Opteron vs. EPYC Benchmarks & Performance-Per-Watt: How AMD Server Performance Evolved Over 10 Years

[torsionbar28]

Holy crap I didn't know Opterons sucked that hard. No wonder AMD wasn't making any money at all.

Uh, No, you're forgetting that these are 10 year old models from a 12 year old platform. Compare to intel Xeon of similar vintage and you'll find roughly the same performance. AMD was reasonably competitive with intel back then. Kids these days.... SMH.

This article is comparing an ancient DDR2 platform to the latest DDR4 platform. The three generations of DDR3 Opteron that came in between are missing from this comparison.

[edwaleni]

2.5 inch NVMe "drives" are a compromise to some degree. Datacenter service managers who didn't want to lose access to storage from the front and to manufacturers who didn't want to make wholesale changes in their rack designs.

Also an issue is signal loss using the kluge M.2 to U.2 cable adapters. Having 2 physical connectors between the NVMe storage and the signal bus raises interference and can cause signal loss. Seems to be a compromise in design.

We have recommended that they instead use NVMe trays for the front bays of modern servers that hinge down and expose the M.2 slots for insertion and replacement. Someone (we think Intel) doesn't like that as they say it exposes the planar to random static shock from DC floor staff performing maintenance or replacement. They prefer "drive" looking NVMe hardware that is shock and static resistant and provides a better cooling profile. We think it wastes space and provides no better density or cooling than M.2.

All that said, it appears there is room for innovation for NVMe storage formats, especially for hardware that can accommodate several NVMe devices in a single chassis.

[leipero]

Let's add some cringeworthy post, now that's an EPYC improvement .

[bug77]

Holy crap I didn't know Opterons sucked that hard. No wonder AMD wasn't making any money at all.

mbello Intel only lost that leadership on paper, their silicon still packs more transistors per sq mm than others'. But their leadership has eroded over time and yes, single threaded performance doesn't see as much improvement as it once did. Then again, x86 and x86_64 have been set in stone for decades, there's a limit of how much can be squeezed from any given architecture.

[mbello]

This is quite interesting to see.

What strikes me the most is the evidence that single-threaded performance is not scaling as fast as it used to a decade ago. The only benchmark that explored single threaded performance saw the new Epyc roughly 2x as fast as the old Opteron. The frequency is about the same 10 years ago and now, but the memory bandwidth and latency was vastly inferior. Yet, 10 years later, the shiny new processor on a single threaded basis and with much better interconnects is only 2x the speed.

Not being able to increase single threaded performance faster than the competition means Intel need an edge on process technology (manufacturing technology) more than it ever did. The problem is, Intel just lost its edge on manufacturing (to TSMC, with Samsung and GlobalFoundries not too far behind).

So Intel is no longer the leader in process technology and its single threaded performance is not advancing as fast as it used to giving its competitors (Ryzen, Power 9, ARM) ample time to catch up. And Ryzen proved they mostly did.

Intel now is increasing single-threaded performance on specific applications by deploying accelerators to its many SKUs. AVX-512 is one of them, but there is also QuickAssist in its many versions. AMD is not as rich as Intel and cannot have an SKU portfolio as wide, so it is focusing on cheaper packaging and is keeping their processors free of these new adons (although they already have many). x86 CPUs are general purpose processors afterall, there is a point in not adding bulk to it in the form of accelerators and instead let other system components do that (GPUs, NICs, etc).

Very interesting times ahead. I believe AMD has a solid business model, and the news out there is that Ryzen 2 already taped out at GloFo 7nm, so the war is just starting.

[Michael]

I was Waiting for the Article since it was announced .. but hell ? Why only 2 2300 Series CPU no 6100/6200/6300 ? I like most of the article but this one is nearly useless with the current count of CPU. IMHO thats not a evolving thats only how the performance was 10 years ago and now ... a real performance evolving article is like this one: https://www.phoronix.com/scan.php?pa...ronix-13&num=1 or this one: https://www.phoronix.com/scan.php?pa...ry-power&num=1 there we could see the evolving betwween the generations.

Greeting

Because the 2300 is the only opterons I had, as I believe I said in the article. Obviously if I had some 6000 series Opterons, I would have included them but unfortunately I don't have any and never did.

[Ground0]

I was Waiting for the Article since it was announced .. but hell ? Why only 2 2300 Series CPU no 6100/6200/6300 ? I like most of the article but this one is nearly useless with the current count of CPU. IMHO thats not a evolving thats only how the performance was 10 years ago and now ... a real performance evolving article is like this one: https://www.phoronix.com/scan.php?pa...ronix-13&num=1 or this one: https://www.phoronix.com/scan.php?pa...ry-power&num=1 there we could see the evolving betwween the generations.

Greeting

[Michael]

You got this one from mozilla's article on Quantum, didn't you Michael ? I don't recall this term being used on phoronix before.
I am not sure it is a valid use in this case, though. Could someone comment on this? To me, it seems only embarrassing to the dev who has to program it.

Now, on topic, it's great to see actual figures for these improvements. I will use these to point out that frequency != performance.
Any word on the price point of the system tested here? Are they comparable?

You got this one from mozilla's article on Quantum, didn't you Michael ? I don't recall this term being used on phoronix before.
I am not sure it is a valid use in this case, though. Could someone comment on this? To me, it seems only embarrassing to the dev who has to program it.

Now, on topic, it's great to see actual figures for these improvements. I will use these to point out that frequency != performance.
Any word on the price point of the system tested here? Are they comparable?

No. It's been used on Phoronix... That "EP.C" in NPB is short for embarrassingly parallel, C data set size.

[M@yeulC]

embarrassingly parallel

You got this one from mozilla's article on Quantum, didn't you Michael ? I don't recall this term being used on phoronix before.
I am not sure it is a valid use in this case, though. Could someone comment on this? To me, it seems only embarrassing to the dev who has to program it.

Now, on topic, it's great to see actual figures for these improvements. I will use these to point out that frequency != performance.
Any word on the price point of the system tested here? Are they comparable?

[chuckula]

I've got an idea: Intel shouldn't release anything new in the server space for 10 years. Then in 2027 we can be all shocked & amazed when their 2027 Xeon parts beat up on their 2017 parts and act like it's a big deal.