AWS Graviton4 96-Core Performance vs. AMD EPYC & Intel Xeon CPUs

[phoronix]

Phoronix: AWS Graviton4 96-Core Performance vs. AMD EPYC & Intel Xeon CPUs

Last week I published some initial benchmarks of the Amazon/AWS Graviton4 processors now available within the EC2 cloud using the new "R8g" instances. That initial comparison was a 64 vCPU comparison of Graviton4 against AMD EPYC and Intel Xeon 64 vCPU AWS instances. In today's article is a look at the 96-core Graviton4 bare metal performance using the "r8g.metal-24xl" AWS instance type. The Graviton4 r8g.metal-24xl performance was then compared in today's article against various bare metal AMD EPYC, Ampere Altra Max, and Intel Xeon processors in the lab at Phoronix.

https://www.phoronix.com/review/graviton4-96-core

[ddriver]

So Intel is now officially third place and confidently striding towards turd place!

[schmidtbag]

That's REALLY bad when ARM manages to outpace Intel in high-performance workloads, especially when you factor in the lower core count. Interestingly, I couldn't find any info on Graviton4's wattage, but I'm sure it never exceeds 300W.

[coder]

Wow! These are even more interesting than the previous round of Graviton 4 benchies! It's great to see such diversity among the contenders! That said, I'd have probably missed Altra the least, but it does somewhat complement Sierra Forest in showing how well (or poorly) you can compensate for lack of single-threaded performance with sheer core count.

[Adarion]

Oh, well.
I second what Michael wrote, too bad there are no power consumption numbers. That would really give a full picture.
Then, some things might not scale super-well on many cores, like x265, so the 64 core with high frequency is in front.
The results so far are still interesting, Epyc slaughtering nearly everything else in most disciplines. And after all these years I am especially happy about it.

Meanwhile, intel seems to have issues with its recent desktop line. (if one is interested here's some video https://www.youtube.com/watch?v=oAE4NWoyMZk with people talking about it - tl;dr is that the most recent CPU generation seems to have hardware issues, for all OSs, gamers and server people alike, and it is also not the fault of boards or memory.)

[Jumbotron]

I was actually banned for a while by Michael or the algorithm on this site for touting the fact too many times that the Age Of ARM is upon us and that the writing is on the wall for all x86 based technology going forward and particularly by 2030. My prediction was that by 2030 50% of all new computers NOT a tablet or cell phone would be ARM based. I was including Macs and ChromeOS along with ARM based Windows computers which we are seeing roll out now, but still the paradigm shift in CPU platform from x86 to ARM even for consumers will be self evident. Of course when you combine that with the complete takeover of ARM in the portable, wearable and IoT space and aincreasingly in the HPC space, it's hard not to believe that what I say is true. You can also see this in the dev space as more and more ARM based work is being done in most compilers like GCC and LLVM, languages and toolkits. It is still my belief that even though there is rather little of CISC in today's x86 chips it is still impossible to make an x86 chip as low power as ARM but still retain high performance. In other words, the power to performance metric will always be in favor of ARM because they have had to slowely but surely engineer higher amounts of performance while still keeping the power demands to as bare minimum as possible, something the x86 world has never had to really think about until now. X86 world have always had to whittle down their tech to meet a power demand equal or better than ARM but in the end always lose performance to a greater degree becasue of the very architecture itself. The only way to change that would be to abandon x86 altogether and that's not going to happen. And if it did then Intel and AMD would be producing something very much akin to ARM or RISC-V. And it certainly would be a clean sheet Intel or AMD version of a RISC chip in the end.

[coder]

Oh, well.
The results so far are still interesting, Epyc slaughtering nearly everything else in most disciplines.

But the results (especially if you exclude the X3D cache model) are much closer to Graviton 4 than in the previous 64-core benchmarks!

[coder]

I was actually banned for a while by Michael or the algorithm on this site for touting the fact too many times that the Age Of ARM is upon us and that the writing is on the wall for all x86 based technology going forward

For realz? I've been broadly supportive of ARM, but never had any trouble with mods or filters.

My prediction was that by 2030 50% of all new computers NOT a tablet or cell phone would be ARM based.

To be honest, if you'd have asked me like 5 or 7 years ago, I thought it would've actually happened by more like 2025.

along with ARM based Windows computers which we are seeing roll out now,

It's much too early to declare victory on that front! AMD and Intel won't allow x86 to go quietly. Snapdragon X is impressive, but it's not the absolute revolution originally promised by Nuvia. Let's see how their 2nd gen turns out.

Also, I think it might help the Windows-on-Arm market to have some real competition. IMO, it was ultimately self-defeating for Qualcomm and Microsoft to enter that exclusivity agreement, which is finally about to lapse.

The only way to change that would be to abandon x86 altogether and that's not going to happen.

I'd say APX closes the ISA gap by at least 50%.

if it did then Intel and AMD would be producing something very much akin to ARM or RISC-V. And it certainly would be a clean sheet Intel or AMD version of a RISC chip in the end.

Well, rumors suggest AMD is launching an ARM-based CPU next year. Also Nvidia. Mediatek will round out the WoA laptop SoC market. I expect AMD also to target the cloud with ARM-based CPUs. They're losing business to in-house ARM CPUs and it would make more sense for them to do an ARM CPU core, if they're also able to sell it into the cloud, and not only in laptops.

I honestly wonder if Intel will just skip ARM and go straight to RISC-V. In the battle between ARM and RISC-V, the big wildcard is China. Because of them, I think RISC-V is going to dominate a lot sooner than it otherwise would've. I'll say by 2035, but maybe I'm now erring on the side of being too conservative.

[skeevy420]

I was actually banned for a while by Michael or the algorithm on this site for touting the fact too many times that the Age Of ARM is upon us and that the writing is on the wall for all x86 based technology going forward and particularly by 2030. My prediction was that by 2030 50% of all new computers NOT a tablet or cell phone would be ARM based. I was including Macs and ChromeOS along with ARM based Windows computers which we are seeing roll out now, but still the paradigm shift in CPU platform from x86 to ARM even for consumers will be self evident. Of course when you combine that with the complete takeover of ARM in the portable, wearable and IoT space and aincreasingly in the HPC space, it's hard not to believe that what I say is true. You can also see this in the dev space as more and more ARM based work is being done in most compilers like GCC and LLVM, languages and toolkits. It is still my belief that even though there is rather little of CISC in today's x86 chips it is still impossible to make an x86 chip as low power as ARM but still retain high performance. In other words, the power to performance metric will always be in favor of ARM because they have had to slowely but surely engineer higher amounts of performance while still keeping the power demands to as bare minimum as possible, something the x86 world has never had to really think about until now. X86 world have always had to whittle down their tech to meet a power demand equal or better than ARM but in the end always lose performance to a greater degree becasue of the very architecture itself. The only way to change that would be to abandon x86 altogether and that's not going to happen. And if it did then Intel and AMD would be producing something very much akin to ARM or RISC-V. And it certainly would be a clean sheet Intel or AMD version of a RISC chip in the end.

I kind of expect either Rosetta-like or dual-arch solutions. Take desktop users, workstation users, and gamers, 8 worth a shit x86 cores are more than enough for their needs in regards to x86 software that doesn't have a FOSS equivalent or native Linux version. My X3D barely exceeds 20% usage in modern games like Elden Ring while the GPU, 6700 XT, is in the 90-100% range (however MangoHud reports). Instead of adding more and more cores, AMD could pivot over to some sort of Arm/x86 hybrid like what they do with the PS4...although don't judge that solution by the PS4's abysmal OS performance...

At this point in time all the major operating systems have ARM and x86 support so regardless of which way anyone goes, translator or dual, there's still going to be the need for software and schedulers to be aware of how to handle different architectures so that's not much of an argument for or against either one.

And feel free to replace ARM with whatever architecture you please. I'm just hypothesizing the need to have lower power devices and the need to run older, proprietary desktop software that can't be recompiled. I suppose the translator option would be better when factoring in older PPC software from desktop iMacs.

Ok, so now we're gonna need a triple-arch solution with ARM, PPC, and x86

Shit, Solaris workstations used SPARC. Queue up the quad-arch .

I do think it's rather interesting that we're at a point in time where some sort of hybrid architecture clusterfuck chip is something that can be considered over adding more and more and more of the same cores. Not everything scales that way and having workload-specific nodes on our CPUs might be a way to go forward.

[skeevy420]

coder
Sometimes the Phoronix auto-mod is weird. You can write the most pleasant, nicest post anyone can muster and it'll flag you. You can post things that most rational people would consider borderline hate speech to get a rile out of folks and it'll wave you through.

Sometimes you can just post too many words and the auto-mod is like, "Ain't nobody got time to read this shit!" so it'll flag you out of caution. That looks like what happened to Jumbotron based on the posts he's been writing lately.