New AMD Epyc vs ARM and Intel Benchmarks

[IceDigger]

Who doesn’t like a good fight?

Usually it is Intel Xeon vs AMD Epyc BUT this time around Phoronix is throwing the new IBM Power9 into this server CPU fight!

A fun break from the Ryzen vs Xeon.

“The Talos II server with dual IBM POWER9 22-core servers ended up delivering performance around that of the EPYC 7551 previous-generation Naples processor. But overall the Talos II POWER9 had quite a respectable showing compared to the x86_64 CPUs. While not at the same performance level as Intel Cascadelake or AMD Rome, the benefit of the POWER9 route is the open ecosystem around the processor and for the entire system from the likes of Raptor Computing Systems where they provide schematics and open-source code down to the BIOS/microcode. As for the ARM performance, the Ampere eMAG came out ahead of the Cavium ThunderX for this range of single and multi-threaded benchmarks as while it has less cores than Cavium, it does offer much higher clock speeds. There’s also the matter of the software support as it stands today with x86_64 certainly at the advantage with more software packages being tuned for it. However, there is increasing exposure to ARM and POWER by open-source developers that may yield more untapped potential moving forward. So that’s where things stand today with regards to the latest Intel/AMD CPUs against IBM POWER9 and ARM for currently-released products and the current Linux software stack – we’ll certainly repeat these tests in the months ahead.” – Phoronix

Full Benches and Story

 

[Zarathustra]

Interesting data on the POWER9 system.

I haven't dabbled much with ARM in the server space yet, but I'd imagine it can be limiting from the perspective of many of the applications I'd like to run.

I wonder what proportion of the software I run on my server has ARM versions.

 

[Grimlakin]

I bet mssql. Does but it's the same per core licensing. Lol

 

[IceDigger]

I bet mssql. Does but it's the same per core licensing. Lol

Ouch, I forgot about the liscensing per core cost for a lot of software. That has got to be awkward when pushing it in the future.

 

[Grimlakin]

Imagine those 128 core arm processors. No hyperthreading. At 10k per 2 cores. 640k in licensing just for the sql enterprise. So that server better be running a metric ton of sql instances to consolidate down to that one piece of iron. At least VMware is still per socket.... For now.

 

[Zarathustra]

Imagine those 128 core arm processors. No hyperthreading. At 10k per 2 cores. 640k in licensing just for the sql enterprise. So that server better be running a metric ton of sql instances to consolidate down to that one piece of iron. At least VMware is still per socket.... For now.

From a home user perspective I always hated per socket licensing, because I am often usi g older cheaper hardware with fewer cores, so I'd be paying twice in lice sing fees for dual hexacores as some Enterprise would for a single 12 core. Seems wrong somehow.

If anything, per-core licensing seems more fair.

 

[Grimlakin]

From a home user perspective I always hated per socket licensing, because I am often usi g older cheaper hardware with fewer cores, so I'd be paying twice in lice sing fees for dual hexacores as some Enterprise would for a single 12 core. Seems wrong somehow.

If anything, per-core licensing seems more fair.

Trust me you want per socket licensing. And you do not want per core licensing to catch on especially with the way things are going today. Each new generation is increasing core count and speed as well as IPC. Sure you might not upgrade for a couple years but when you do you're going from 4 to probably 8+ cores at a bump. Where if you had per socket licensing your software license as an end user cost would not increase.

Per socket licensing is better.

Of course we are talking enterprise software licenses here. If you are running enterprise software on your personal desktop you probably have there werewithall to have a VM with fewer cores. But that's just a thought.

 

[Brian_B]

Interesting results.

The only thing that may paint a more complete picture would be energy consumption.

Not a big deal that ARM (or whatever arch) is slower, but if it can use less total energy to get at the same result.

Note that total energy would be energy demand multiplied by time required to perform the work: That would take into account the argument of "race to idle" that has become popular. I wouldn't be surprised to see some of these faster chips are using less total energy, even at a higher peak energy demand / TDP - especially when you are looking at speeds that are 2-6x faster.

A few ways to accomplish that while your running the same benchmark suite - it wouldn't necessarily mean a lot more benchmarking time

I would think most people who are dealing with rack servers would be very interested in that metric.

 

[Zarathustra]

Of course we are talking enterprise software licenses here. If you are running enterprise software on your personal desktop you probably have there werewithall to have a VM with fewer cores. But that's just a thought.

Only thing I've ever paid a per-socket license for was a hypervisor. I eventually stopped that subscription because I was pissed about paying for two sockets and being "punished" for using older more affordable hardware.

I have no problem with per core licensing, as long as it isn't a cash grab.

If a company says "hey, we currently charge $400 per socket, and our typical customer uses 8-core CPU's in those sockets, so we will shift to a $50 per core model" I have no problem with that.

 

[David_Schroth]

If a company says "hey, we currently charge $400 per socket, and our typical customer uses 8-core CPU's in those sockets, so we will shift to a $50 per core model" I have no problem with that.

That's exactly what they want you to think though. Then the next generation of chips you upgrade to, you'll have a 16 core chip that you're now paying $800 to license.

 

[Grimlakin]

That's exactly what they want you to think though. Then the next generation of chips you upgrade to, you'll have a 16 core chip that you're now paying $800 to license.

This is exactly the point I was trying to make.

 

[Zarathustra]

That's exactly what they want you to think though. Then the next generation of chips you upgrade to, you'll have a 16 core chip that you're now paying $800 to license.

Meh. With every generation of their project they decide what to charge for a licence what the market will bear. They raise and lower prices accordingly based on what the competition will do. I don't think moving to a per core license changes overall costs long term. If their prices are unfavorable compared to the competition they will adjust them, just like in per socket licensing. It just changes the distribution of who pays what.

 

[Grimlakin]

Meh. With every generation of their project they decide what to charge for a licence what the market will bear. They raise and lower prices accordingly based on what the competition will do. I don't think moving to a per core license changes overall costs long term. If their prices are unfavorable compared to the competition they will adjust them, just like in per socket licensing. It just changes the distribution of who pays what.

Yea because nobody prices an enterprise product low to drive competition out of the market then jacks up prices once they have a comfortable footprint in the market space. Nope nobody at all. Never been done. /s

 

[Zarathustra]

Yea because nobody prices an enterprise product low to drive competition out of the market then jacks up prices once they have a comfortable footprint in the market space. Nope nobody at all. Never been done. /s

Yeah, but they can do that with per socket licensing and per core licensing alike.

They are always seeking to maximize profit, so you set the price at the highest price people (or enterprises) are willing to pay. At some pricepoint the value it provides is no longer worth the price, and they want to avoid hitting that.