Raspberry Pi demolished by monster 128-core ARM CPU!
HTML-код
- Опубликовано: 20 май 2024
- The collab nobody knew we needed. I took my 6-node Raspberry Pi cluster to @ServeTheHomeVideo and faced it off against a 128-core monster ARM server.
Huge thanks to Micro Center for making this possible!
Shop Micro Center’s Top Deals: micro.center/1yz
Check out Micro Center’s PC Builder: micro.center/2vs
Submit your build to Micro Center’s Build Showcase: micro.center/6i9
Visit Micro Center’s Community Page: micro.center/dpq
Support me on Patreon: / geerlingguy
Sponsor me on GitHub: github.com/sponsors/geerlingguy
Merch: redshirtjeff.com
2nd Channel: / geerlingengineering
#Sponsored #ARM #RaspberryPi
Mentioned in this video:
- ServeTheHome (website): www.servethehome.com
- ServeTheHome (RUclips): / servethehomevideo
- The video Patrick has is the "Supermicro ARS-210ME-FNR", with the M128-30 CPU
- DeskPi Super6c [Amazon]: amzn.to/3AH8kNh
- DeskPi Super6c [Direct]: deskpi.com/collections/deskpi...
- DeskPi Super6c Mini ITX case [Amazon]: amzn.to/3XnJxHP
- DeskPi Super6c Mini ITX case [Direct]: deskpi.com/collections/deskpi...
- Guide for Raspberry Pi CM4 NVMe boot: www.jeffgeerling.com/blog/202...
- DeskPi Super6c Ceph Storage Cluster video: • 6-in-1: Build a 6-node...
- Top500 Ansible Automated Benchmark Playbook: github.com/geerlingguy/top500...
- Blog post with more benchmarking results: www.jeffgeerling.com/blog/202...
Contents:
00:00 - 24 cores - but not the same
00:52 - Apples to Apples
03:50 - Size comparison
04:52 - ARM in the cloud? Edge?
05:55 - ARM v ARM: top500 showdown
06:44 - New new bramble
07:38 - No time for microSD, we go NVMe
08:56 - Super6c, Super improved
10:27 - Linpack on the Pis
12:02 - Linpack on the Ampere Altra Max
12:46 - Price to performance
13:53 - Benchmarking methodology
14:56 - Is ARM the most efficient?
16:46 - GPUs: efficiency kings
17:48 - Platform maturity vs x86
19:38 - So what's the verdict?
20:39 - Surprise Stolen Supermicro 
Наука
What a super-fun collaboration exploring Arm server options! Thank you for stopping by Jeff!
It was so fun to do this video... maybe we can figure out another fun project for next year?
Hope you got back the Server he stole lol at the end we all saw you Jeff. lool 🤣🤣🤣
We need more of those!
Keep doing more, that was fun
@@JeffGeerling You could set up a mastodon instance on a pi cluster. See how it holds up:)
2021 Linus flexes doing videos on million dollar home, 2022 Jeff flexes showing off he ownes multiple Rasberry Pi units.
i make my pies in owen too lol
@@lordjaashin stupid auto correct
he has the last batch of RPIs while they were still made.
But the PI entered the halls of history, it was good while it lasted.
@@monad_tcp I think he is the reason for the shortage. He just uses them once and throws them away after each video lol
They that rare? I traded someone a pin zero w for like a pair of wire strippers and a 32v 20a power supply
Patrick and Jeff. In one video. Talking ARM servers and RPIs. Awesome
Truly excellent crossover episode!
Amazing that he was able to find 6 compute modules let alone regular Raspberry Pis
The secret is to order them in late 2020 😭
It just the RaspberryPi trend since the beginning. I remember the months I needed to get my hands on the first one...
Call this Pi server "unobtainium"
I can still get them rather easily, but at 3-4 times the price. It's insane.
@@JeffGeerling Now you tell us.
It seriously makes my day when you're feeling good enough to make a video and post it. Glad to see you upload again Jeff.
Thanks for that-this video will end up being a bit of a swan song for 2022, sadly... I have to get surgery, and though overall I'm feeling a bit better than summer, I've had complications enough that I have to get a pretty major surgery in a week that'll knock me out until next year :(
But hopefully after I recover, I'll be a lot better off! Maybe I'll be able to feel 100% again, it's been a while since I was feeling my best.
@@JeffGeerling Really hope you feel better soon and that the surgery helps.
You have done amazing work, and we all appreciate what you do.
In fact, you inspired me to setup a Raspberry Pi Compute Module 4 and I/O Board with native-boot NVMe, which is what I am writing this message to you on right now. :)
Thank you for all that you do Jeff, you have made the world a better place and are an inspiration to all!
@@JeffGeerling 🤞🏻that the surgery goes well, hopefully even better than expected 🤞🏻Wishing you can feel 100% again (it's been over 30 years since I had a day without pain so I don't really remember how 100% feels)
@enrique amaya coming from an atheist, not only does he live you, he loves everyone.
@3V0 person here, who types words that people read when in front of their face. I'm an atheism denier. I don't believe that people don't believe in skyman. You can't prove you exist, so I win. Wish maker-of-video man all the good things and speedy recovery. I like how you speak words out loud and I hear them. Subscribed.
When I saw this thumbnail I literally went to "The Collab nobody knew they needed" only to find it was the first line of the description. Great minds and all that...
Good stuff y'all!
Haha you know it!
That was a great collab. The scripting was fantastic and you guys interacted seamlessly.
I love both channels, looking forward to more colabs!
Great collaboration from 2 of my favorite tech channels.
Two of my favorite RUclipsrs! I'm glad to see what Patrick's studio looks like!
What an awesome vid guys! Not only did i learn new stuff, I couldn't be but happy to see you both school us on a badass project. One of the best vids I watched coming from your end. Thank you both!
Great job with this video.
A lot of information packed in here to work with.
Jeff nicking the server at the end is the cherry on the top. Great video guys, I enjoy both of your channels :)
This collab is great! Love it!
Loving the content, good Sir. Definitely the best collab of '22!
Owesome ! Thanks to both.
Hope your getting beter Jef ,
good to see a colab with you and Patrik :D
Loved this video. Will love a follow up on the LTT screw driver even more!
Hey, this is great! Y'all make great crossovers.
Have been a fan of STH for almost 10 years now... Has always been a great website, and the forum is great as well. The hardware sale section is an amazing place to find deals on server grade hardware that is coming out of production if your looking to setup a good home server cluster.
This is an amazing with Patrick. Thank you both
Woah, I'm so interested to see the results!
love love love the collab!! More please!
This is monstrous. Thank you for the learning, Jeff and Patrick, too! Cheers!
I seriously love ARM so much, so this got me really excited!! Great video :D I've waited years and years for someone to fully push out ARM to the mainstream desktop for "regular" consumer usage, and while the Pi's + others are capable, nothing was ever the same as your usual x86-64 for typical tasks / light gaming. Well, when Apple announced their M series of chips, I got excited and purchased one right away after reviews! Been happily using my M1 Max as a daily driver, all because I love ARM and really want to support it as much as I can, even if it's coming from Apple, who doesn't truly *need* the money but ended up making a product that I had been looking for for quite some time.
I'm just happy that a major company took the plunge, and I'm happily following suit. Maybe someday we'll all be on ARM!
not for much longer. the way arm is going it is killing itself by spitting in its customers face by forcing major companies to use its gpu along with its cpu. after this many companies are looking at risc v. hopefully, risc v takes off as it is open source
For what it's worth, afaik, the big selling point of ARM cpu designs was in power savings not so much raw performance. -- Even if a particular design/SoC performed at a mere fraction(say 50%) of an Intel x86/x86-64 system the former would kill in terms of power savings. // Things have improved a lot and now that even Intel chips are SoC-like...
@@jnharton Performance as well. At the socket level Ampere M128-30 matches or exceeds top bin Epyc Milan for Data Centre workloads whilst burning less energy. Anandtech and STH have published stats.
@@vikingforties I was really talking about the broader history and not so much the current state of affairs. Also, given that x86 development was, for a long time, presumably based on single core, single thread performance and ARM is a relatively new architecture...
@@jnharton Ahh. got you. I take your point.
Wow, I cannot believe how compact the STH studio is, it really stood out when both were in there
Fantastic video Jeff, hope to see you guys collab again!
Great video Jeff 👍👍
I might pick few tips to build my first PI cluster.
Serve the "large scale, high performance data center" home. Great collab guys.
Two of my favorite channels in one video. This awesome!!
I like those collaborations, it gives more insights and different views on the topics we love.
You guys are both great! Fun to see you together.
Thank you for the great video, Jeff! I would like to inquire, if there are any current alternatives to RPi that are capable of running Xen?
I love the collab with STH! Keep up the great content! I have been running my own "server" (TS3) since the early 2010's. Never realized that was homelab-ish territory. Now I run a bunch of servers off of RasPi's, Local PC's, and a NAS. Totally Cosplaying as a SysAdmin. Want to do more in the future and you are helping me do that!
can you imagine a raspberry pi with that monster 128 core arm processor in it?🤣
love the collaboration, hope u doing well man. love the videos
Great video and I love the direction that ARM is going!!
yes, the collab we never knew we needed. good thing red shirt jeff must be busy somewhere else.
Just don't watch to the end 🫣
@@JeffGeerling i commented before finishing the whole thing... oh no....
@@n0madfernan257 We are still trying to find the Ampere server
@@ServeTheHomeVideo Hope he appears secretly at your channel and still return it in one piece... and hope red shirt jeff does not use his favorite grinders and whatnots
Brothers in ARM's
I just made the connexion between you and the fantastic ansible roles everyone use.. THANKS A LOT FOR THAT !
I loved the transition you made to the STH clip :D
I love the labels on your fan drawers.
We should compare Wh for fixed unit of work. E.g. the same compilation/render and just look at the power consumed to do the work.
That's what GFLOPs/W and MIPS/W benchmarks represent on a standard scale that should convert to other loads, provided everyone uses the same benchmarks without cheating. Of cause, additional numbers are needed for idle and near idle power consumption.
Awesome! Thanks man!
Another great video Jeff!
I was already impressed by your 6-pack Pi. Time for a new T-shirt, "My other computer is a 128 core Ampere"? :)
Hey it's Patrick STH! He always has the coolest and latest "toys" to play with. Nice comparison guys!
RE: first bit about 24 RPi cores vs 24 x86_64 cores. A, say, 1GHz ARM and 1GHz x86 are VERY different. ARM will always win on the power consumption front, but even if using a reduced instruction set(IIRC) the x86 will walk all over an ARM twice before it even realizes what happened, and a few more times before it hits the ground. Yes, application matters. Programming matters. Auxiliary processing load matters. But straight clock for clock ARM and x86 are not comparable, its x86 on top all day long. You cannot replace a x86 desktop with an ARM and expect the same performance. But that was never what they were supposed to do, its not their design case. Remember, the right tool for the right application ;)
I've been trying to get a review sample of a ARM based server for a while and it's great to see this one here!
Seeing a video with this two makes me very happy, great guys
x86 servers also have a mechanism for attesting the boot process that simply doesn't work right now with an arm server.
You mean the lockdown to Microsoft signed boot images?
18:09 You should try out Armbian. Their support of older arm boards is phenomenal. My tiny cluster of 4 Orange Pi pc's form ~2015 is still running perfectly with kernel 5.15 thanks to Armbian.
seconded, armbian is pretty great
Love Love LOVE my MicroCenter in St Louis Park. Excellent sponsorship! Beats the fiasco of Kamikoto knives and Scottish Titles (what a scam those were).
i was waiting for this video :) im so excited for all the amazon stuff to trickle into our home labs!! Pat and Jeffy G!!! what a team combo!!
What. CM4 modules cost now 165$. This is insane.
Great video Jeff.
Also, I am bit on the edge about getting DeskPi Super6C. It only has 2x1Gbit uplinks, and even utilising two of them is tricky due to the type of switch chip used. If it would had 10Gbps, or maybe even 2.5Gbps, it would be an interesting option. Also lack of some BMC to access serial console and power state of CMs is not great. They could have added a small stm32 micro with a 100Mbps ethernet to just do some telnet to interact with few things on board and serial ports, and maybe small http server to view some minimal metrics, it would be useful. And where is RTC battery. I can live without it probably for cluster setup where I would be using NTP, but come on.
I think the Turing Pi 2 is probably the more flexible option, and hopefully their production run will get underway soon (they just said they're tooling the line for a first 1,000 unit run).
in case you think of using one of them for virtualization it would be nice to have the performance per core. that is because next year there will be a shift in licencing from cpu to core (read VMware here).
@Greyson Sounds like a reason to use qemu.
@Greyson thanks for the comment, yes but, if you own a million dollar company you want to stay on the safe side and then you buy support from the vendor (read RedHat or Suse). if you get a problem installing a SW (SAP, Dynamics, VDI) and you have problems to whom are you asking, do you have enough people in IT to go through the opensource code and find a solution....
Being technology enthusiastic, having an home lab is completely different in managing company/corporate infrastructure, you want reliability and support. if you look around many applications are born as opensource but there is always a "pro" version, for the support.
@@g.s.3389 Just to be picky, you miss stated your example. RedHat and Suse are opensource. Opensource and free are not the same. I get what you mean though. I have 2 sides in my DC. The internal side runs on things like docker, haproxy, a lot of Ubuntu for simplicity. The client side runs on kubernetes (openshift), F5, and RedHat for added support. I have no issues supporting either, but when you are dealing with things in the enterprise, they want to have a company support to be able to contact for some types of things.
@@skechergn It depends on the pricing. If the core pricing ends up equaling what the CPU pricing would be at a certain point, it is just targeting the new processor setups. There is a large difference in the amount of vms you can run on 24 cores/512GB ram and 140 cores/2TB ram.
Especially when you start looking at containers on VMs on a hypervisor. Containerization is really hammering VM virtualization. I have a pair of IBM power servers (for redundancy) running Openshift. It's running the same workload that use to sit on a 12 system ESXi cluster because of the number of different VMs it needed and the number of servers to make it n+1 for the hardware. Those 2 power servers still have plenty of resources to add a ton of additional containers to it before it would even comes close to high utilization on one node in a failed scenario.
I can see why VMware would need to go to the core level overall to not only make more money on the higher end server chips but also allow them to be price competitive on smaller core numbers. If the price per CPU is based around say what will now cost for 40 or 64 more cores, then if you come in needing 24 or 36 cores, then it would be cheaper then the old CPU cost.
@@skechergn Your still assuming it will be a higher (then inflation) cost with the new licensing. We can't say it will or won't until we see it put in place across the board. It could be a way to enable charging more across the board and it wouldn't surprise me. But it could also be a way to make the costs more cost effective also. If it becomes cheaper on the lower end and more expensive on the higher end, then they are really just trying to be cost effective against competition.
Again, I am not saying it will work out that way, I am just saying we are assuming that it will raise prices (beyond inflation) for customers without any evidence yet.
Thanks for your sharing
Great job with this video
Two guys that I don't watch at 2X speed. Great pairing, great content!!
Thank you for the manual captions and timestamps.
One of the best collab intros i have watched so far... :)
Have you been served?...in the home?
By Grabthar's Hammer...Cool Content!
Yay! Both of my favorite channels!
I wish Patrick greeted Jeffrey in third person "This is Patrick"
I was in the market for an ARM VPS but couldn't find any good ones. Honestly I am so ready for ARM to just take over the server space more and more - its super efficient, less cluttered than x86 and has some great nicities. Unfortunately, software support is still a little hit or miss - especially when talking about Phones and Tablets where SystemReady isn't implemented properly - be it on purpose or not. Honestly I'd love to play around with one of those! These ARM servers are awesome as heck! ^_^
Try Oracle. They give 4 core ARM Amperes for free.
AWS EC2 m6g, uses AWS graviton.
I'm using Graviton, as Arm 64 makes sense to match my M1 Max laptop. However, AMD Milan is way more efficient than any Arm server chip up to this point, you don't get much in terms of efficiency if you are on AMD, right now.
@@GustavoNoronha Epyc is crazy, it's amazing how they just leapfrogged everyone else.
@IngwiePhoenix Did you not pay attention? Literally said in the video 16:28 that ARM isn't inherently more efficient than x86. In fact, the lack of libraries and application support can make it LESS efficient. What apple did with the M1 came down to building an ARM chip made specifically for OS X, down to the point where it theoretically should be good for gaming, but isn't since apple doesn't put as much thought or support behind it. Also, apple decided efficiency at all costs, to the point where the damn thing throttles on certain tasks to keep the battery life well and the cpu cool.
both of these guys are in their element. I can see their enthusiasm and excitement
why doesn't patrick have more questions about how you got into his recording studio?
/s
Jeff is always welcome.
Besides of being an incredible interesting subject, i love how you pulled a page out of the influencers that contaminate RUclips and the media with Nvidia GPUs visible just because (LTT, TechSpot and others) with your strategic placement of AMD CPUs.
Here is a like going your way!
The subtitles are the best. "red shirt jeff acts suspicious" hahah love it!
I remember years ago Top Gear did a race between a Prius and an M3. All the M3 had to do was keep up with the Prius, while the Prius thrashed it's way around the track.
At the end they measured how much fuel was used and the V8 M3 was significantly more efficient at the speed they were driving around the track.
I think the pi is great as a testbed platform, low power edge deployment, etc. But it never really stood any chance when real compute capacity and io connectivity is required.
Those linpack benchmark results were super interesting, didn't know m1 max chip was that far ahead. I wonder how long will it take for competitors building arm64 processors to catch up
Comparing different things here. An M1 client CPU to CPUs designed to operate in a DC (e.g. Ampere Altra & Graviton). Data Centre and cloud workloads tend to force the economics for your compute & I/O efficiency per rack of servers and storage. M1 still needs a focus on per core performance because there's way more single threaded stuff that it has to deal with.
An other indicator how much 'node size' matters. A they mentioned in the video: the M1 uses a newer TSMC process.
Does the Super6c require all 6 compute modules to be installed, or can you have some unpopulated without damaging anything?
I have 5 pi in mine... Works fine. I can't wait to find one more to fill it up. Note that only the first one has extra ports...
^ This
Very cool and inspiring, the idea of the pi concept has grown. I like how it compares to other similar products, I should have kept mine. But I have real need for real applications. No time to play and learn unnecessary studies that I already know without a computers input.
Yes my cousin invested the computer chip, but there is more to this story then what is shown, but, its when, it's truth is seen, that's when it all changes.
That was a great collab!
Let us hope Red Shirt Jeff does not use an angle grinder for that end clip....
Oracle Cloud has some really generous "always free" quotas for VMs on their Ampere machines.
Search for "OCI Ampere A1". 4 vCPUs, 24GB RAM, multiple distros, adequate storage and lots of regions. What's not to like?
Such an awesome collab :)
Wow, two of my favorite tech chans together, awesome!
Jeffy G?
Apparently I'm Jeffy G and Jeff from Craft Computing is Crafty Jeff XD
That means you’re ….not crafty. 😗
11:30 - seems jeff is into only fans :D
🤣 ikr
Very cool. Major improvement.
Very cool, great video
Don't forget that I/O bandwidth is next to nothing on the PI and the real server probably has excessive lanes. Don't feel bad though, even desktops have crippled I/O compared to what they should have.
Yeah, like my desktop Ryzen CPU gets crippled by x1 and x4 PCIe 3.0 slots even though the CPU has plentiful PCIe 4.0 available :(
You just barge into his studio when he's recording?
This is the way.
Nice video! I really do hope that the Pi is available soon. I would love to put the CM to use.
That's some pretty crazy stuff! I have a couple Raspberry Pi 3s running Stretch. They weren't cutting it for me so I just decided to drop $$$ for an Apple M1 Mac Mini with the 16GB of RAM and base everything else. External SSD is a lot cheaper although slower. The plan is to run it headless with a direct 1GB/s Ether connect to my router which goes to my cable modem. The router gives my house WiFi. It's kinda old, but I'm not sure what to update it with yet.
That's my lazy solution to a local server. It's double plus cool to see that there are still hobbyists out there building custom computers. It's like the 1970s with 2020s tech. Good stuff.
At 0:30 , 3960x 24-core is old technology.
A modern 16-core Ryzen 7950x offers greater performance, for just $550 USD.
Might be useful if raspberry pis were actually obtainable one of these years
Nice benchmark comp Jeff! Rasp Pi's have become like gold now - amazing. That ARM system is a beast. He who dies with the most teraflops, wins!😂😂
Maker center does that bring me back reminds me of my youth going to radio shack to pick up components to put a well anything together.
Atlas might not be the best blas library. There are others which may give significantly better performance, including openblas and the libraries bundled with official arm development compiler suite. You can also usually win a bit more performance by experimenting with the relevant compiler flags.
If the point is /benchmarking/, then you use the same program on the different hardware configurations because you want a comparative sense for which is better. Otherwise you use whatever can perform /best/ on the hardware you have, as you are suggesting.
@@jnharton The linpack test does use the same source code on each system. Blas is a standard library, so does the same thing on every system. Clearly, you need to compile and link the code in the way that gives the best performance for the target hardware.
Microcenter is legit. I stopped in for the first time in ages this summer. Not only did they HAVE the 8gb Pi4, it was on sale for $60.
rpi available and on sale, wow
I would really like to see you cluster multiple of these together, imagine the power of 4-8 of these
I've been around for many years in the computer world (granted more on the software side, but built my fair share), so I was pretty surprised to hear some terms I have never encountered before! :)
What really concerns me, the realization that red shirt Jeff is not confined to the Geerling basement. We don't need that harbinger of chaos roaming free to spread this madness!
So I really enjoy the large box labeled "Only Fans" 🤣. Also it's pretty surprising your efficiency went up when overclocking, apparently the opposite usually happens on desktop pc's. Wonder if that means there's more headroom in the silicon?
Naked processor of the day, for only $20/month.😂😂
Watching both channels, always useful in a different areas. Would enjoy a further cooperation with Patrick. Well done.
Double trouble and doubly awesome! More!
you could also get dual xeon for 50$ and have 20 core beating those shitty rpi
What a savings.... LOL love it! Just bought the motherboard and trying to get the CPUs
I've had an order in with DigiKey for a month or so... hopefully they come through!
Loved this reference
YEAHHHH!!! THIS IS AWESOME !!!!