This insane 80-Core ARM CPU easily beat a 64 Core Threadripper
HTML-код
- Опубликовано: 19 май 2024
- More about Hetzner RX Servers:
derbauer.hetzner.com/en
---------------------------------------------------------
Support me on Patreon:
/ der8auer
---------------------------------------------------------
Save 10% on your iFixit purchase: DER8AUER10
eustore.ifixit.com/der8auer
Pro Tech Toolkit: bit.ly/2JOFD8f
---------------------------------------------------------
Find my products at Caseking:
Delid Die Mate 2: bit.ly/2Rhv4y7
Delid Die Mate X: bit.ly/2EYLwwG
Skylake-X Direct Die Frame: bit.ly/2GW6yyC
9th Gen OC Frame: bit.ly/2UVSubi
Debug-LED: bit.ly/2QVUEt0
Pretested CPUs: bit.ly/2Aqhf6y
---------------------------------------------------------
My Equipment:
USB- Microscope*: amzn.to/2Vi4dky
My Camera*: amzn.to/2BN4h2O
(*Affiliate Links: If you buy something, I will get a part of the profit)
---------------------------------------------------------
Music / Credits:
Outro:
Dylan Sitts feat. HDBeenDope - For The Record (Dylan Sitts Remix)
---------------------------------------------------------
Paid content in this video:
- Ad, at the end of the video to promote RX line
Samples used in this video:
- /
---------------------------------------------------------
Timestamps:
0:00 Intro
0:22 Ampere Altra ARM CPU
0:57 Server motherboard & CPU installation
2:36 CPU cooler, TDP & Info about Altra-CPUs
4:07 ARM v8 vs x86
6:03 Power draw & Testsystem
8:14 Geekbench & Efficiency
10:54 Ampere Altra delidded
12:50 Die size & Package
14:13 Conclusion
15:04 ARM RX-Server from Hetzner
15:46 Outro 
Наука
This man is a legend he does every video in 2 languages which is almost double the work
thanks :D
@@der8auer-en Bro how do you manage that? Just shoot the same scene twice? Or make a voiceover in German afterwards?
@@MarioGoatse if your watch carefully you can see that the English version is overdubbed
@@DavidUnderhill no
@@DavidUnderhill Lmao.
It is cool to see, that ARM CPUs / SOCs can be used for other than smart phones and small single board computers.
Get used to it, it's the future.
@@BroBytesBhai Yes, I also hope that RISC-V will be there.
The previous champion of the Top500 supercomputer is powered by ARM CPUs
@@BroBytesBhai
I'm fed up with nISC monopoly
Give me high performance VLIW for mainstream users!
@@borisvokladski5844
And i hope VLIW is coming back to mainstream computer, especially due to improved programming and so
I'm currently interning at Ampere in logic design (writing RTL code for an upcoming chip as we speak, my little brother sent me this vid). I mean it when I say that this is the smartest group of engineers I've ever been around and the caliber of innovation we are achieving is really exciting to be a part of. Thanks a bunch for brilliantly showcasing the hard work of my coworkers!
All the best, I hope Ampere beats X86 and makes ARM based software more popular.
Happy for you man, it’s intoxicating working with brilliant people. Hang on to that internship!
@@prashanthb6521 Actually in terms of ecosystem, ARM still has a long way to go. x86 is still very dominant software wise.
And Geekbench has always been pretty ARM favored tbh, in other application, you wouldn't really see that much of a lead. ARM is especially weak in large vector space matrix multiplication. ARM v8 solve that but it also made the instruction much more complicated.
These Ampere CPUs are excellent fit for cloud server use. They simplifies the in-core structure as much as possible so that each core takes up very little die area, while enhancing a few certain functionalities such as int8 add/sub and some scalar calculation capabilities. That's why they're almost unmatched in terms of VM applications. But if you're building a HPC server, these chips probably aren't the choice.
@@F1ve-seveN well said and accurate
It's my hope that one day ARM cpu's will be the standard in all our devices. Intel and AMD both hold license's, it would be interesting to see their attempts with todays technology. The rate we are going with GPU power consumption, we need all the power reduction we can get anywhere we can get it, without sacrificing performance.
1) ISA is ARMv8.2 ... 64-bit high performance RISC with very advanced loose memory model. Designed in 2010 at clean sheet for high IPC
2) ARMv8 is not binary compatible with 32-bit ARMv7 - using different encodings, backward compatibility via both decoders (allows to remove 32-bit and save transistors)
3) ARMv8 has 32 general purpose registers and 32 FPU/SIMD registers (compare to 16 / 16 for x86)
4) CPU core used in Ampere Altra is license core Neoverse N1 derived from Cortex A76 (1st desktop class core from 2018, IPC between AMD Zen1 and Zen2/Skylake)
5) Amazon has 64-core Graviton 3 CPU (5nm TSMC) based on much powerful next gen Neoverse V1 from 2021 ..... IPC +42% over this N1, or IPC +15% over AMD Zen3 (yeah, even Zen4 will have lower IPC than today's Graviton 3)
6) This year's Cortex X3 with 8-wide integer core (6xALUs + 2xBranch units is similar to Apple M1 8-wide design) will provide about +30% IPC over AMD Zen3 (4xALUs + 1xBranch Unit = 5-wide design similar to Intel's Golden Cove)
7) ARM has new ARMv9 ISA which is supported by X2 and X3 cores (this cores has removed 32-bit mode, saved transistors for 32-bit decoder, all invested into pure 64-bit power).
8) ARMv9 ISA supports new SIMD vector instructions SVE2 (scalable vectors) and SME (scalable Matrix) math
9) SVE2 / SME supports up to 2048-bit registers/vector length (x86 has 256-bit AVX2 and for a decade tries to move to 512-bit AVX512)
10) SVE2 / SME is length agnostic - CPU has to support 128-bit up to 2048-bit (with 128-bit step) while SW compiled for SVE2/SME support SIZELESS type (for autovectorizing an arbitrary long vectors - pretty smart). SW will chop vectors according to available HW resources in CPU, dynamically on demand. x86 has problem to adopt AVX512 because it's too beefy for laptops or e-cores (where 128-bit or 256-bit version would be fine). This caused slow AVX512 adoption and ISA fragmentation.
11) Idea of scalable vectors came from Japanese Fujitsu - they designed ARM CPU with SVE for super computer Fugaku. The most powerful SC on the world in 2021. Only CPU based, no GPU and yet beating other Nvidia GPU based super computers. Pretty amazing how those SVE vectors are powerful. Before Fugaku I was convinced that GPU cannot be beaten by CPU only machine at similar cost and power consumption. Obviously it's possible.
Pretty good stuff from ARM in ISA features and also in cpu design. Upcoming Neoverse V2 core based on X3 (ARMv9+SVE2 ISA) will be absolute x86 killer.
Great info !
@@ShrKhAan 1st 5nm server CPU is 64-core ARM Graviton 3 .... all that half year before AMD Zen4.
This means that Graviton 3 must be made in HUGE numbers otherwise TSMC wouldn't give them exclusivity before AMD, Nvidia etc.
When Amazon deployed 7nm Graviton 2 they stopped buying x86 server CPUs. AWS has 33% cloud global market so this is HUGE. Graviton 2 made about 10% global server market. It looks like Graviton 3 aims at much higher numbers.
Graviton 3 can cost approximately $500 while is better than $5000 AMD Epyc. The math is pretty simple why ARM grow in huge numbers. This is the power of open ARM world where anybody can buy a license and manufacture his own CPU. Outdated monopolistic proprietary x86 cannot win this battle. Even free license for x86 ISA will not help. X86 ISA is old, complicated, for high IPC you need many workarouds which cost transistors and kill efficiency. It doesn't make sense to start developing new x86 core even if that would be possible for new comers. That's why China killed x86 collaboration with AMD and started their own RISC-based MIPS64-derived ISA called LoongSon. RISC is the future. It's always been.
@@arm-power that's cool and all but while we can find ARM CPUs in smartphones, notebooks, tablets and now servers, we still can't have an ARM CPU that's meant for workstations. As someone who works with their PC and is currently making a Ryzen Threadripper build it sucks to be forced to invest money on an already obsolete product.
What would you deem a reasonable timeframe for ARM CPUs to become available in the professional/prosumer segment?
Before Fugaku I was convinced that GPU cannot be beaten by CPU only machine at similar cost and power consumption. Obviously it's possible.... DOUBT
In regards to how small that die is considering how many cores it has - keep in mind those are non-SMT cores. They will have half the execution units as the 2-way SMT cores we are used to in x86 land and will be physically smaller.
I'm not up on the latest arch, but in principle an SMT core only has one execution unit/pipeline, shared by 2 sets of registers, and the register banks are not a dominant part of the core like they used to be in simpler times.
@@MattOGormanSmith Pretty sure modern x86/x64 CPU cores have multiple execution units per core which is why SMT provides such a boost to performance. I remember AMD going short in pipeline and wide in execution units in their cores back in the early Athlon64 days and Intel following suit a few years later with their Core architecture.
@@MattOGormanSmith S in SMT stands for simultaneous, so there needs to be at least as many execution units as threads. There are other forms of MT though. There are likely multiple execution units per thread, so it may not necessarily be half of an amd64 core but probably less.
What matters to data centres is performance per watt. If 80 single thread ARM V8 cores can do more work for a given number of watts than a 64 core x86/x64 chip. That is less power/cooling/space in the data centre.
@@patdbean Data centres care about performance density first and foremost. The more performance you can pack into a given space the more money you can make from that given space.
The face you pulled when the delid happened, I was pulling the same face thinking you'd absolutely ruined it. Nice job again Der8auer!
Awesome video as always! I would love to see the temperature and power draw of this CPU with liquid metal applied!
*THANK YOU!!!!* for putting the delidding part into the same video instead of milking it for all the money possible by doing a cliffhanger and then a second video with just the delidding. It is so incredibly much appreciated that you do not split it up into two parts and force us to wait for the epicness to commence! Instead you immediately give us the delidding part of the video, thank you so much for this dawg! Is super mega incredibly appreciated, thank you 👍👍😁
Thanks. 😁💪
Of course I can't speak for Roman, but I think most people realize that Roman doesn't do this for the money, but because he loves what he is doing, at the same time is the whole cooling theme of the channel a great way to spread the knowledge and information about Thermal Grizzly.
Very impressive seeing some of this high end server equipment etc. what your home pc has pales in comparison to what these servers are doing day in day out. Great video to show Der8auer 🥰👍. Love these behind the scenes look at server farms and what is being managed by these massive setups.
The noise reduction in the mic mixing is really impressive! Well done editor.
I absolutely love seeing hands on with enterprise gear!
Also incredible how small 7nm ARM cores are compared to 14nm x86
ssshhhh your making those 14 nm cores jealous
even the x86 cores are relatively small, it's the cache that takes up most of the space
@@Petr75661 do arms chips have any cache? I know x86 is loaded with cache but not into arms specs
@@raven4k998 @petr75661 It can vary from chip to chip but they generally have a lot of cache. The Altra seems to have 128KB L1 and 1MB L2 cache per core. An Apple M1 core has 320KB L1 and 3MB L2 per P-core. Meanwhile an Alder Lake P-core gets 80KB L1 and 1.25MB L2. L3 usually depends on the CPU model, but I’m pretty sure AMD has everyone beat there
"14nm" x86 would be closer to 10nm if you used the same system that TSMC uses to measuring transistors. As someone else said, most of that space is cache and SA.
Roman, I absolutely love how you delid everything you can.
This videos are the best. Truly.
nice! my brother actually works for Ampere and they seem like they're an insanely talented team.
That's a really interesting CPU, one more wedge in ARM's slow datacenter march. Thanks for showing it off and especially the delidding!
You need to switch form RPi to this ^^
@@guy_autordie I would love to find a way to get an Ampere-powered server in my rack!
Pretty big wedge, Ampere's in Azure, Google, HPE, OCI and the Chinese hypers.
@@JeffGeerling Funny you should mention that. Will contact you offline Jeff.
Where else can we see a 4000 euro CPU delidded using a bench vice?! Awesome to see enterprise gear - the numbers are insane, and fascinating to see such power in such tight spaces. derBauer has the coolest job around. Sheik approved, too!
4000 EUROS?!?! WOW.😱
@Godot yeah that's true...
It was horrifying to see the way he de-lidded it though. This guy has no respect for precision engineered parallels.
@@godot3422 still 4000 bucks tho
Im personally waiting for ARM dedicated servers to come to the budget market. Right now I have a server with Kimsufi - a OVH redist - for 18€/month. Its a 2nd Gen Intel CPU and the only reason I keep it is the storage XD So hopfuly cheap ARM CPUs will help the lowend a little! ^^
That’s a super interesting IHS and heatsink design. I never expected the SOC to be recessed like that.
If you consider problems that 12th gen intel has with cpu bending, it makes plenty of sense. Apply force to the thick part of the IHS and reduce deflection.
im hoping we will be able to get hands on these type of processors for home servers in 5-10 years, cant wait!
This was exactly my thought.
Nice - I work for Hezner South AFrica, now Xneelo, and it is good to see this - thanks :)
Awesome video, very interesting to learn a bit about ARM processors.
This was quite interesting to see because we're seeing a huge shift towards arm CPUs so it was interesting to see how it would play out in the data center
Playing out quite well, Azure, Google Cloud & Oracle plus the Chinese hypers all have @AmpereComputing offerings.
@Godot Yes! I have to imagine Arm's efficiency seriously piqued interest in the data center.
If I remember correctly, the reason arm is so efficient is because it doesn't support older instruction sets like x86_64 is expected to. X86 is great because you can buy an Intel processor today, a brand new raptor lake chip and fire up DOS like it was running on a 486. Very neat for automation or manufacturing where the software or hardware hasn't received updates in years, or the user decided not to upgrade for one reason or another. Either way that same user can again, buy the latest Intel chip, hook it up via serial to his 25 year old whatever-it-is and it just works.
What I think would be a welcome solution is a new processor built on old instruction sets and architectures, manufactured in a moderately sized amount.
@Godot Thing is the Enterprise market demands it and that is probably the bread and butter of most of the company's sales. Its fairly easy to tell a home user that "no your photo viewing program from 2001 will no longer work". Unfortunately you cannot tell Mega Big Corp "Sorry your billing system from 1993 will no longer work" because they are a huge customer and their CEO calls your CEO and you get told to keep their stuff compatible.
Fantastic stuff - will you be replacing the thermal paste with Grizzly and retesting so they can see how much better it is? Possible diversification in sales there ;).
Why stop there? Liquid metal cooled 80 core cpu!
@@craigmohr Why stop there? Liquid nitrogen cooled 80 core cpu!
@@pearce05 I mean, it's only $4000 CPU, why not just smash it with a hammer in slow-mo ? :-P
@@guestimator121 or send it to the waterjet channel?
Very cool! Thanks for showing this impressive piece of engineering.
Very interesting CPU, great overview!
It'll be really cool to see this trickle into the consumer space, especially with Windows on ARM. I'm excited mainly about the better power efficiency.
you see what is possible when you cut the cpu cache you can add more cores
I use ARM for software development on the Mac Mini for my job. It runs absolutely buttery smooth compared to my top-of-the-line x86 Intel laptop. Doesnt help that Windows sucks ass, because there are a lot of platforms that would run better with ARM and Unix-like systems over windows and x86
@@merthyr1831 arm fanboy
@@merthyr1831 When people run Windows on ARM in a VM on M1 Macs, it runs better and smoother than it does on actual WoA hardware like the Surface Pro X SQ2. To me this says that Windows isn't the problem, Qualcomm is.
I miss my Acorn Risc Machine (ARM) based desktop computer. ((every video like this gets loads comments about ARM being mobile only but it never was))
I had an Acorn A3010 with a single core 33mhz 32bit ARM cpu with 1mb of ram and a 3.5" floppy from 96-98
That thing was a beast in the early 90's and still pretty awesome when i had it.
We've come a long way since then but no one seems to remember that ARM was desktop first (mobile came much later)
Arm became pretty much exclusively mobile and embedded as the only places you would find them for quite a while though, so for many it is all they have ever heard of or seen using ARM - makes sense so many folks wouldn't know ARM as being anything but Phones and Pi's...
I really would like to get my hands on such a powerhouse ARM system, or a RISKV system of similarly high end specs, though last I checked RISKV wasn't really ready for that sort of use yet.
The good old days, I had an acorn machine as well. As did my school at the time. Looks like ARM based CPUs are going to make a comeback for personal computing maybe? Could be interesting.
Arm can't mulitask.
Source: every phone, ever.
@@ZeroHourProductions407 Source: a stupid statement made by someone that has 0 experience with an ARM Pc
@@ZeroHourProductions407 Arm can multitask perfectly well - for instance I run a Raspberry Pi4 as my go to always on desktop, it does basically everything my bigger computer can do*, just more energy efficiently and that includes having many things running at once. Don't confuse the hardware with the limitations of the operating system you happen to be using!
Either AMD and INTEL can't multitask either, or Arm can - as they fundamentally function very very similarly when it comes to multiple processes at once - none of them are the dedicated (or FPGA) massively parallel in real time logic systems that you could truly consider real multitasking at the hardware level.
*just moderate your expectation of how fast it can do them all, and avoid anything that needs more than the ram capacity - for most things its about as fast as any other computer, even really expensive powerful ones as the limitations are you the user not the system anyway, but it is a tiny cheap SBC, not a supercomputer...
Thank you very much for taking the time to make the videos a second time, but in English.
Very nice reporting! Thank you! I subscribed immediately. Also, that chip is about the size of my face! Alles guten wünsche. Danke. P.S. Your cat is beautiful.
Literally any benchmark other than Geek Bench would be more useful - how long does it take to compile Firefox. How long does it take to render a scene. Geek Bench just tells us that a fake benchmark notorious for favoring ARM ... favors ARM.
Sssh, that would be against the narrative. ARM is the best and beats everything else in everything - no questions asked. Stop asking for benchmarks or proofs of this.
fAkE bEnChMaRk because it tends to run better on one architecture.
Anandtech and phoronix have articles, and benchmarks about the Q80 and Max (128c). Q80 usually doesn't BEAT a 64c Epyc, but is on par, trading blows
I'd post a link, but, apparently, only bots are allowed to post links on YT smh
I was thinking "man I wish I could see the die on this thing" when you tell me you are going to delid it. WHAT A TREAT!
Not that this will ever been seen, but if you have a lot of time to wait and are delidding multiple CPU's I've found that soaking them in acetone or gasoline to be fairly effective in softening the silicone glue that holds the IHS to the PCB. Once softened the glue not only cuts easier and cleans from the PCB easier but in some cases you only need to cut around 3/4 of the glue and it'll just open like a hinge. The glue may even be soft enough that the IHS could be pushed to the side with no cutting, but I haven't tried that. I soaked an old G1820 in a 50/50 blend of gasoline and acetone for about a week and the IHS came off very very easily.
I love this channel. Sooo many interesting vids
The radeon 7970 had a sunk chip too, though it's not hand replaceable, so maybe it doesn't count.
The i7 8709G also has a similar thing with the igpu being lower than the cpu (or the opposite, don't remember)
4 years ago we had the Cavium Thunder X2 and the Qualcomm Centriq, both were powerful ARM v8 chips that beat the best x86 silicon. Sadly neither made much impact on the market and went away.
Woah great video. Amazing they let you decide it.
cant wait to use it in my setup 10 years from now
De8auer you forgot the Hetzner sponsor info in the description! It’s just “xxxx”
fixed :) thanks
@@der8auer-en no problem 👍
I wonder if AWS would be open to doing some videos with you? Maybe at AWS Frankfurt? It would be amazing to see Graviton 2 and 3 up close 🤩
Graviton 3 is an incredible chip. The most impressive server chip on the market right now, IMO.
It's really amazing, at work we switched like 100 instances from Xeon based to Graviton 3 and the performance is the same (as far as we can see) but the costs are waaaaay lower.
Thanks for this video. Very cool processor
The sound that made when the lid came off was gnarly
Just a PSA, you cannot compare geekbench 5 results between windows and linux systems.
Linux performs much better in the benchmark, for example my 8700k at 4.9ghz and not the best memory setup gets 8600 points on linux vs only 6700 points on windows.
This is especially true for cpu's with a lot of cores, so I would refrain from comparing it to hwbot scores as linux is banned there.
Indeed. I get a large boost in GB5 from just moving to Windows and Linux.
True. I also think runs on Linux are a lot more consistent since you don't have hidden background processes randomly eating up CPU cycles.
@@AD34534 Depends on configuration, I see lots of unexpected kernel threads and daemons that I fight to remove on my Linux servers.
Well, yes, you can compare the scores. The comparison makes Windows look bad.
@@tjthill well no, because the point is comparing CPUs, not Windows to Linux. Comparing a CPU on Windows to a different CPU on Linux isn't valid results since the Geekbench scores themselves will be different on the *same* CPU.
The die size is somewhat of 1/4 to Sapphire Rapids 4 tiles solution.
So it looks like good for highly parallel workloads.
It seems 1 core would be equal to Intel's Gracemont core. But Sierra Forest is only in couple years from now.
So interesting, if these Ampere CPUs are popular and a demand is high, why wouldn't Intel to spread their Atom family with more cores solutions now...
Ahhhh, excellent point! Id love to see an atom competitor!
Intel is also working on RISC-V cores. It could be that the successor to Sierra Forest is RISC-V-based.
@@ChristianStout Yeah, I think they want to develop risc v as opponent to ARM and keepin in mind Jim Keller's opinion that risc v is currently the most powerfull instructions sets, I think Intel can do very good with risc v accompanied foundry services.
@@Slav_space they also mentioned combination packages. So they could come out with something with a few xeon cores and a bunch of ARM and/or RISC-V cores. Would probably ease the transition considerably.
And with Zen 4c, AMD is also trying to compete in that space, it'll be a crowded space in the next couple of years
thanks Hetzner... so cool to see
ARM (Advanced Risk Machined based in Cambridge UK) also designs the circuits (not just the instruction set). Thye are designed with efficiency in mind. Some people (Apple) have taken ARM's circuit design and modified it - for their phones and PCs.
Der8auer: "You may recognize this location from a year ago"
Me not remembering what I did last week...
That processor is a bit over twice as fast as my EPYC 7601 (32C/64T). SC: 894. MC: 19070. Single core is about the same. MC is 2.2x faster.
For CFD work, both need more memory bandwidth. 8 channels is not enough.
To help me understand, I needed to look up the specs, heh. For the per-core stuff:
AMD EPYC 7601 = 3.2 GHz | Ryzen Zen1 | 894 @ 3.2 GHz
Ampere Q80-30 = 3 GHz | Neoverse N1 | 882 @ 3.0 GHz
To compare, N1 has +5% higher 1T GB5 IPC vs Zen1. Arm *claims* the 2022 N2 has +40% IPC gains. Iso-power, +10% perf (though with 8MB vs 4MB L3). Iso-perf, -30% power (both are iso-process, but presumably N2 will be on some 5-nm nodes).
[as my EPYC 7601 (32C/64T)]
The "MY" is key here. 15:33
@@Atheist7 Nope, that 7601 1T GB5 score is entirely representative. See the thousands of results at Geekbench: when unthrottled, all peak at a very, very similar ~890 points. And, clearly, it is Zen1 @ 3.2 GHz: it is very expected.
While anecdotal, it falls smack dab where the thousands of other results are.
This is cool! I would to play around with those someday
Excellent video and great to see Garfield the cat at the end.
I remember thinking in 2011 that it would only be a few years before ARM chips like this pretty much took over the whole data center server area, I am still bewildered that it hasn't happened yet.
It’s very shocked I agree! It’s weird how slow the industry has been to adopt switching over but I think it’s starting to happen. With Apple recently showing just how damn powerful ARM can be, I think that will help add some momentum. *fingers crossed*
Most code is written in x86_amd64.... So what do you expect.... Neighter vmware nore Microsoft are actively pushing this technology... And a lot of DC is all about virtualisatuon, and running stuff on it... And no a fuxking docket container is not that kind of virtual machine... 🙈
@@jolene6911 It needed the software support to allow it to happen. That's slowly come in to place now, hence widespread adoption of Ampere kit.
@@justacomment1657 For data centres goes without saying just use Linux where all its all open source/recompilable for your hardware and Linux OS? People using MS in the data centre aren't really in the game, IMO.
@@outcast6187 company dcs hardly use Linux only. Compatibility layer is almost always somekind of vmware or hyperV - mainly because the applications you run are written in x86 or amd64.what good is a postgresql database on an arm Plattform if your application is not written for arm... You now have to run x86 and arm hardware... Great... Not..
And yes, big data, shure... they do run on arm if the user had enough cache to buy something verry niche.....
Multipurpose systems... Like your ordinary 300-600 user infrastructure mostly does not.
I would love to see how it performs with all workloads vs Intel / AMD.
I would love to put my hands on that kind of system one day!
I hope that one day in the future, we will be able to use ARM cpus in our main system. I like to see motherboards be able to take in ARM CPUs from different manufacturers aswell as the ARM PC being able to take in different types of hardware such as GPUs
plenty of ARM cpus take gpus. PCIe is not an x86 exclusive communication bus, with even the Raspberry Pi having PCIe lanes. the only limitation currently to running big gpus on ARM is software and drivers.
@@zachbrown7272 I think OP means standard ATX desktop PCs, not server parts. PCIe lanes on low power SBCs are not always standard, requires an extension (daughter or motherboard) and they are quite limited in number and bandwidth.
During the first roadmaps of Zen, the AMD K12 family was supposed to use AM4 socket and the same motherboards. This was a promising step but AMD never released K12.
I remember when we had multiple mainboard chipset vendors, ALi, SiS, VIA, plus of course Intel; before that C&T and a bunch of others. And how Socket 7 was a cross vendor CPU socket. That was nice.
But things seem to be moving into the opposite direction where things are more integrated and more insular. Some suggest that discrete GPUs may disappear as well in the near future.
Still, some variant of a Broadcom server hub chip was used on last gen game console to glue IO to an AMD SOC via HyperTransport. Now all the MB logic is basically PCIe based. Would be curious if someone just standardised a CPU socket, or AMD licensed AM5. You could have a standard for AGESA like modules and a utility which integrates it into the mainboard firmware, which can be installed via Flashback, or you could have two separate flash chips, one being in the mainboard and another in the CPU package. A lot of work to untangle things.
There are bunch models support GPUs. Like Gigabyte, Supermicro, and Foxconn.
Ian over at TechTechPotato got a 45059 on Geekbench 5 with a TR 5995WX 64c/64t, so no SMT, with SMT was lower. Geekbench doesn't really like Ryzen as much as other architectures but the 80 core ARM is still doing well.
Yes very interesting performance of the 5995WX :) we shot the video about 6 weeks ago so we didn't have this info for comparison
@@der8auer-en Nice video but quite misleading. I don't really see the point of comparing a headless server platform with a workstation one, especially considering they will often run on very different operating systems.
Although many server tests were made few years ago when Ampere systems started to be available and it appeared that virtualised hosting and cloud services were the strength on those ARM processors, other types of server loads were not that better and sometimes they even performed much much worse on Ampere than Xeon or Power.
That is a shiny toy! Not a wonder it could kick a Threadripper. Thanks to you and Hetzner for being willing to show it off!
luv the video, thats a case for some conductonaut right there.
The orange angle pieces for the bench vice are technically called "soft jaws".. Merely to protect the item being clamped of the knurled clamp face being imprinted on the piece being clamped otherwise potentially causing damage.
In the past, strips of aluminium plate would be used for this as it is soft enough to conform to the shape without causing additional damage to whats being clamped, but since plastics now the norm..
aluminum would deform the copper heatspreader, so plastic is the right choice here. an annealed copper or brass would also work.
Thank you for a great video, for arm CPUs. Everyone says it is the future of the data centers. But I have 2 questions: 1-what OS you installed to test with geek bench, you mentioned Linux, but what type? 2. You deleaded the CPU for over clocking? You showed us the temp is not too high even at full load about 6ixties or even 5ivties. So why dealead?
They probably use Xubuntu in Hetzner testing bench. It should work on every distro (with similar performance) as most Linux distros today support ARM.
Regarding delidding, I guess that was done to show the CPU itself.
Hetzner is awesome! And that ARM cpu is a Monster! 80-cores 250w 56'c just lovely!
310W and 65C... Ohh and for that u need to delide it(bye bye warranty)... And still using thermal paste in 2022... Intel soldering too and its still sucks, but not that much... AMD much better and 128 core on the way in the WX5000's!!! Bye Bye shitty 80 core ARM
@@MadGangstaHUN doesn't need a delid.... Learn to use your brain
@@MadGangstaHUN Are you a child? Stop spouting nonsense and just appreciate newer hardware innovation. I'm sure the people deciding to use thermal paste are much smarter than you and would have good reasoning behind it. You'll also find 300w over such a large die won't have the cooling issues, like a 12900k with a die only the fraction of the size but similar current draw. The wattage per square mm would actually be rather small.
@@MadGangstaHUN All of this is false. AMD 5000 series tops out at 64 cores and 128 threads. 7000 series is rumored to reach 96 cores and 192 threads.
The whole system used 310W at the higgest peak, but the CPU power along was only around 200W or so. Other components such as memory, VRM and PSU losses, and IO amd storage cards can easily make up the difference.
Deciding was not needed to cool the CPU. Watch the install process again. They just stuck it in there and mounted the cooler. The delid was performed afterwards to showcase the size of the die compared to the size of the package.
Stop spreading misinformation.
@@MadGangstaHUN 310W was the whole system. The SoC was only 170W.
Exciting stuff!
Uberwundershön!! 🎉
The greastest thing about this is that ARM can be licensed by almost any company who wants to make a CPU and not locked as x86 is. This means more brands in the future for CPUs and that's good for consumers
It won't matter shit. This doesn't impact supply... They will have to fight for manufacturing quotas at TSMC at the same prices as anyone else. Do you think whatever brand has the benchmark leadership is gonna stop taking their premium?
@@haakonht no they won’t stop doing theirs best and it will take some time but for sure in the future there will be more than just Intel and AMD
@@haakonht, considering how much less die space is required for the same number of arm cores, starting your own production at smaller scale first, would be much cheaper and you won't need as much silicon. So I think it's entirely possible.
@@rinrin4711 pardon...? . Starting your own production.... Yeah.... Works really well for the Chinese... And they have a complete economy and the government behind it...
@@justacomment1657 More than that, PRC can't get licences from ARM anymore and that made many fabless companies to migrate to Risc-V.
That's certainly why Dinkleberg said "licenced by ALMOST any company"
The selling point of x86 isn't just performance, it's also the software support and documentation it has accumulated over decades of use.
X86 was really behind on performance on 10 years of intel doing nothing so AMD bring it back to live with zen cores, if Apple had the M1 before AMD Zen so today x86 would be dead
32-bit compatibility (which is the real x86) is just disappearing: Microsoft themselves have dropped 16-bit applications support from Windows (and, earlier, all remaining traces of DOS compatibility) just because the OS was becoming impaired by its own complexity and neede serious discards to become just maintainable. At the same time, Red Hat and Canonical went even forward by stopping compiling their Linux systems for 32-bit machines, and some ARM releases of CentOS and Ubuntu were already available years ago; not to mention FreeBSD which has its own project ("of course", ARM "runs NetBSD" ;-) ).
That perfectly makes sense to me: x86_64 is NOT x86, for sure it's going to stay, but that doesn't mean closing the door to alternatives; especially on the server market which has no need at all for some proprietary must-have desktop applications like, say, Skype or AutoCAD.
At the beginning of the 64-bit PC era, I saw 64-bit machines sold with 32-bit versions of Vista just because they didn't have the right drivers for the chipset. That, especially on the open source ecosystem, is going to be more and more a minor problem as time advances. I expect to see the same happen within the ARM ecosystem.
@@rdothegeek 32-bit compatibility (Compatibility Mode) is built into the x86_64 standard. The reason MSFT did not include 16-bit compat in 64-bit versions of Windows is because V8086 mode does not exist in Compatibility Mode at all. NTVDMx64 is not the same as NTVDM and emulates x86_16 or runs it in a VT-x VM. Has nothing to do with complexity and NT is very modular and can run different OSes executables through subsystems. The MS-DOS Subsystem (NTVDM) is still present in 32-bit versions of Windows. Other subsystems include WOW32 (win16 compat), WOW64 (32-bit compat), Win32 (32-bit/64-bit), OS/2 1.x, POSIX, and Linux. The difference between IA-32 and x86_64 is that we got 48-bit addressing, more registers, more paging levels, enforced flat segmentation, syscall instead of sysenter, and lost a few older instructions.
@@DorperSystems I understand all that, but as long as software can be compiled from anything at a higher level than Assembly it does not sound relevant. I would feel x86 as an advantage only if and where a platform (hardware+o.s.) appeared totally from scratch; which may pose the question of how much work is required to get a GCC (and maybe a JDK) in there. It does not seem to be the case here, and it is my opinion that nobody would waste a beast of an ARM server by running Windows on it.
Anyway, I wouldn't underestimate the important change in stability and efficiency that Windows achieved by (also) dropping backward compatibility, since the times of, say, XP.
@@rdothegeek Linux was specifically designed for x86 computers. It takes very little work to port GCC. It's like 100 lines to port GCC to a new OS. Windows Server provides the best experience for Active Directory and File Sharing. Using the Linux equivalents (OpenLDAP, Samba, NFS) is like castrating yourself with a rusty fork. You have to have the source code to recompile a program.
Liked your Cat in the Video, appreciate any Cat in a Video. Always love for the Cat.
What a beast, reminds me of the giant Alpha cpus :) my kind of thing
Can you compare benchmarks across 2 different version of the benchmark software? Like when you run the benchmarks on AIDA64, it warns you not to compare benchmarks gained on different version numbers of AIDA64.
Depends on the benchmark. I'd say that comparing the render times of say cinebench or blender is valid so long as the same frame(s) are rendered as putting "scores" aside, the only thing that matters (IRL) is render time, which makes it pretty platform agnostic
Depends on the benchmark itself, I wouldn't compare two off the shelf benchmarks across 2 architectures, however one can easily develop some benchmarks to test the performance between 2 architectures to calculate raw performance.
A benchmark like a blender test essentially times how fast blender takes to render a scene, it wouldn't measure the architecture's raw performance but it's what a user can expect in real life from that architecture
Different version numbers is not the same as same code built for different architectures.
In general there's no reason why runspeed of same code built for different architectures with recommended compiler for each shouldn't be compared. That would show how average software is going to run without need for any hand-tuning (which is expensive or even impossible if you intend to run some closed-source software).
@@JaenEngineering that's what I was thinking, if it was something that could be compared independently from the benchmark, render times, or Mips, then that would be meaningful. It was more the Geekbench score I was meaning, obviously that would be hard to compare across different architectures.
You certainly can. AIDA64 is just different software variations so optimising the software and reinstalling the newer version would obviously get a better result however in that instance you would be running the test on the same hardware.
Assuming you are running the bench on two different hardware systems there may be certain hardware advantages that allow for higher performance so you should also note to use the latest and most optimised piece of software for use for the best comparison between two different hardware systems.
What a beast! Excited for ARM-based gaming PC/home desktop in the future. Imagine the power. I mean there's apple mac m1, but imagine ARM-based windows desktop for gaming. That would be cool.
Wow, now I really want to be able to build an ARM-based PC similar to an x86 one, like that.
I am already using Ampere Altra ARM base server processor on Oracle Cloud and its performance is too good. I use ubuntu 22.04 for enterprise business application and I am happy with there performance as per cost :)
Awesome video! ARM based systems are certainly starting to flex their muscle; and competition is always good.
On a side note; if you compared that record-holding Threadripper score (34,735) with 64 cores and scaled up for an even number of cores (ie. 80 cores; comparable to the Ampere Altra Max), you're looking at about 43,400 points... so nearly the same as the Ampere. The TR absolutely thrashes the Ampere as far as single core performance, with 1236 pts vs 882 pts... or approximately 40% faster on the TR. Albeit, the efficiency is pretty horrible on the TR; given it was OC'd.
I do wonder how much the increased memory channel / throughput is benefiting the Ampere / ARM design, and helping to push the score higher...
Not bad against the old Zen 2 based Threadripper! But the 64 core Zen 3 Threadrippers do better. GB5 scores - Zen 3 Threadripper Single Thread - 1552 - Multithread - 45059 - Not too shabby for the same/similar (280w TDP) power envelope, thermals and 16 cores less than this ARM CPU! But I can't help but feel that these ARM "Server" CPU's are very underwhelming, and very niche, after all the sky-high promises a few years ago. Thanks for the fascinating video der8auer!
I feel like ARM is a couple years away from completely dominating the server market. hundreds of Low cost, low power cores are exactly the kind of hardware most servers need compared to Intel and AMD that have higher power single cores with higher energy and price costs.
I can run a 1080p streaming and file server from a raspberry pi 4, which has 4 relatively crappy ARM cores. For servers needing to handle massive numbers of clients with very little server processing, ARM is a very scalable and economical solution - at least until RISC V matures
@@merthyr1831 Most software is still single threaded. And a lot of the enterprise market uses virtual machine software that isn't available on ARM. Windows Server isn't available on ARM either. And "low cost?" None of Amperes stuff is low cost. A base model with 32 ARM cores (which are less powerful per cycle than x86 cores), 8 GB of DDR4, no hard drive, no NIC(!!), would cost you $5,686. You can get a Threadripper based 32-core based server for $300 less with 4x the cache. And that cache will be important when your load-store architecture is constantly pulling 8 instructions for what would be 1 instruction on x86!
@@merthyr1831 these chips are not lower power when you look at density, which with Zen4 Epyc out I don’t think AMD is going down easily. Not to mention CDNA3 also being a thing, something ARM has yet to replicate. It will be a tough battle as long as AMD sticks with x86
14:15 finally cat from stray finds his home on desk on front of camera 😀
Thank you!!!
With how power hungry systems are getting now, with Zen3+ and Core 12th Gen+, I wonder how long before we see a paradigm shift in the market away from x86.
It's going to be messy but surely it's better to do it sooner rather than later. Look at how much advancement has been made on ARM and Apple ARM systems, going head to head with high end Intel/AMD CPU's with a fraction of the power requirement because of the efficiency improvements in the instruction set not having to deal with 40 year old x86 compatibility.
Just look at the Apple M1 Max and M2 SoC's and how they dominate any SoC produced by Intel & AMD for total system performance. Its enough to make me at least want a Macbook Air for a laptop.
You need go look deeper, M1 SoC are basically an ASIC. They're very good in certain tasks and very bad at others. Geekbench also prefers Apple architecture, someone already prove this. This is why most serious desktop performance reviewers rarely use Geekbench.
@@bltzcstrnx The very bad at others is in just 3d stuff. X86 GPUs, both from AMD and NVIDIA have hardware acceleration for raytracing which apple's don't and that's why anything that requires rendering raytraced 3d scenes is multitudes of times faster on PC. Apart from that, mention anything that's faster on PC or that at least gives anywhere as good performance per watt.
One correction: x86 is only 4 years older than ARM. ARM has been in existence for over 36 years now. It's just on a different architecture that is much more efficient with much more simplified instructions compared to x86's.
@@fidelisitor8953 not only ray tracing, M1 is very good in ML workloads. Although gaming workloads leaves much to be desired. There are reviews that even the M1 Max having half the performance of RTX 3080 in non gaming ray-traced games. This is with a system that cost less than the M1 Max, yes you can build a RTX 3080 that cost less than M1 Max system.
Edit: also a lot of features for modern game 3D rendering aren't there. I think UE5 Nanite is one of the engine effected by it. There are rendering features that are not available yet common on other 3D accelerators.
@@bltzcstrnx Apple silicon isn't made for gaming and Apple makes that very clear by having close to zero support for anything gaming and you have to understand why.
The Apple silicon macs are made for different set/category of people, mostly with creators in mind so they're basically a creators machine with heavy focus on stuff like photo/video editing, music production, data engineering (with NPU hardware acceleration) and the likes. So it's not like Apple can't produce the same level of performance and efficiency for gaming or 3d stuff but rather they don't see a profitable need to move into that kind of stuff because they'll absolutely steal the market if they do.
direct die/LM? if this thing can even be overclocked, maybe you'll be able to match the zen 3 epyc
at least from the stock BIOS no modifications were allowed
Thanks for the video, would you be so kind to share what benchmarking and monitoring software you are using on Linux?
The item used in the jaws are simply called 'soft jaws'. They're the sacrificial layer that protects what ever item you're clamping from the vice itself.
Going back to my old datacenter days, a RISC system would eat any CISC systems for lunch at any time of the day. Intel only won because it was much cheaper than a Sparc system (and Oracle bought Sun Microsystems to take over Java). Now ARM is once again proving RISC is superior by not having all the bloat of CISC architecture. Apple bet on the ARM platform and Ampere (or any other server grade ARM CPU) on the datacenter world will cause some stir - and that will be great for everybody.
I don't know when your days were, but Intel P6 workstations from Pentium Pro onwards destroyed Sun's RISC crap for 1/7th the price. That's why Sun dried up and blew away.
The first ARM chips where also faster than Intel at the time and at a fraction of the power (late 80's)
It's less about the architecture as what you optimize for. Apple optimized for desktop with lots of special purpose hardware, AMD and Intel largely optimize in sharing one architecture from mobile to server while alterra only does server workloads with as much cores as possible.
The reason while ARM CPUs are faster than x86 is because that's currently the only competition. They can't use x86 and RISC-V isn't ready yet on the software side.
like @Mr. Moto said, if you haven't tried Pentium Pro's and later systems, they were really a lot faster (in most workloads) than any current sparc's of the time, while being alot cheaper and energy effecient. CISC's were not at all bloated at that time since every little thing they had was in use. now, of course they are, because they still retain a lot of things from the past for compatibility reasons. but keep in mind that arm also cannot change that much without it too becoming not risc.
@@giornikitop5373 Also x86 can just implement many of those mostly unused instructions in microcode, so they don't have to be on the silicon.
Is he using gauge blocks as vice jaws? Oh my god 😆
Yeah, when he showed the plastic soft-jaws I thought he was just holding bits of aluminum, but when he described gauge blocks .... 💀
Awesome power and efficiency of ARM.
What Linux is needed for that CPU?
They apparently also have an 128 core cpu, i’d love to see how that goes
Yep, the 128 is called Altra Max, socket compatible. Single threaded and same cache per core, so it scales linearly.
@@vikingforties yes, but linearly wouldn’t quite be correct since there’s always diminishing returns with going big unless they get everything perfect
it goes weeeeeeeeeeeeee
@@nguyenminh7780 Agreed, generally the CMN 600 mesh would handle the extra core load but you could potentially run out of PCIe lanes or start to have stalls if the workload is memory bound.
I always thought that all ARM CPUs are always soldered permanently to the Motherboard.
So I was wrong
Here I am thinking arm cpus are for phones and tablets lol
you should be soldered permanently to a motherboard
it's all about what it's engineered for, if you need a ton of power and upgradeability and aren't as worried about the size then any cpu can be made socketable
the only reason all phone CPUs and most laptop ones are soldered is that the whole phone motherboard is smaller than this CPU's PCB
You probably weren't wrong a short 2 years ago, it is just new applications calls for different packaging
The word you were looking for, as pertaining to your vice, is called a "jaw". Yes, the same as your jaw, which is made to bite down on things. Hence the name.
Thanks for the video, Roman! Excellent work, as always! o7
the part he was talking about was actually a "jaw pad" or "jaw cover"
Can’t wait for stuff like this with RISC-V
that upright cat sleep state
That's insane. I bet if I glued 80 3ghz Pentium 4s together they'd have about the same performance... (although, at 60+ watts per chip...)
the orange cat was a nice suprise
Great video. Thx a lot? What about RISC-V CPUs ?
ARM stands for Advanced RISC Machine, but when it was Acorn property it was Acorn RISC Machine. Point of note, but I’ve seen some dubious false knowledge of it being touted on the internet which would surely upset Steve Furber and Sophie Wilson who both worked on its creation. At least they’re both around to dispute the claims of it being an American design.
You could make a whole video about its origin if you felt like it.
*Roger Wilson
@@desktorp That was Sophie Wilson’s given name, but we won’t spend too long deadnaming her now will we. Most transgender people would rather others didn’t. Thank you in advance for your expected manners in this _delicate_ matter.
Sophie Wilson is a computing legend. She wrote the BBC BASIC interpreter in 6502 assembly language for the BBC Micro at the beginning of the 1980s. Later she designed the ARM instruction set and wrote an emulator, again in 6502 assembly language, which ran on a BBC Micro (with 6502 second processor), which proved the concept. Steve Furber then implemented the design in hardware and worked with VLSI to manufacture the silicon. The original ARM1 was used in a second processor for the BBC Micro to further prove the concept. The second iteration, ARM2 was used as the basis for the Acorn Archimedes home computers, along with three support chips, also produced by VLSI, called MEMC, VIDC and IOC, which provided memory, video and general IO support, respectively. The interim ROM-based single-tasking operating system was called Arthur and early buyers got a free upgrade to RISC OS when it was released. I bought an A440 Archimedes with 4MB of RAM, an 800KB 3.5-inch floppy drive and a 20MB hard disk. I moved on to a RiscPC later and I still have both. The best thing about BBC BASIC is that it included an inline assembler, first for 6502 and, on the Archimedes and RiscPC, ARM assembly language.
@@johnm2012 Now if only I could get my hands on the coveted ARM second processor they used to test it on, now _that_ would be something. Story has it that it used so little power that when they accidentally connected it up wrong it ran on just the power coming in through the IO pins’ protection diodes.
That must have been a foreshadowing moment!
@@Jenny_Digital I seem to remember Steve Furber confirming that little anecdote in an interview he gave about the SpiNNaker project. I'll try to find it - no promises though - and post a link.
I've never seen an ARM1 second processor for sale - they must have only made a handful for internal use - but there's an open source project to implement a generic second processor for the BBC Micro, using a Raspberry Pi. The Pi connects to the 2MHz "Tube" interface and runs code that allows it to emulate all manner of second processors, including 65C02, Z80, Torch Z80, 32016, 80186 and ARM.
Yes ARM is impressive, but for the rest of us to get interested, they need to run just as fast while emulating x86 so the x86 software can be run.
I know in my use case as a user, I am not willing to burn everything and run fully ARM, and even Apple realized this and built a semi x86 / ARM CPU with the M1 chip where they have some x86 functions in hardware, but most of it is done in emulation, with issues.
The only REAL reason why there is very little available in MS Windows and ARM is that Microsoft won't grant MS Windows licences to ARM processors unless under specific circumstances like for their own SQ1 or SQ2 CPUs etc. Microsoft's literal problem is they make too much money to care, in other words MS's stock price needs to drop considerably before we see more Windows on ARM CPUs, and that might take another decade.
The m1 is not semi x86 at all. The only thing it has in hardware to support for x86 like 'total store ordering', and that's a super minor change in hardware. The rest of rosetta 2 is just well implemented binary translation from x86 to arm.
There are also rumors of intel and microsoft to have a backdoor deal in blocking ARM proliferating to desktops because with ARM being such an open platform it's adoption would allow many other manufacturers other than Intel and amd (TSMC in background) to enter the market and make desktop CPUs. So Microsoft's Windows ARM version will always be artificially blockaded.
This is both to keep the windows+x86 monopoly and also geopolitically to block China and others to catch-up with semiconductors manufacturing.
That's a lot of cores. I'm not sure, but I have a feeling that in day-to-day use it might be hard to efficiently parallelize whatever one is running over all those cores.
Then again I guess processors with such a large numbers of cores are mostly intended for really heavy duty processing.
2:53 Так же можно установить кастомную крышку или просто снять её и использовать рамку, а так же стандартный "квадратный" куллер.
Even with more cores, you'd be losing performance with the translation layer for the x86 applications. For memory intensive stuff, Epyc has twice the amount of memory channels and wins again. Also, the architecture is a bit concerning. ARM chips tend to have asymmetric core steups in order to deal with temperature targets.
I doubt that x86 translation will play muh of a role here. A few major Linux distros already have native ARM support. It's not unrealistic to see cloud datacenters running their own in-house software on these just like they do on x86
Geekbench is a useless piece of irrelevant software. According to it, a Raspberry PI4 would fit on the same scale as x86 CPUs of the last 15 years.
If you want a more relevant and accurate standardized benchmark for servers that need to perform calculations, memory operations, etc., then you can run the built in benchmark in Wolfram Mathematica that is native to Windows and Linux x86, but also Linux and OSX ARM. It runs natively on Raspberry Pi Zero, Apple M1, Epyc, Xeon and everything in between.
This 👍
I would be verry interested in such a comparison
nice video, oh did the CPU survive ?!
Is the Hertzner CPU package power tool available somewhere as open source? I've had a snoop around their public repositories and can't find anything. Looks handy, not sure if it's just limited to server CPUs/monitoring interfaces.
Which is why Nvidia entering arm cpu for desktop, data center, and A.I. markets is appealing....
arm desktop lol, unless NV makes an insanely good rosetta-like software that translates x86 to arm instructions i dont see it having any success
Nvidia already entered arm cpu for desktop a decade ago. Then they stopped.
Geekbench is GARBAGE for comparing across different ISA's. Use REAL WORLD TESTS! Also, try and compare it to a current Zen 3 Threadripper Pro 5000 part instead and I doubt that even Geekbench would look advantageous. This isn't NEARLY as impressive as you seem to think... The TR 3990X is OLD!
Just a heads up..
Current world's fastest supercomputer is ARM based. Don't remember the exact location, but the country that host it is Japan..
The score you see here is from a single machine whereas super computers can be 1,000's of machines.
Damn I love that cat :* Reminds me very much of a cat I used to have, apart from the white boots, neck and privates.! He was a "Real Ginger", feisty, fierce, loving, beautiful and independent.
The fact that 80 cores and 128Gen4 lanes along with 8 channel DDR4 fits on that die is BEYOND insane.
AMD Epyc spends a massive 400mm2 just for a for the same amount of IO on their IO die using an equivalent process and a great deal of this is spent on PHYs.
The big kicker between them is the Cache. AMD has 32MB of L3 per 8 core cluster for a total of 256MB of L3. But these ampere SoCs only have 16MB of L3 shared amount ALL the cores.
This has huge implications for core-core communications and data sharing, which is why the CPU is going to be more special purposed than the more general AMD offering.
In its niche however, it should be amazing.