Nvidia announces better ARM support and new ARM CPU - x86 not the main player anymore

[Bombastinator]

6 minutes ago, Vishera said:

ARM CPUs are not new,even in the consumer space...

That might have been soon after windows RT failed.  Would be a reason for something like that to be believed.  People are wrong a lot.

[Vishera]

2 minutes ago, Bombastinator said:

That might have been soon after windows RT failed.  Would be a reason for something like that to be believed.  People are wrong a lot.

Yeah,and if you want one with an NVIDIA GPU - just buy a Jetson.

[Bombastinator]

1 minute ago, Vishera said:

Yeah,and if you want one with an NVIDIA GPU - just buy a Jetson.

Just don’t watch the jetsoms movie.  It’s even worse than one might imagine.  96min I’ll never get back.

[RedRound2]

30 minutes ago, Vishera said:

ARM CPUs are not new,even in the consumer space...

Im not sure which part of the sentence claimed that ARM CPUs were new. The implication there was that it would never be a thing in mainstream PCs like your regular everyday laptops and desktops.

[RedRound2]

On 4/14/2021 at 3:20 AM, StDragon said:

ARM will stay relegated to Phone, the Macbook, and Surface Pro X. You won't find and ARM based PC short of some AIO unit.

 

Due to licensing, I'm thinking RISC-V will explode in adoption in China and India. That momentum will be carried over with possibly Intel and AMD creating a RISC-V chip for the PC market.

 

PC will jump from X86 to RISC-V over the next 10 years. That's my prediction.

Apple is completely transitioning to ARM on their whole lineup including the Mac Pro, so that's already a provably false statement.

 

RISC V is nowhere near mass adoption, especially in high performance applications. Remember, people have been working with ARM processors for cutting edge performance for over 10 years. We haven't even begun with that with RISC V

On 4/14/2021 at 7:49 AM, leadeater said:

ARM CPUs as fast as or faster than Intel Xeons have existed for many years. Apple only did consumer performance ARM first not actual performance ARM as a whole first.

That just looks like a way how the same people today would spin it today. 

Even in single core performance (I'm not sure just asking)? And Xeon isn't really what you're supposed to be crowning as the king of performance

On 4/14/2021 at 7:49 AM, leadeater said:

It's never been an issue with ARM not having or being capable of having the performance, nobody wants a complete an utter software break in a proposed switch and not even Apple users want that either which is why Rosetta exists and has existed in the past for the same reason. So until that and ongoing software is seen as a viable ecosystem on Windows a switch simply won't happen.

But that can be fixed. It's not a physics or technological limitation that is stopping here. So it's quite dumb to think that software changes won't happen at all - because err, developers are lazy.

 

Once competitive ARM processors enters the PC market, more and more mainstream softwares will be written natively. Sure, microsoft can never do it nowhere near as fast as Apple, but to say that ARM and it's performance implications haven't piqued developers interest is inaccurate. I can see ARM laptops easily being a viable alternative to x86 laptops in about the next 5 years time. And hopefully microsoft will get their shit together and actually come up with a plan and strategy to make ARM version of Windows competitive with a proper plan like how Apple did with universal binary and Rosetta 2

[Vishera]

12 minutes ago, RedRound2 said:

The implication there was that it would never be a thing in mainstream PCs like your regular everyday laptops and desktops.

But they are in everyday regular laptops,desktops are a different story.

If you want an ARM laptop with Windows,then there is the Microsoft Surface Pro X for example,there are also Chromebooks,and more.

[Bombastinator]

19 minutes ago, RedRound2 said:

Apple is completely transitioning to ARM on their whole lineup including the Mac Pro, so that's already a provably false statement.

 

RISC V is nowhere near mass adoption, especially in high performance applications. Remember, people have been working with ARM processors for cutting edge performance for over 10 years. We haven't even begun with that with RISC V

That just looks like a way how the same people today would spin it today. 

Even in single core performance (I'm not sure just asking)? And Xeon isn't really what you're supposed to be crowning as the king of performance

But that can be fixed. It's not a physics or technological limitation that is stopping here. So it's quite dumb to think that software changes won't happen at all - because err, developers are lazy.

 

Once competitive ARM processors enters the PC market, more and more mainstream softwares will be written natively. Sure, microsoft can never do it nowhere near as fast as Apple, but to say that ARM and it's performance implications haven't piqued developers interest is inaccurate. I can see ARM laptops easily being a viable alternative to x86 laptops in about the next 5 years time. And hopefully microsoft will get their shit together and actually come up with a plan and strategy to make ARM version of Windows competitive with a proper plan like how Apple did with universal binary and Rosetta 2

Having a platform supplant another platform in its space has happened, but when it’s happened it’s been a very slow process.  Multiple years. Arguably decades in some instances.  There is a potential shortcut if one system is able to run the software of the other though. 

Edited April 15, 2021 by Bombastinator

[leadeater]

1 hour ago, RedRound2 said:

But that can be fixed. It's not a physics or technological limitation that is stopping here. So it's quite dumb to think that software changes won't happen at all - because err, developers are lazy.

No they aren't lazy, there isn't a economic or business driving reason to change out an already working toolset and also require a lot of professional development to upskill employees (which isn't free) so what would be for them no gain at all.

 

You have a very simplistic view of a very complex situation and also not realize that for the software ecosystem side of the equation switching has no direct benefit to them at all. Windows ecosystem is in at least, very minimum, a 3 way standoff with nobody willing to shoot first. Apple on the other hand was at most in a 2 way standoff but also had the much bigger gun.

 

1 hour ago, RedRound2 said:

Sure, microsoft can never do it nowhere near as fast as Apple, but to say that ARM and it's performance implications haven't piqued developers interest is inaccurate.

So you are going to ignore the fact that there actually isn't a direct performance gain switching to ARM at all? What's inaccurate is saying ARM has greater performance to x86.

[leadeater]

1 hour ago, RedRound2 said:

And hopefully microsoft will get their shit together and actually come up with a plan and strategy to make ARM version of Windows competitive with a proper plan like how Apple did with universal binary and Rosetta 2

They did and literally everyone yelled and screamed and tossed their toys out and refused to adopt it, UWP.

[GDRRiley]

19 hours ago, leadeater said:

Nvidia's much deeper in to this market than most people outside of it are aware. Nvidia is the ODM/OEM for all the server vendors that use the 8 GPU and 16 GPU NVLink configurations. So for example if you look at an HPE Apollo 6500 that board inside of it that has the SXM GPUs on it is designed, made and supplied by Nvidia and everyone designs around this. Only PCIe systems are fully designed by the server vendors.

thing is for supercomputers I think HPE is doing something else because they aren't using a PCIE Switch

7 minutes ago, leadeater said:

So you are going to ignore the fact that there actually isn't a direct performance gain switching to ARM at all? What's inaccurate is saying ARM has greater performance to x86.

grace is looking to be slower than 2nd let alone 3rd gen epyc. at least based on the score they claim

[leadeater]

5 minutes ago, GDRRiley said:

thing is for supercomputers I think HPE is doing something else because they aren't using a PCIE Switch

Yep, depends on platform system. But HPE is super in to Photonics and composable infrastructure, some very forward thinking stuff that I'm not sure when will hit the market, even in your area/type of clientele.

 

Had HPE come in a few times and give presentations on that sort of stuff they are doing research in to, all super cool and super interesting, and super out of our league lol

[leadeater]

2 hours ago, RedRound2 said:

That just looks like a way how the same people today would spin it today. 

Even in single core performance (I'm not sure just asking)? And Xeon isn't really what you're supposed to be crowning as the king of performance

No it's not a spin, it's literal fact. And we're talking about years here, when Xeon was the crown and yes HPC ARM CPUs have excellent single core performance which btw is completely irrelevant for that use case but whatever they still have that. The only TSMC 5nm ARM products on the market are Apple's, they have a very distinct silicon process advantage to other companies i.e. Fujitsu.

 

The Fujitsu A64FX as an example was released in 2019 with SVE 512bit with performance as high as lower-mid/mid range GPU but is a CPU and is able to do all the things a GPU cannot while also having higher Double Precision performance than Single Precision which is basically unheard of. This CPU currently powers the worlds fastest supercomputer with more than 3 times the Rmax of the second place which is powered by Nvidia V100 GPUs. A100 powered systems are in the list but the very big ones aren't up and running just yet (or not submitted) but it'll lose the number one spot this year. Of course Fujitsu has been working on a new and faster CPU.

 

Ampere Computing as seen on previous page with their Altra CPUs are also very fast, released in 2020, with a 128 core version coming soon.

 

These are after all chalk and cheese design requirements, but as to the point high performance ARM has existed for a long time, even before these, and anybody could design and manufacture a high performance mobile or desktop class product during that time if they so wished but nobody did because their wasn't a market for it. Apple M1 design is very good at what it's targeted for but it in no way proved that ARM could offer high performance, this was proven long before that (even by Apple themselves).

 

P.S. As an aside the A64FX CPU is faster than the GPU in the M1 and each core uses ~3.9W which is very similar to Apple Firestorm cores in the M1 at ~3.45W, and for vector processing the A64FX cores are faster than Firestorm cores (it was purpose architected for vector math after all and Fujitsu created the SVE specification for ARM).

[Vishera]

55 minutes ago, leadeater said:

while also having higher Double Precision performance than Single Precision which is basically unheard of.

Well,The IBM PowerXCell 8i could do the same thing,I have one myself as part of my collection.

[leadeater]

13 minutes ago, Vishera said:

Well,The IBM PowerXCell 8i could do the same thing,I have one myself as part of my collection.

Well it's actually the same for SP and DP for that one but still equally as unheard of 

 

Edit:

Or perhaps not?? Conflicting documentation lol

Quote

This gives the eight SPEs per chip a combined double-precision peak performance of 102.4 GFLOPS or exactly one-half the chip's single-precision performance (~204.8 GFLOPS)

https://www.ibm.com/developerworks/rational/cafe/docBodyAttachments/2636-102-1-3388/technology_cell_IDC_report_on_PowerXCell.pdf

[LAwLz]

3 hours ago, leadeater said:

So you are going to ignore the fact that there actually isn't a direct performance gain switching to ARM at all? What's inaccurate is saying ARM has greater performance to x86.

It depends on which ARM processor and which x86 processor you're comparing.

I am getting kind of tired of people on this forum not being able to think in nuances and always dealing with absolute black and white thinking.

Can we please stop talking about "ARM vs x86" and instead talk about architectures? For example Firestorm vs zen3, or N1 vs Ice Lake SP.

It makes way more sense to do that than to try and generalize like 40 years of processor architectures into a single performance number and then try and make arguments based on very specific numbers from that generalization.

 

Unless of course we are talking about very general things like, how does the ISA of a certain architecture affect the way the decode engine looks, as an example.

 

 

 

3 hours ago, leadeater said:

They did and literally everyone yelled and screamed and tossed their toys out and refused to adopt it, UWP.

UWP was/is nothing like Rosetta 2 or universal binaries.

UWP was and still is garbage and doesn't get used because it sucks ass, not because Windows developers dislike change.

 

If Rosetta 2 is a fantastically baked apple pie, then UWP and the WoA translation layer is an apple muffin filled with broken shards of glas, and rotten, wormfilled apples.

You can't compare the two and go "well Windows users clearly don't like apples". I like apples, apple pies and apple muffins. What I don't like is eating shit and then getting blamed for not liking apples.

 

 

  

3 hours ago, GDRRiley said:

grace is looking to be slower than 2nd let alone 3rd gen epyc. at least based on the score they claim

What scores are you referring to?

[leadeater]

1 hour ago, LAwLz said:

UWP was/is nothing like Rosetta 2 or universal binaries.

UWP was and still is garbage and doesn't get used because it sucks ass, not because Windows developers dislike change.

Not true for Universal Binaries, and in a way Rosetta but in the reverse direction. UWP allows iOS and Android Apps to run on Windows (API Bridges). UWP itself was designed for multiple targeted platform so a single UWP application could natively run on x86 or ARM, if actually compiled that way and if UWP got any traction. UWP was Microsoft's plan for this.

 

Quote

Universal Windows Platform (UWP) is a computing platform created by Microsoft and first introduced in Windows 10. The purpose of this platform is to help develop universal apps that run on Windows 10, Windows 10 Mobile, Xbox One and HoloLens without the need to be rewritten for each. It supports Windows app development using C++, C#, VB.NET, and XAML. The API is implemented in C++, and supported in C++, VB.NET, C#, F# and JavaScript.[1] Designed as an extension to the Windows Runtime (WinRT) platform first introduced in Windows Server 2012 and Windows 8, UWP allows developers to create apps that will potentially run on multiple types of devices.[2]

 

 

And yes I wasn't implying it was that good or anything but nobody bothered to give it a chance either or help improve it so it's basically dead. But UWP is a thing. Rosetta is a different situation because it serves a specific need that nothing else does or could where as UWP was sitting aside an existing application model that was decades mature while also being feature inferior in every way. So everyone just ignored it and refused to adopt it and I can say confidently that UWP cannot improve without adoption that is required to aid in improving it. So it's crap and will forever be crap while this is the case, maybe Microsoft can make enough improvements over time to get people willing to try it again but progress will be slower.

 

WoA is not required in a fully matured UWP ecosystem, caveat being every application being UWP otherwise WoA is required for Win32 on ARM.

 

1 hour ago, LAwLz said:

It depends on which ARM processor and which x86 processor you're comparing.

I am getting kind of tired of people on this forum not being able to think in nuances and always dealing with absolute black and white thinking.

Can we please stop talking about "ARM vs x86" and instead talk about architectures? For example Firestorm vs zen3, or N1 vs Ice Lake SP.

Yes and that was literally my point as well. Just saying ARM has better performance or has better power efficiency is an incorrect statement. It's also incorrect to say that ARM has a unique capability to have superior performance to x86.

 

What makes ARM attractive is it's openness and anyone's ability to contribute to it and also extend upon it and make custom designed for a specific purpose like the Fujitsu A64FX. Fujitsu could never have achieved that with x86, so that's where ARM is most attractive compared to x86 or POWER. But having this capability does not mean every company has the capacity, capability and customer base to do it.

[leadeater]

1 hour ago, LAwLz said:

What scores are you referring to?

 

Quote

But other than that, all the company is saying is that the cores should break 300 points on the SPECrate2017_int_base throughput benchmark, which would be comparable to some of AMD’s second-generation 64 core EPYC CPUs

https://www.anandtech.com/show/16610/nvidia-unveils-grace-a-highperformance-arm-server-cpu-for-use-in-ai-systems

 

However I don't think this is very meaningful as the platform design for this is all memory subsystem and interconnect for the purpose of most efficiently getting work to the GPUs, the CPUs are primarily a facilitator of that so even though the cores in Grace aren't groundbreaking performance the system as a whole will be much faster which was the design purpose. I doubt you'll be able to buy Nvidia Grace CPUs as an individual product.

[LAwLz]

45 minutes ago, leadeater said:

Not true for Universal Binaries, and in a way Rosetta but in the reverse direction.

What do you mean? Universal binaries and UWP are two separate things. One does not imply the other.

UWP was an application platform that could (optionally) support multiple ISA binaries, but it was way more than that. Rosetta 2 and universal binaries are not at all like UWP. They can't and should not be compared.

 

 

52 minutes ago, leadeater said:

UWP allows iOS and Android Apps to run on Windows (API Bridges).

Again, what are you on about? This has nothing to do with Rosetta 2 or universal binaries.

Also, the UWP API bridges were terrible. The Android one was completely scraped and Microsoft just went "just develop with Xamarin instead", and the iOS ones aren't (from what I've heard) that pleasant to work with either. It doesn't even support Swift for crying out loud.

 

 

1 hour ago, leadeater said:

UWP itself was designed for multiple targeted platform so a single UWP application could natively run on x86 or ARM, if actually compiled that way and if UWP got any traction. UWP was Microsoft's plan for this.

Yes, but UWP was also far more than just a universal binary, and since it was ass in a lot of regards nobody bothered.

You're comparing apples and oranges.

 

 

1 hour ago, leadeater said:

And yes I wasn't implying it was that good or anything but nobody bothered to give it a chance either or help improve it so it's basically dead.

Stop blaming people for not wanting to eat an apple muffin full of broken glass.

"Oh people should have tasted it and then helped Microsoft refine the recipe". No. Microsoft shouldn't have served it to begin with before it was decent.

 

 

1 hour ago, leadeater said:

I can say confidently that UWP cannot improve without adoption that is required to aid in improving it.

Why not?

Sounds to me like you're just blaming users for Microsoft making a shitty product.

 

 

1 hour ago, leadeater said:

WoA is not required in a fully matured UWP ecosystem, caveat being every application being UWP otherwise WoA is required for Win32 on ARM.

I assume that you mean the binary translation layer when you say WoA, correct?

I'd still want a binary translation layer even if UWP was the standard on Windows. UWP is not necessarily ISA-independent, and there will always be cases like old programs that haven't been updated needing to run, even if all the actively developed software had been ported to UWP.

 

 

1 hour ago, leadeater said:

Just saying ARM has better performance or has better power efficiency is an incorrect statement.

Yes, just like saying x86 has better performance or better power efficiency is incorrect. It depends on which architectures you are comparing.

 

1 hour ago, leadeater said:

It's also incorrect to say that ARM has a unique capability to have superior performance to x86.

I think that depends on what exactly is being discussed. ARM has benefits over x86 that can translate to higher performance.

 

1 hour ago, leadeater said:

What makes ARM attractive is it's openness and anyone's ability to contribute to it and also extend upon it and make custom designed for a specific purpose like the Fujitsu A64FX. Fujitsu could never have achieved that with x86, so that's where ARM is most attractive compared to x86 or POWER.

That's certainly one of the things that makes it attractive. I wouldn't say it's the only one though.

[LAwLz]

15 minutes ago, leadeater said:

https://www.anandtech.com/show/16610/nvidia-unveils-grace-a-highperformance-arm-server-cpu-for-use-in-ai-systems

 

However I don't think this is very meaningful as the platform design for this is all memory subsystem and interconnect for the purpose of most efficiently getting work to the GPUs, the CPUs are primarily a facilitator of that so even though the cores in Grace aren't groundbreaking performance the system as a whole will be much faster which was the design purpose. I doubt you'll be able to buy Nvidia Grace CPUs as an individual product.

Ah I see. Thanks for the link and quote.

I agree with you as well. Judging Grace by the CPU performance is like judging a fish by its ability to climb a tree.

 

It would be very interesting to know which CPU architecture it uses, at what frequency and how many cores, to get that 300 point score.

ARM Neoverse has already shown with the N1 that it can compete with both Intel and AMD for datacenters, so it will be very interesting to see how big of an improvement the next gen core brings.

[igormp]

1 hour ago, LAwLz said:

It would be very interesting to know which CPU architecture it uses, at what frequency and how many cores, to get that 300 point score.

ARM Neoverse has already shown with the N1 that it can compete with both Intel and AMD for datacenters, so it will be very interesting to see how big of an improvement the next gen core brings.

It's based on Neoverse.

[WolframaticAlpha]

On 4/12/2021 at 10:46 PM, Drama Lama said:

I‘m wondering if Nvidia might try to make their own game console/ pc at some point

I guess not. They are earning money with the switch

[Drama Lama]

47 minutes ago, WolframaticAlpha said:

I guess not. They are earning money with the switch

Yea but with their own platform Nvidia would get the higher margins (they don’t have Nintendo between them and the consumer) and Maybe even more important is that Nvidia would be able to collect a 30% or maybe lower cut of all games, DLCs and other purchases.

Though honestly I don’t think it’s likely because the console market is full of a few dominant players and huge investments would be required to get in the market.

[LAwLz]

1 hour ago, igormp said:

It's based on Neoverse.

Yeah I know, but Neoverse is a product family, not an architecture.

Right now we have two different architectures in the Neoverse family. N1 and E1.

 

Grace will probably be based on N2 (not released yet). What I am interested in is how much better N2 is compared to N1.

ARM's own figures claims ~50% higher IPC which would be fantastic, but that remains to be seen.

[leadeater]

3 hours ago, LAwLz said:

ARM Neoverse has already shown with the N1 that it can compete with both Intel and AMD for datacenters, so it will be very interesting to see how big of an improvement the next gen core brings.

 

1 hour ago, igormp said:

It's based on Neoverse.

 

So is the Altra Q80-33 (N1) and other models in that series. You can see a review for it on Anandtech, though to get the most up to date comparisons you should read the Ice Lake-SP review which has the data in it for the Altra and EPYC 7003.

 

Just as a note though the Altra is targeting very high core count (80 now 128 soon) and parallel workloads so hasn't done with a very dense high IPC core design, but does have some very strong points anyway.

https://www.anandtech.com/show/16594/intel-3rd-gen-xeon-scalable-review

https://www.anandtech.com/show/16315/the-ampere-altra-review

https://www.servethehome.com/ampere-altra-max-m128-30-128-core-arm-server-update/ (Soon^TM)

 

 

(Note this is measuring single core but under full all core load)

 

Multithread graphs are a bit big so go to source to see them, the Q80-33 is very competitive there.

[Bombastinator]

43 minutes ago, Drama Lama said:

Yea but with their own platform Nvidia would get the higher margins (they don’t have Nintendo between them and the consumer) and Maybe even more important is that Nvidia would be able to collect a 30% or maybe lower cut of all games, DLCs and other purchases.

Though honestly I don’t think it’s likely because the console market is full of a few dominant players and huge investments would be required to get in the market.

They’d have a branding disadvantage and they’d have to attract devs to their platform.  If it was 10 years ago buying, say, Sega for their brand and doing a console that way might have been an option.  Sega fans are getting older though. Console games are a different market than chip building.  If they want in on Nintendo’s cut they can always just buy some.

Edited April 15, 2021 by Bombastinator