Review Raspberry PI (ARM A72) vs EPYC for DC study. Interesting results.

[Markfw]

OK, now this is a very specific test, but as much as they are disparate, I don't think other scenarios would be that different.

So I got an 8 gig Raspberry PI (8 gig, 32 gig HD, A72 v8 4 core 1.5 ghz). Its does the Open Pandenics covid 19 WCG unit, 4 cores times 7.5 hours

My EPYC 7742 does 128 of these same units at 2 ghz in 2.5 hours.

So it would take 32 PI's to do the same work, but 3 times slower. An 32 PIs at 5 watts each is 160 watts, vs the 7742 @250 watts. (motherboard, memory and all) So more power and 3 times longer for the same work is 2 times the power usage. (1200 watts total vs 625)

Cost... 32 PIs at about $60 each (including power supplies and cables, probably more) would be at least $1920.

The EPYC is about $4000+480+580 or $5060. (I paid $3000 less, that retail on ebay)

So the EPYC effectively less because it cost 2 times less electricity for the same work.

This seems WAY different than I am seeing when people talk about ARM and efficiency and power usage. Please tell me where I am mistaken in my math., But don't be a jerk if you find the errors of my ways. The run times are real.. The PI is at 7.5 hours and hasn't actually finished a unit yes, no other tasks running on the PI. And I am looking at several units on the 7742, one is at 97% in 2:21.


Edit: The first unit finished in 7:50, almost 8 hours. The next 2 units are at 8 hours and still running at 96 and 98%.

And my comment on "wins by a large margin" has to do with total power usage, which in data centers, and for me is a big deal.

 

[Markfw]

The problem is that Mark then applies his DC project specific conclusions to everyone and every application. If it isn't the best for his DC products, then we are wrong no matter what our usage is.

Thats just not only wrong, but borders on trolling. I don't like ARM since at the moment it has very limited uses. Phones, and proprietary servers.

 

[Thala]

Actually the problem is very clear to everyone. It is impossible to buy a competent (read mainstream, not some joke effort from Q vendor) Arm workstation with Windows or Linux. Once Apple is out with ARM Macs, that problem will go away.

Technically we already have Surface Pro X and other Windows ARM devices, where anyone can run standard Linux benchmarks. - sure we are limited with Cortex A76 at the moment but sure better than Raspberry PI Cortex A72 or other ARM SBCs.

 

[piokos]

Thats just not only wrong, but borders on trolling. I don't like ARM since at the moment it has very limited uses. Phones, and proprietary servers.

Not only it's true, you've just agreed with it.

"Very limited uses": phones (LOL).

"Priority servers" - I don't even know what that means.
As of today pretty much any service offered by AWS (estimated 33% market share) could be running on ARM. Often you don't even know what's underneath (which is exactly how it should work). They're only limited by the number of chips that TSMC can make them.

 

[MrTeal]

You can get single board computers with A73, but I'm not aware of any with a newer core like the A77. I'd actually be interested if anyone else knows of one.

 

[piokos]

You can get single board computers with A73, but I'm not aware of any with a newer core like the A77. I'd actually be interested if anyone else knows of one.

Maybe if you search long enough, in some very niche products. But mainstream SBCs use older cores, mostly A73/53. I've seen some A55 lately. They're just cheaper and perfectly functional. SBCs aren't that focused on performance.
Top performing cores are useful in devices that need the potential: smartphones and tablets.

 

[Markfw]

Not only it's true, you've just agreed with it.

"Very limited uses": phones (LOL).

"Priority servers" - I don't even know what that means.
As of today pretty much any service offered by AWS (estimated 33% market share) could be running on ARM. Often you don't even know what's underneath (which is exactly how it should work). They're only limited by the number of chips that TSMC can make them.

Please read ... Priority <> proprietary

 

[Doug S]

If you need to build a server and Apple does not sell you the CPUs, why would you pay attention to Apple?

If you want to judge the performance of ARM vs x86, that would be a hell of a lot better than comparing based on a five year old A72 in a RPi and starting a thread claiming EPYC wins by a mile.

 

[Doug S]

Actually the problem is very clear to everyone. It is impossible to buy a competent (read mainstream, not some joke effort from Q vendor) Arm workstation with Windows or Linux. Once Apple is out with ARM Macs, that problem will go away.

It will still be impossible to buy an ARM workstation with Windows or Linux after Apple releases the ARM Macs. There's approximately zero chance they release Windows or Linux drivers for their GPU, or provide enough info for someone else to write one, and likely no way to install a third party GPU. Linux and Windows will be available on the Mac only via VM.

 

[Markfw]

If you want to judge the performance of ARM vs x86, that would be a hell of a lot better than comparing based on a five year old A72 in a RPi and starting a thread claiming EPYC wins by a mile.

I was mistaken in several areas. I said right in the first post "Please tell me where I am mistaken in my math., But don't be a jerk if you find the errors of my ways. ". Well, assuming ARM v8 was the latest (which apparently its not, there is 8.1,8.2,8.3 etc...) was a mistake. Not knowing it was on 28nm was a mistake. I thought that ARM was very efficient, I thought these were new (within a couple of years), since the are PI v4. Rome is 2-3 years old. So I thought I was comparing recent to recent tech.

APPARENTLY NOT.

And almost nobody went by request to be nice pointing it out, I suspected something was wrong, which is why I said that.

 

[Doug S]

And what if it turns out that Apple only makes ARM chips up to high-end mobile level (so MacBooks, Mac Minis)? No workstations?

There's no chance of that. If they were planning on dropping the Mac Pro they wouldn't have bothered to introduce the new one recently.

 

[dullard]

Thats just not only wrong, but borders on trolling. I don't like ARM since at the moment it has very limited uses. Phones, and proprietary servers.

I don't think we need to go into detail on your first statement, just for the rest of us, please look up your history of posts with open eyes (especially before they were edited).

As for the second, do you even understand what an ARM processor is and what products use them? Sure, phones and a few servers. But also tablets, watches, entertainment systems, smart devices. Have you ever used a chip-and-pin credit card at a gas station, grocery store, or restaurant? Have you heard of Fitbit, wi-fi dongles, cable TV boxes, GPS units? Have you used a USB drive? I can't honestly fathom your "very limited uses" phrase.

 

[blckgrffn]

Well, Anandtech did review this:

Ampere® eMAG® 64bit Arm-based Workstation

Ampere eMAG 64bit Arm Workstation

store.avantek.co.uk

Ampere workstation (CPU name)

Avantek's Arm Workstation: Ampere eMAG 8180 32-core Arm64 Review

www.anandtech.com

"What’s actually more important than the current generation eMAG Workstation is the possibility of Avantek and Ampere creating an updated successor based on the new Altra processor. " <-- from the conclusion.

Yeah, for most of us x86 is just going to be the answer for the next many years because the margins for it to exist for general computing are going to basically soak up all the silicon they can mint. That and it is a lot easier to get server side apps to be (nearly) hardware agnostic. You might still need some compilation or other software architecture tweaks to get things just right - x86 has been around for a long time and is like the jack of all trades.

New ARM CPUs with different architectural tradeoffs (lots of cache, less cache, tons of memory bandwidth, crazy PCIe lanes, etc) than what Intel/AMD are interested in selling in volume can fit niches where the right software will be able get demonstrably more perf/w than more general purpose hardware or even other ARM platforms.

Also, from my data center background, the abstraction from "this server in slot 7, rack 14, row 12, room 3" to "it runs on cluster 1 unless it's running on cluster 4" has helped separate developers from brand/architecture loyalty to some degree. People kept their servers like pets and had preferences for servers like other people have preferences for car brands.

That's still getting ground out of existence for on prem stuff, but when you are talking about Azure vs AWS, you've already flipped your mindset. It's about TCO. All of a sudden Graviton vs Intel is less about brand loyalty and showing which can sustain higher throughput for your particular app with lower hourly costs.

Anyway, that's a lot of words Mark to say until a DC project releases a project they recommend running on a certain ARM platform vs x86 I don't think it's worth trying to figure it out in your house.

 

[LightningZ71]

@Markfw if you're looking for the best ARM performance that you can get your hands on for a not obscene price (though, still high), the best you're going to do, at the moment, is Nvidia's Jetson Platform. The Xavier NX is a six core Arm V8.2 product built on a "modern 12nm" process. The larger, and twice as expensive ($699, as opposed to $399 for the NX) Xavier AGX also exists. It has 8 cores, and a higher TDP. Both feature rather large Volta GPU sections with a lot of CUDA cores for an embedded.

Enterprise Marketplace | NVIDIA - robotics-edge

Explore NVIDIA Enterprise Hardware, Software, Cloud Solutions, Data Center Offerings, RTX in Professional Workstations, Embedded Systems, and Networking

www.nvidia.com

I don't see anything else on the market for less than 5 figures that you will get better performance or thermals from. Also, if you can find a way to engage the gpu side of things, you might get interesting DC results from it.

 

[MrTeal]

You could apply for an N1 SDP too, though I'm not sure how selective they are.

Neoverse N1 System Development Platform (N1 SDP) Request – Arm®

The Neoverse N1 System Development Platform (N1 SDP) is an enterprise class reference board based on the Neoverse N1 core.

www.arm.com

 

[piokos]

There's no chance of that. If they were planning on dropping the Mac Pro they wouldn't have bothered to introduce the new one recently.

On the contrary.
1st gen Mac Pro lasted 7 years, 2nd gen: almost 6 years.
If 3rd gen survives until 2025 (and it's a fantastic design, with some proper upgrades possible - surely not something they planned for 2-3 years)... I don't think "workstations" will still be necessary. And who if not Apple to start such a revolution?
It's well known that Apple is expanding its DC infrastructure as we speak. ARM MacBooks will probably turn out to be relying on cloud for many tasks. It's even easier to do with desktops.

 

[blckgrffn]

@Markfw if you're looking for the best ARM performance that you can get your hands on for a not obscene price (though, still high), the best you're going to do, at the moment, is Nvidia's Jetson Platform. The Xavier NX is a six core Arm V8.2 product built on a "modern 12nm" process. The larger, and twice as expensive ($699, as opposed to $399 for the NX) Xavier AGX also exists. It has 8 cores, and a higher TDP. Both feature rather large Volta GPU sections with a lot of CUDA cores for an embedded.

Enterprise Marketplace | NVIDIA - robotics-edge

Explore NVIDIA Enterprise Hardware, Software, Cloud Solutions, Data Center Offerings, RTX in Professional Workstations, Embedded Systems, and Networking

www.nvidia.com

I don't see anything else on the market for less than 5 figures that you will get better performance or thermals from. Also, if you can find a way to engage the gpu side of things, you might get interesting DC results from it.

That AGX system is pretty beefy (32GB of LPDDR4X!), but a solid premium. 512 "cores" - is that cuda cores? So about 25% of a 2060?

But if it is a research project, that seems to be one of the most approachable options.

https://rockpi.org/rockpi4 <-- Even "upgrades" from the Pi like this tend to be IO improvements vs straight CPU upgrades. (still A72 based, same CPU offered from Pine in their similar boards)

 

[Markfw]

@Markfw if you're looking for the best ARM performance that you can get your hands on for a not obscene price (though, still high), the best you're going to do, at the moment, is Nvidia's Jetson Platform. The Xavier NX is a six core Arm V8.2 product built on a "modern 12nm" process. The larger, and twice as expensive ($699, as opposed to $399 for the NX) Xavier AGX also exists. It has 8 cores, and a higher TDP. Both feature rather large Volta GPU sections with a lot of CUDA cores for an embedded.

Enterprise Marketplace | NVIDIA - robotics-edge

Explore NVIDIA Enterprise Hardware, Software, Cloud Solutions, Data Center Offerings, RTX in Professional Workstations, Embedded Systems, and Networking

www.nvidia.com

I don't see anything else on the market for less than 5 figures that you will get better performance or thermals from. Also, if you can find a way to engage the gpu side of things, you might get interesting DC results from it.

First, the most useful post so far, thank you !

They have a $99 unit, but don't say anything about the CPU. Also, the $699 unit does not say what OS it can use. Do you know anything more ?

 

[Markfw]

You could apply for an N1 SDP too, though I'm not sure how selective they are.

Neoverse N1 System Development Platform (N1 SDP) Request – Arm®

The Neoverse N1 System Development Platform (N1 SDP) is an enterprise class reference board based on the Neoverse N1 core.

www.arm.com

I applied. No idea how this will end. Good constructive post.

 

[blckgrffn]

First, the most useful post so far, thank you !

They have a $99 unit, but don't say anything about the CPU. Also, the $699 unit does not say what OS it can use. Do you know anything more ?

This link beats the pants of the nvidia page for details:

Jetson AGX Xavier - eLinux.org

elinux.org

Looks like it runs a version of Linux supplied and updated by nvidia.

I started looking into that too but the the docs started to get pretty deep in a hurry.

Welcome — Jetson Linux<br/>Developer Guide 34.1 documentation

docs.nvidia.com

 

[Markfw]

octal-core NVIDIA Carmel ARMv8.2 CPU

For the ARM gurus, before I make a fool of myself, how recent is this core ? and on what process node ? It looks like 12nm. First tested Sept 2018, so should be more recent than Rome.

But how fast is it compared to some you have been mentioning here ?

 

[LightningZ71]

The two Xavier's, the AGX and cut down NX, can use Ubuntu 18.04 and have an upgrade path to 20.04 (extremetech had a look I believe). The CPUs, while performance, don't really run away from the pack as much as you'd think in day to day stuff. The real gem is the integrated CUDA and Tensor cores. If your DC program can use them under linux, these will probably score very well in output per watt. Of your project is very memory throughput sensitive, the AGX has more throughput than an RX 560.

Now, for the "rest" of the pack, the Odroid N2 has a quad core A73 core processor fabbed on a bulk 12nm node. It's got the best compute performance on the core out of the rest of the SBCs out there. The N2+, which now runs the A73 cores at 2.2Ghz, is still less than $89 for the kit. Judging by the benchmarks over at medium.com from earlier this year, it should be able to best the NX in a few specific sub-tests and, save for things that can take advantage of the large caches of the Xavier, not get totally beaten. Given the cost of the full dev kit for a Xavier NX, you should be able to get three or four N2+ units that should be able to out perform it for the money, though they may draw more watts per unit performance. The Mali gpu in the N2+ isn't horrible, though it doesn't hold a candle to the Volta's CUDA and Tensor capabilities.

 

[NTMBK]

octal-core NVIDIA Carmel ARMv8.2 CPU

For the ARM gurus, before I make a fool of myself, how recent is this core ? and on what process node ? It looks like 12nm. First tested Sept 2018, so should be more recent than Rome.

But how fast is it compared to some you have been mentioning here ?

It's Nvidia's custom CPU core. Anandtech tested it, and found it about on par with the A75 CPU: https://www.anandtech.com/show/13584/nvidia-xavier-agx-hands-on-carmel-and-more/3 But far behind the best ARM cores (i.e. the Apple CPUs). But it should be a big step up from your Raspberry Pi 4.

 

[LightningZ71]

Bascally, an undervolted gpu of appropriate spec will outperform any of the ARM processors on DC projects that were properly written for them. Their highly specialized nature makes them extremely efficient for those tasks. For things that can not run on a gpu, it's going to come down to code optimization. If they support AVX2, then I expect any of the x86 processors that support it to be very competitive. If they don't use the AvX units, or have a properly optimized and compiled ARM binary, with a modern ARM chip, then youcan probably get some very competitive solutions using the NX or the N2+.
It all depends on each project.

 

[LightningZ71]

If you want to try this, I would get one Xavier NX dev kit and test it thoroughly. That's a manageable outlay to use on something for testing. The Odroid N2+ is cheap enough to just play with.

 

[Markfw]

That AGX system is pretty beefy (32GB of LPDDR4X!), but a solid premium. 512 "cores" - is that cuda cores? So about 25% of a 2060?

But if it is a research project, that seems to be one of the most approachable options.

https://rockpi.org/rockpi4 <-- Even "upgrades" from the Pi like this tend to be IO improvements vs straight CPU upgrades. (still A72 based, same CPU offered from Pine in their similar boards)

All of those are OOS BTW