designed right from the start, with an ability to run x86 apps?

i mean, why not put an X86 decoder on a RISC type CPU and run that? has anyone even thought of that before?!? i doubt it!

 

[ThatBuzzkiller]

That doesn't mean that ARM vendors are opposed to SVE2. There has been no public opposition to the standard. I don't think you're carefully considering what it DOES mean. SVE2 certainly isn't a "zombie standard".

Most of the ARM vendors use standard designs from ARM holdings. Even Samsung has ended their custom line of ARM SoCs. Qualcomm, Samsung, Rockchip, Huawei (in mobile), and Amazon (Graviton/Graviton2) use or will use reference ARM designs. Not sure about Ampere, but they may be using reference ARM cores as well. SVE2 will be implemented when ARM Holdings finally rolls designs that implement it. Neither A76 nor A77 feature any variant of SVE, so obviously, you aren't seeing it today. The last word I heard from anyone is that SVE2 would be standard for reference ARMv9 designs that should show up in silicon in 2022. Custom ARM vendors are free to implement it now. They just haven't. Fujitsu has implemented SVE in A64FX.

So long as a significant majority of ARM vendors continue to copy reference designs, SVE2 should show up in consumer products in 2022.

I see no indication that either SVE or let alone SVE2 will be standardized. It's been over 3 years since SVE has been introduced and the first system to incorporate the feature will be available in 2021!

At large the industry standard for SIMD functionality will remain as NEON for ARM and AVX/AVX2 for x86 systems for the forseeable future and even x86 might converge to AVX-512 before we see any of ARM's biggest customers integrating SVE into their product portfolio.

If ARM's most popular designs still haven't made SVE/SVE2 standard by 2022 then it'll be a major political failure and conditions will be worse than previously thought.

Heck, x86 vendors will beat ARM vendors to the punch with even more advanced features by standardizing heterogeneous compute in the far future. What AMD is doing with ROCm or Intel with their oneAPI is primitive compared to what they'll introduce 5 years down the line. CUDA was groundbreaking for GPGPU however it wasn't what I'd initially describe to be suitable for heterogeneous compute but ROCm or oneAPI in the future could become the models for heterogeneous compute for many years to come. Out of all the ARM vendors only Nvidia possesses the political capacity to compete in heterogeneous compute while both x86 vendors are well on their way to developing a unified heterogeneous programming solution.

 

[GodisanAtheist]

People have provided some excellent discussion and answers in this thread, so here is my contribution:

Things that can happen don't for a combination of three reasons:

People are lazy
People are cheap
People fear change

Overcome those hurdles and anything is possible.

 

[NostaSeronx]

The upcoming Cortex-A7X, Cortex-A5X, and Cortex-A3X will have SVE2 upgrade options.

Qualcomm has been confirmed for their next CPU version to support SVE2. Samsung's next CPU design will support SVE2. Huawei/HiSilicon's N5P might also be supporting SVE2, while the N5 version will have it disabled.

2020/2021 is the year of SVE2. With the later Cortex-A7Y, Cortex-A5Y, and Cortex-A3Y having SVE2 being mandatory.

 

[DrMrLordX]

I see no indication that either SVE or let alone SVE2 will be standardized. It's been over 3 years since SVE has been introduced and the first system to incorporate the feature will be available in 2021!

SVE is pretty niche. It's mostly aimed at HPC. ARM Holdings isn't including it in their own reference designs for that reason. NEON is more-suitable for consumer applications. Since SVE2 basically covers both arenas quite well, it made perfect sense for them to effectively mothball SVE and work towards integrating SVE2 into their reference designs instead.

You must remember that SVE/SVE2 aren't being pushed on anyone. ARM Holdings introduced the instruction sets. They provide resources to teach you how to code for those standards. They have an emulator to let you test code for functionality. Eventually, when they feel like it, they'll start licensing reference designs that utilize SVE2. There's every indication that it is already standardized. ARM Holdings just has to decide when it's in their own best interest to start including it in designs.

 

[ThatBuzzkiller]

You must remember that SVE/SVE2 aren't being pushed on anyone. ARM Holdings introduced the instruction sets. They provide resources to teach you how to code for those standards. They have an emulator to let you test code for functionality. Eventually, when they feel like it, they'll start licensing reference designs that utilize SVE2. There's every indication that it is already standardized. ARM Holdings just has to decide when it's in their own best interest to start including it in designs.

I would not argue that an available extension is an indication that it is industry standard functionality. 3DNow! once existed widely on past AMD CPU architectures but today it's relegated as legacy functionality.

Only when the corporations decide to actually converge on technical grounds should a new feature become the industry standard. With SVE or SVE2, ARM vendors are in a prisoner's dilemma ...

The dominant strategy is for them to never adopt SVE/SVE2 because currently they lose out to other competing vendors who also decide to not adopt it. If some ARM vendors do adopt SVE/SVE2 but the others don't then those who didn't will win out by paying lower licensing costs and have less implementation complexity compared to those that did. Especially since there are more players in the ARM processor market it's a far higher likelihood of other ARM vendors cheating each other out.

Currently, most of the ARM vendors are only interested in appealing to the same high-profit margin customers as Apple does so if they ever want to gain the upper hand against Apple in processor design then they'll rightly forget about the idea of implementing SVE2 in the first place.

ARM vendors will never be able to replace x86 in desktops because they are not up to the task of politically cooperating with each other to do this. All I ever see from the many ARM vendors is them trying to cheat each other out in some way which practically kills all sorts of developing technical standards compared to what we see from x86 vendors.

 

[DrMrLordX]

I would not argue that an available extension is an indication that it is industry standard functionality. 3DNow! once existed widely on past AMD CPU architectures but today it's relegated as legacy functionality.

3DNow! was an ISA extension pioneered by a still-minority x86 CPU vendor to cover up for a non-pipelined FPU. AMD in no way, shape, or form dictated future standards for the x86 ISA and its extensions. Definitely not until they introduced x86-64. And that was a major coup for them.

Only when the corporations decide to actually converge on technical grounds should a new feature become the industry standard.

Welcome to the ARM world. ARM Holdings dictates what will or won't go into future ARM standards. Anyone not liking their decisions can use a design license (ala Apple). Anyone willing to throw in with ARM Holdings can just use off-the shelf cores, which is what many of the big players are still doing. Most of the design shops that were producing custom ARM cores have stopped. Apple is one of the few holdouts.

I guarantee you, once ARM Holdings includes SVE2 in one of their designs, the major players will use it. Samsung, Qualcomm, and (in their consumer products) Huawei will all be on board.

The dominant strategy is for them to never adopt SVE/SVE2 because currently they lose out to other competing vendors who also decide to not adopt it.

What EXACTLY are the competing vendors going to use if not SVE2-capable designs? Do you think they're going to go back to custom cores just to avoid it? Why would they? NEON and SVE2 will co-exist in future designs. If you support one, you will support the other. There's no reason to deliberately try to strip out support from SVE2 from future ARM designs.

Do you think they're going to come up with their own ISA extensions? Apple has sort-of gone this way (they have proprietary logic blocks for things like AI). Practically nobody else is doing this. Even Marvell seems to not be taking that approach. They seem to be using fairly-standard ARM designs, albeit with SMT4. What Huawei is doing is unknown since they haven't said much about the differences between the v110 and A76. Never mind that Huawei is still using A76 for the Kirin 990.

 

[beginner99]

I see QNAP sells NAS units with Ryzen 1600's. I bet those are sweet.

And I was pretty sure these cost an arm and leg....so I checked- Indeed the cheapest one (4 drives +2ssd bay) costs you >$1500 with only 8GB of RAM. Even these consumer units are extremely prices for the hardware. You pay for the software and the compact form factor.

I just use my pc as NAS. I mean windows can easily be used as media server or what I prefer trivial file shares.

 

[lamedude]

i mean, why not put an X86 decoder on a RISC type CPU and run that? has anyone even thought of that before?!? i doubt it!

PowerPC 600 - Wikipedia

en.wikipedia.org

 

[beginner99]

PowerPC 600 - Wikipedia

en.wikipedia.org

You completely missed the original posts smiles at the bottom:

i mean, why not put an X86 decoder on a RISC type CPU and run that? has anyone even thought of that before?!? i doubt it!

The joke here is that all modern x86 CPUs are exactly what he says. a RISC cpu with a x86 decoder at the front. RISC has won. Just not in the way originally intended.

EDIT:

And ARM is the same but with a "cleaner" ISA with less baggage. And hence in general lower power use. Hence the win of ARM in mobile. However once you scale up, this difference become marginal at best and then there is the huge sunken cost of all x86 software and windows. Corporations mostly run on windows and hence tons of legacy windows native business apps, such apps that do not run on windows ARM. All this stuff would have to be ported to make it compatible with ARM. Not going to happen. Having less battery life in latops and laptops beign a bit bigger for the needed cooling doesn't cost the company anything. Porting the software...probably trillions worldwide.

 

[Richie Rich]

The upcoming Cortex-A7X, Cortex-A5X, and Cortex-A3X will have SVE2 upgrade options.

Qualcomm has been confirmed for their next CPU version to support SVE2. Samsung's next CPU design will support SVE2. Huawei/HiSilicon's N5P might also be supporting SVE2, while the N5 version will have it disabled.

2020/2021 is the year of SVE2. With the later Cortex-A7Y, Cortex-A5Y, and Cortex-A3Y having SVE2 being mandatory.

A78 Hercules should be announced in couple of months. If having SVE2 support that would be nice but ARM written in SVE2 presentation this:

These new technologies are not yet part of any announced product roadmap.•But guidance to developers to prepare for future Arm architecture and CPU products.

Page 5: https://s3.amazonaws.com/connect.linaro.org/bkk19/presentations/bkk19-202.pdf

And an announced roadmap is A76, A77 and Hercules(A78?). So no SVE2 silicon till 2022 and no SVE2 for A78. But who knows about Apple, they push super hard and their A15 might have SVE2 one year before Cortex core (possible reason for x86->ARM transition delay, Apple with SVE2 can replace Xeons and increase performance). IMHO there is strong link between ARMv9 and SVE2 announcement and Nuvia begining. Gerard Williams realized that with this advanced extensions an ARM is really going to take server and HPC majority over x86 soon. He had an excellent unique lifetime opportunity to design new Nuvia core based on ARMv9 and SVE2 from scratch and he grab that. Because everybody has to start from scratch at least from FPU side. The combination of Ex-apple engineers and new ARMv9+SVE2 2048-bit FPU makes from Nuvia potentially the most successful server company on the world. Potentially very disruptive to server market. Something like Tesla in automotive.

SVE2= SVE+NEON...extended to 2048-bit........ it means whole FPU must be reworked (it takes about 4 years of intensive work). So no wonder that we need to wait until 2022 for first SVE2 silicon. But since then x86 is screwed because even if Intel'd announce AVX1024 this year they will need 4 year to implement and you know.... it's still 1024-bit only. AMD still wasn't able to implement 7 years old AVX512 (2013). So IMHO only 4 years waiting for SVE2 silicon is very good.

 

[DrMrLordX]

But since then x86 is screwed

Says you. Supporting SIMD wider than 512b will be a monstrous undertaking. I don't think that will be the deciding factor in which ISA predominates even 10 years from now.

The beauty of SVE2 isn't that it can support up to 2048b SIMD. The beauty of it is that it can support variable widths with the same code base. Supporting AVX, AVX2, and AVX512 kinda sucks. You code to different targets each time. With SVE2 you only code to one target.

 

[MangoX]

And I was pretty sure these cost an arm and leg....so I checked- Indeed the cheapest one (4 drives +2ssd bay) costs you >$1500 with only 8GB of RAM. Even these consumer units are extremely prices for the hardware. You pay for the software and the compact form factor.

I just use my pc as NAS. I mean windows can easily be used as media server or what I prefer trivial file shares.

Exactly! I know people who adore their QNAPs and Synology's, but I've always scoffed at the prices. You really pay for the simplicity and convenience. I'd rather do a DIY pc with internal drives on top of Windows. But these days finding a nice chassis with plenty 3.5 bays is a rarity.

 

[beginner99]

But these days finding a nice chassis with plenty 3.5 bays is a rarity.

True. the case (and motherboard) is the limiting factor. You can't get a DIY as compact as a commercial NAS.

 

[Richie Rich]

Says you. Supporting SIMD wider than 512b will be a monstrous undertaking. I don't think that will be the deciding factor in which ISA predominates even 10 years from now.

It's way more than 10 years. Vector width of 2048-bit seems to be reasonable maximum (sweet spot) for CPU from the efficiency point of view. More than 2048-bit vectors are faster and efficiently executed on parallel engines like GPU. It's based on kind of IT law. Same way like in the mechanical engineering you have fuel consumption clash of turbines vs. piston engines. This results in piston engines being more efficient up to 1MW and turbines above that (that's why vehicles have piston engines and power plants have turbines). Turbine has higher heat loss at the same size but with bigger size it's getting better because volume rise with power 3 but surface (which drives heat loss) with only power 2. Big piston engines suffer from higher friction (mechanical rings sealing vs. labyrinth contactless sealing at turbines). That's why there is a inflection point dividing the two territories. IMHO same for vector width for CPU vs. GPU. Intel with AVX512 and it's messy 17 subsets was aimed to screw AMD in a first place (it didn't work because that mess was not adopted in SW ... x86 vector extension is stuck at dead point as a result). While ARM's SVE2 is aimed to be future proof benefit for all.

So if we never see any wider vectors in CPU than 2048-bit because of that principle .... why somebody would like to invest his money into AVX512 ISA which will change soon anyway? It's much wiser invest into optimization for 2048-bit vectors which has vector width designed for 20+ years.

The beauty of SVE2 isn't that it can support up to 2048b SIMD. The beauty of it is that it can support variable widths with the same code base. Supporting AVX, AVX2, and AVX512 kinda sucks. You code to different targets each time. With SVE2 you only code to one target.

I agree, the variable width is important one.
But more important is ARM's focus at auto-vectorization. No need costly hand optimization, compiler can extract a lot of low hanging fruits automatically. SVE has introduced sizeless types which can change vector size on the fly by CPU/FPU. This gives huge power to HW to decide dynamically on the fly how to execute in the most efficient way. Little core like new post-A55 can execute SVE2 at one slow 128-bit FPU pipe and BIG core will execute SVE2 at two 256-bit pipes (having 4x higher FPU throughput while keeping same SVE2 compatibility). This is something what Intel cannot do with his little Atom cores....

 

[ThatBuzzkiller]

Welcome to the ARM world. ARM Holdings dictates what will or won't go into future ARM standards. Anyone not liking their decisions can use a design license (ala Apple). Anyone willing to throw in with ARM Holdings can just use off-the shelf cores, which is what many of the big players are still doing. Most of the design shops that were producing custom ARM cores have stopped. Apple is one of the few holdouts.

I guarantee you, once ARM Holdings includes SVE2 in one of their designs, the major players will use it. Samsung, Qualcomm, and (in their consumer products) Huawei will all be on board.

ARM has many designs so it's more likely that it's customers won't use it if they don't deem it desirable. It's far too premature to claim that any one vendors will support SVE2. Again, show me one of their roadmaps and only then could you attempt to make a meaningful argument in favour of it's standardization.

What EXACTLY are the competing vendors going to use if not SVE2-capable designs? Do you think they're going to go back to custom cores just to avoid it? Why would they? NEON and SVE2 will co-exist in future designs. If you support one, you will support the other. There's no reason to deliberately try to strip out support from SVE2 from future ARM designs.

Do you think they're going to come up with their own ISA extensions? Apple has sort-of gone this way (they have proprietary logic blocks for things like AI). Practically nobody else is doing this. Even Marvell seems to not be taking that approach. They seem to be using fairly-standard ARM designs, albeit with SMT4. What Huawei is doing is unknown since they haven't said much about the differences between the v110 and A76. Never mind that Huawei is still using A76 for the Kirin 990.

Given that Moore's Law (more like an observation) is dying they may never get to use SVE2 in their consumer products. It's hard to predict exactly how things will play out 10 years from now on. All we can say for certain so far is that the post-silicon world is coming since physics will prevent us from minimizing our transistors any further.

Apple isn't the only one doing their own thing with the ARM architecture but Nvidia is as well. Nvidia is already facing hard competition in the high performance computing sector against x86 processors and that's before both vendors are willing to converge on AVX-512 which could cut the value proposition of their main products in the long term. If Nvidia lobbied for ARM's SVE/SVE2 success then they would stand to face even more competition against other ARM vendors as well in the HPC market. It's plenty hard enough for Nvidia to convince ISVs to use their GPUs when the performance gains might not justify it but it'll be even harder for them to justify high prices when faceless ARM vendors with industry standard SVE/SVE2 implementations could achieve comparable performance with their many core designs too.

As for Qualcomm, I don't think you understand their mindset very well. To them the ARM architecture is in the back of their heads. They think connectivity will be more important in the future for consumer applications where they hold some key strategic advantages in their baseband technology to capitalize for that purpose. ARM is just their starting point, not an end to their means.

With Huawei and the Chinese state it's hard to figure out what their true intentions are with technical standards. Their government might be counting on them to create a high performance alternative with the ARM architecture or instead they could be counting on Zhaoxin to create high performance x86 cores for military purposes. We can't exactly tell for certain either way since the communist government is secretly calling most of the shots with high profile projects.

None of the vendors I mentioned above necessarily have to share the same interests with ARM Holdings or with each other which is why it's important to get a grip on the political situation at hand. SVE2 has potential to be derailed because of a lack of cooperation between vendors.

 

[DrMrLordX]

ARM has many designs so it's more likely that it's customers won't use it if they don't deem it desirable.

The main two designs for consumer devices (read: not embedded, though the A5x cores can be used in embedded) are currently:

A7x line - the "big" cores.
A5x line - the "small" cores.

Both lines currently support NEON. Both will support SVE2. What will be interesting to see is how pipes vs. registers will work out in the final designs.

There are no cores from ARM Holdings with more compute power than the A7x line (currently A77, next is A78). That's it. The only option ARM vendors would have at that point is to spend boatloads of extra money doing custom cores for no reason than to strip out SVE2 (which is pointless, since integration of the standard will be painless from the ARM vendor's point-of-view) or to use old cores like Huawei did for the Kirin 990 (instead of A77, which is still a NEON core so, whatever Huawei).

If they want to be stuck on A78 forever, or switch to x86 (hahaha!) or RISC-V, they're free to do that as well. In the real world, companies like Qualcomm will move to SVE2-capable cores just the same way they've been quick to adopt A76 and A77.

There is the odd possibility of ARM Holdings introducing a third line, which might correspond to Hercules, that offers full 512b SVE2 that might not be appropriate for a lot of mobile devices. But we'll have to wait and see what they do.

I'm going to pretend you didn't mention nVidia's ARM efforts. They're a shining example of how not to do custom ARM cores.

None of the vendors I mentioned above necessarily have to share the same interests with ARM Holdings or with each other which is why it's important to get a grip on the political situation at hand. SVE2 has potential to be derailed because of a lack of cooperation between vendors.

I keep trying to get you to stop talking in circles and instead show me something evidence-based. Give me some inductive reasoning here. Huawei has taken a shot at ARM Holdings by refusing to adopt the A77 core for Kirin 990. Publicly they've said that A76 is "fast enough" for their customers and that they're waiting for 5nm to do A77; in contrast, some device manufacturers won't use Snapdragon 865 due to power draw. Coincidence? I think not. So there IS some tension between ARM Holdings and the ARM vendors over the progression of the standard core designs wrt power usage; namely, some vendors (like Huawei) are opting to save money by using a design they've already licensed instead of licensing the new design and trying to tame its power draw.

SVE2 won't exacerbate that problem since the ARM vendors can easily just stick to a single 128b pipe which will be no more or less powerhungry than running one NEON unit today. The only problem ARM Holdings might face is if they keep pushing for wider core designs that use more power by default. But it's pretty easy for ARM vendors to drop clocks or cut features from the standard designs (see Graviton2 and the much-delayed-and-probably-downclocked Rockchip 3588).

 

[Carfax83]

ARM has many designs so it's more likely that it's customers won't use it if they don't deem it desirable. It's far too premature to claim that any one vendors will support SVE2. Again, show me one of their roadmaps and only then could you attempt to make a meaningful argument in favour of it's standardization.

I think this is a salient point. Most ARM CPUs are used in mobile or low power devices, so what would be the point for say a company like Apple to put huge vector units on their CPUs that are going to be used in smartphones and tablets? Wouldn't that defeat the purpose of a mobile CPU, which is to be as power efficient as possible?

I could see why companies that are designing ARM based CPUs for HPC or servers would want to implement SVE2, but they are a minority if I understand correctly.

 

[ThatBuzzkiller]

I keep trying to get you to stop talking in circles and instead show me something evidence-based. Give me some inductive reasoning here. Huawei has taken a shot at ARM Holdings by refusing to adopt the A77 core for Kirin 990. Publicly they've said that A76 is "fast enough" for their customers and that they're waiting for 5nm to do A77; in contrast, some device manufacturers won't use Snapdragon 865 due to power draw. Coincidence? I think not. So there IS some tension between ARM Holdings and the ARM vendors over the progression of the standard core designs wrt power usage; namely, some vendors (like Huawei) are opting to save money by using a design they've already licensed instead of licensing the new design and trying to tame its power draw.

You're guilty of this just as much as I am. You keep unilaterally insist that ARM will somehow be only offering standard core designs with SVE2 despite none of their roadmaps having committed to it thus far.

SVE2 won't exacerbate that problem since the ARM vendors can easily just stick to a single 128b pipe which will be no more or less powerhungry than running one NEON unit today. The only problem ARM Holdings might face is if they keep pushing for wider core designs that use more power by default. But it's pretty easy for ARM vendors to drop clocks or cut features from the standard designs (see Graviton2 and the much-delayed-and-probably-downclocked Rockchip 3588).

Limiting SVE2 to 128-bit vectors would make it more of a competitor to NEON instead of being a successor so how would it's implementation be justified among customers ?

At that point you don't have a successor to NEON but you now just have a competitor to NEON! FYI, SVE/SVE2 is NOT an extension to NEON and has it's own separate instruction encoding. By comparison SSE/AVX/AVX-512 can all share the same instruction encoding such as EVEX coding so it implies that x86's SIMD extensions shares hardware implementation details with each other.

Justify why ARM must make it's customers pay for the redundancy of supporting both NEON and SVE2 if the latter is limited to 128-bit length vector to begin with ?

ARM themselves even say that SVE/SVE2 is NOT intended to replace NEON which very well means that they expect most of their big customers to keep continuing to only have NEON until the political situation changes and they all collectively desire to converge on SVE/SVE2.

 

[DrMrLordX]

You're guilty of this just as much as I am.

No I'm not. You came out saying there's some hot political battle between ARM vendors and ARM Holdings. I'm asking for evidence. News clips, lawsuits, something, anything. All I'm doing is telling you you're wrong. There's no evidence whatsoever to support that Qualcomm is trying to kill SVE2 (or anyone else for that matter).

I did provide you evidence that most ARM vendors are using standard ARM designs. Each time ARM rolls out a new design, Qualcomm uses it. Huawei was doing this up until A77 (and they've announced they'll use A77 once they move to TSMC 5nm). Samsung killed their custom core shop to go back to reference ARM designs. There's all the evidence you need that the major players are going to go wherever ARM Holdings takes them.

There's nothing about SVE2 that's going to increase the cost of anything in terms of licensing fees or transistor budget. The amount of transistor budget they'll have to spend supporting SVE2 and Neon redundantly is tiny. NEON support won't "compete" with SVE2. It's there for legacy software support. The beauty of SVE2 is that you code once for 128b or 2048b targets, or anything in between. Most vendors who use NEON today, will implement ARM designs with a single 128b SVE2 pipe while simultaneously supporting NEON (again, at little-to-no cost). That way even dinky A59 cores (or whatever) will support SVE2 code just like beefy cores going into future ARM server designs. No need to target SSE4.x, AVX128, AVX256, AVX2, and AVX512 in the same binary, which is what happens in the x86 world.

edit: further evidence of major ARM vendors lining up to support SVE and SVE2:

https://www.reddit.com/r/hardware/comments/bsp3hk

Fujitsu A64FX. My guess it is already being produced.

Marvell/Cavium ThunderX3 ("Triton") is probably going to support SVE.

HiSilicon Hi1630 (Kunpeng 930 series) will be released in 2020. It will have SVE, and SMT.

Though I have my doubts about TX3 supporting SVE. Regardless, if that post is true, there will be at least two ARM server CPUs to support it. I had thought A64FX would be the first and last.

 

[ThatBuzzkiller]

No I'm not. You came out saying there's some hot political battle between ARM vendors and ARM Holdings.I'm asking for evidence. News clips, lawsuits, something, anything. All I'm doing is telling you you're wrong. There's no evidence whatsoever to support that Qualcomm is trying to kill SVE2 (or anyone else for that matter).

That's a mischaracterization of my statement. My main point is that there is a political battle between other ARM vendors, not against ARM Holdings.There's also no evidence that Qualcomm intends to support SVE2 either so you need to stop assuming that it will somehow become an industry standard.

I did provide you evidence that most ARM vendors are using standard ARM designs. Each time ARM rolls out a new design, Qualcomm uses it. Huawei was doing this up until A77 (and they've announced they'll use A77 once they move to TSMC 5nm). Samsung killed their custom core shop to go back to reference ARM designs. There's all the evidence you need that the major players are going to go wherever ARM Holdings takes them.

Just because ARM's customers are using their standard designs does NOT make for a good argument in favour that they'll somehow force the industry to standardize SVE2.

There's nothing about SVE2 that's going to increase the cost of anything in terms of licensing fees or transistor budget. The amount of transistor budget they'll have to spend supporting SVE2 and Neon redundantly is tiny. NEON support won't "compete" with SVE2. It's there for legacy software support. The beauty of SVE2 is that you code once for 128b or 2048b targets, or anything in between. Most vendors who use NEON today, will implement ARM designs with a single 128b SVE2 pipe while simultaneously supporting NEON (again, at little-to-no cost). That way even dinky A59 cores (or whatever) will support SVE2 code just like beefy cores going into future ARM server designs. No need to target SSE4.x, AVX128, AVX256, AVX2, and AVX512 in the same binary, which is what happens in the x86 world.

Now, this is a false statement. How do you expect ARM to make money out of SVE2 ? Are they just going to make the extension free of charge ?

Also, I doubt implementing a separate SVE2 unit is anywhere near as trivial you as speak. On x86 SSE/AVX/AVX-512 shares the same exact unit executing those instructions while SVE2 is totally separate from NEON as evidenced by their incompatible instruction encodings.

While a vector length agnostic programming model is nice for portability reasons it also adds predication overhead too according to the research findings in this paper compared to using a vector length specific instruction set. You can chide x86 as much as you want for their non-portable vector length programming model but there's a very good reason for why the engineers behind the architecture decided against making that trade-off.

Not only is there's no implementation redundancy to x86's SIMD extensions but their programming model has less overhead too so how can you even keep justifying a 128-bit vector length implementation for SVE2 when it could potentially run slower than a NEON optimized application ?!

 

[Nothingness]

NAS units all seem to be low cost. Instead of moving to better SoCs most of the more expensive NAS products just move from low cost ARM to low cost Atom based CPUs...

Look at the link I provided to Synology. It seems to contradict what you wrote as some of their NAS use Intel Xeon chips. Unless you were only talking about non pro NAS?

 

[Nothingness]

Also, I doubt implementing a separate SVE2 unit is anywhere near as trivial you as speak. On x86 SSE/AVX/AVX-512 shares the same exact unit executing those instructions while SVE2 is totally separate from NEON as evidenced by their incompatible instruction encodings.

Do you think having different instruction encoding means different execution units? Or do you think that decoding units are larger than execution units? Or that all extra instructions in AVX-512 don't have a significant impact on SSE/AVX/etc. decoding units?

Not only is there's no implementation redundancy to x86's SIMD extensions but their programming model has less overhead too so how can you even keep justifying a 128-bit vector length implementation for SVE2 when it could potentially run slower than a NEON optimized application ?!

That's a very good point. I guess what matters is having SVE2 as widespread as possible to increase adoption and have that code run faster on CPU that will have wider SVE2 units. That's an option Intel doesn't offer.

I hope ARM & co won't do the same stupid mistake Intel marketing did by fusing off AVX on many of their CPU: in 2020, 12 years after introduction, AVX is still not available on all Intel CPU, which has slowed down its usage a lot. AVX-512 will be even worse. If we follow your logic of a battle between ARM licensees, all I can conclude is that Intel has a severe form of schizophrenia

 

[serpretetsky]

Having the instructions available on paper is different from every CPU actually having all of them implemented,is there even a real ARM CPU that has all instructions available?
RISC CPUs are efficient only if they get rid of instructions they don't (often) use.
RISC started in the 80's instead of the 70' but that's not really a huge difference,the whole ideology of RISC was to only use the most important instructions to reduce overhead.

Berkeley RISC - Wikipedia

en.wikipedia.org

In your same link I quoted below. This is kind of why I was wondering if RISC vs CISC (or ARM vs x86 vs MIPS vs RISCV etc) is important these days or not for the efficiency and performance of cpu. Or if its really all about the microarchitecture underneath all of that.

Techniques developed for and alongside the idea of the reduced instruction set have also been adopted in successively more powerful implementations and extensions of the traditional "complex" x86 architecture. Much of a modern microprocessor's transistor count is devoted to large caches, many pipeline stages, superscalar instruction dispatch, branch prediction and other modern techniques which are applicable regardless of instruction architecture. The amount of silicon dedicated to instruction decoding on a modern x86 implementation is proportionately quite small, so the distinction between "complex" and RISC processor implementations has become blurred.

 

[ksec]

I will give you a straight forward simple answer:

It is not cost effective. ( Yet )

Go and look up the current median salary of Software Engineers in Silicon Valley. Especially those with low level C, C++ Skill Set.

People continuously talk about technical and completely ignore the business side of things.

And there isn't anything new to discuss, there is already a thread about AWS Graviton ( Server Chip using ARMv8 ). And almost everything said there applies here as well. You should read up on it.