Info [Toms, Anand] AMD EPYC Benchmarks

[lobz]

He has a point though about the output. In the athlon days AMD was simply supply limited. And given they have issue with supply here of 3900x I do question how big a dent they can make. Unless 3900x is limited due to clock speed (binning) are because high demand of Rome.

dude... they had their own fabs back then, with _much-much_ lower capacities that what they are able to order. frok TSMC aylt any time, I would bet my entire fortune on them not having any capacity issues this time.

 

[Panino Manino]

He has a point though about the output. In the athlon days AMD was simply supply limited.

Me thinks that AMD was more "bribery limited" than supply limited back in the good old days...

 

[CHADBOGA]

Me thinks that AMD was more "bribery limited" than supply limited back in the good old days...

When Dell started selling AMD after Conroe was out and AMD couldn't supply their retail channels, what hope would they have had of supplying oems, before Conroe came out?

 

[Mopetar]

Bloody hell!

I knew Rome would be good... but wasn't expecting a bloodbath.

Rome calls to mind the Colosseum and it's clearly Intel that's being fed to the lions.

 

[Abwx]

When Dell started selling AMD after Conroe was out and AMD couldn't supply their retail channels, what hope would they have had of supplying oems, before Conroe came out?

How could one build capacity production without the money..?..

Because you think that Intel briberies started with the Athlon 64..?..

They started first with coercition as soon as 1999 against whoever would release an Athlon Slot A compatible motherboard, only two minor vendors released a MB, namely FIC and a firm that took advantage of this opportunity to become quite bigger, that is, MSI.
Asus came a little later with an anonymous white box with no brand marking while the deservedly deceased ABIT publicly stated that they wouldnt release such a MB.

When the Athlon XP was launched in 2001 Intel bribed PC retail shops by paying them half their advertisements in exchange of not selling AMD based PCs, i knew about it from a seller in a shop that ended being trialed for the thing with Intel being fined by the EU (the Saturn stores case...).

As for the Athlon 64 days, well, just Dell received 5bn rebates for not selling Opterons servers and Athlon 64 PCs, HP and the rest were bribed about equally the time it took for Intel to release a competitive, albeit much overestimated, CPU...

Look that you were either not here at the time or eventually lost in Intel s marketing kool aid...

 

[maddie]

Me thinks that AMD was more "bribery limited" than supply limited back in the good old days...

Nah. Their fabs could have never displaced Intel then.

 

[Schmide]

My general thoughts. As to Ryzen3 binning and voltage, without a doubt, there is no reason not to waste a few watts in a sub 100w chip arena. Better yet the 12 core seems to be made up of 2x6 cores that are definitely ugly step sisters in server land.

IO dies with bad links you can bin out to the 48 or less segments.

Threadripper will certainly fit into a soon to be overlapping area.

Did any of you watch the roll out at the Palace of Fine arts? I think that explains where all the dies are going. Google, HP, Cray, etc even the Dell guy was reluctantly there.

 

[R0H1T]

Look that you were either not here at the time or eventually lost in Intel s marketing kool aid...

Kool aid, you think the IDF cares about that? According to them Intel can do no wrong & if AMD disappeared, this was way back before Zen, then Intel would automagically reduce the prices of their CPUs because um they'd suddenly have a change of heart

 

[dacostafilipe]

He has a point though about the output.

Well, if you can't produce, you can't sell ... yes. But giving him credit for a speculation is going too far if you ask me.

 

[Atari2600]

The 56 core intel CPUs should perform significantly faster if you're running a lot of AVX512 code (which is very, very niche right now). There are a couple of other very niche applications where the intel will win. Usually those very very niche markets are used to paying absurd prices for products catered to them so the chip will find a home or two, but really it's more of a PR move than anything.

Initial thoughts are - yes, absolutely.

Then I started wondering what the thermal throttling will be for a Tommy Cooper Lake running AVX512 code.

The unglued alternative has around a 250W TDP (which need everyone be reminded, is Intel's - non AVX - base freq based definition).
So clocks are gonna be down 20% before you start taking a hit for the communications between modules.



Also:

Fig1

Fig2

Note the scaling from 7402P to 7742 (1P) in GROMACS.

The 7742 has ~2.15 the resource of 7402P (based on clock normalised cores), and runs approx 1.9 times faster. So scaling is around 89% - which would be where I would expect HPC workload scaling to be.



Thus a 2P 7742 would use approximately the same power as a 1P Tommy Cooper Lake, and run the solution in the same speed - if not quicker once Zen2 optimisations go in, and it would be far quicker in the non AVX512 code.

Zen2 would not deliver the same density per blade - but crucially - depending on how your rack is cooled - it might deliver the same or better density per rack... and people don't install blades into their building, they install racks.

Tommy Cooper Lake may not be good enough to be even an adequate response in the one area Intel have stacked a lot of transistor budget on.

 

[Zucker2k]

These Rome chips have raised the performance bar so incredibly high chip architects at Intel are going to be losing a healthy amount of sleep for a couple of years to come.

Those rumors of an extra terrestrial residing in the basement of Intel labs better be true,

 

[Atari2600]

When Dell started selling AMD after Conroe was out and AMD couldn't supply their retail channels, what hope would they have had of supplying oems, before Conroe came out?

If AMD were able to supply Opteron into the market at 10 times the ASP of their Athlon line in 2003, do you not think that would have a material effect on the investment into - and thus the capacity of - their FABs?

Stop being silly and perpetuating the myth that Intel's illegal crap had no effect on AMD.

 

[Yotsugi]

Tommy Cooper Lake may not be good enough to be even an adequate response in the one area Intel have stacked a lot of transistor budget on.

It's okay.
The worst part is hands down Cooper being not that far away from Milan.

 

[BigDaveX]

Stop being silly and perpetuating the myth that Intel's illegal crap had no effect on AMD.

I don't think anyone's claiming that, just disputing the idea that Intel's illegal tactics were the only reason their marketshare didn't utterly collapse and allow AMD to take over the market in the mid-2000s.

Yes, AMD could have undoubtedly made more money in 2000-2001 and 2003-2006 than they did, but enough to expand their production capacity to anything approaching Intel's? I don't think they'd have have enough time before Conroe retook the performance lead.

 

[Abwx]

It's okay.
The worst part is hands down Cooper being not that far away from Milan.

At 2x the TDP they are far away efficency wise, that may be marginal with a single server but what about 200-500 when room climatisation has to provide 200W for the 200W excess TDP?

That s 400W difference at full tilt for 200W more TDP per CPU, not only those Platinium are requiring more total power but the climatisation itself has to be 2x bigger or you can only house half the CPU count.

 

[Yotsugi]

At 2x the TDP they are far away efficency wise,

CPX-SP are mainstream parts so expect reasonable power and pathetic clocks.

 

[Abwx]

CPX-SP are mainstream parts so expect reasonable power and pathetic clocks.

Either absolute perf competitive at the expense of efficency or efficency at the expense of performance, either way you split it it doesnt work, besides not only Epyc 2 can be also downclocked but it will gain more perf/Watt for an equivalent frequency downscaling..

How this could be..?.

Because Intel ultra refined 14nm has very good characteristics up to very high frequencies, this mean, and it can be checked, that the power/frequency curve at relatively low frequencies is close to a polynomial with 2 as exponent, on the other hand TSMC process has an exponent that is in the 2.6-2.8 range, say 2.7, at those rates this mean that FI 0.8x the frequency will translate to 0.8^2 = 0.64x the power with Intel CPU while it will be 0.8^2.7 = 0.55x the power for AMD s Epyc.

It was noticed that Zen 2 power increase more rapidly than CFL with frequency, on the other hand it will decrease at the same rate when frequency is reduced...

 

[Yotsugi]

Either absolute perf competitive at the expense of efficency or efficency at the expense of performance

These are mainstream parts.
The key word here is mainstream.
I.e. no unmanageable power or BGA substrates.

 

[Hitman928]

Initial thoughts are - yes, absolutely.

Then I started wondering what the thermal throttling will be for a Tommy Cooper Lake running AVX512 code.

The unglued alternative has around a 250W TDP (which need everyone be reminded, is Intel's - non AVX - base freq based definition).
So clocks are gonna be down 20% before you start taking a hit for the communications between modules.



Also:

Fig1

Fig2

Note the scaling from 7402P to 7742 (1P) in GROMACS.

The 7742 has ~2.15 the resource of 7402P (based on clock normalised cores), and runs approx 1.9 times faster. So scaling is around 89% - which would be where I would expect HPC workload scaling to be.



Thus a 2P 7742 would use approximately the same power as a 1P Tommy Cooper Lake, and run the solution in the same speed - if not quicker once Zen2 optimisations go in, and it would be far quicker in the non AVX512 code.

Zen2 would not deliver the same density per blade - but crucially - depending on how your rack is cooled - it might deliver the same or better density per rack... and people don't install blades into their building, they install racks.

Tommy Cooper Lake may not be good enough to be even an adequate response in the one area Intel have stacked a lot of transistor budget on.

I'm confused, you're basically saying if you throw 2x the amount of Rome CPUs against Intel CPUs, their performance will be roughly on par in AVX512 heavy code. Didn't we already come to this conclusion? What we have in the above benchmarks is 2x64 =128 core Rome vs 2x28 =56 core Cascade Lake and they show roughly equal performance in AVX512 code but better perf/w for Rome. You're saying if we plug in Cooper Lake instead with 56 cores but restrict it to one 1p, the 2p Rome system could still match it in performance in AVX512 code. You're doing the same 128 core vs 56 core comparison though clock frequencies might adjust a bit.

All I said is if you use Cooper Lake instead of Cascade Lake ( implying you don't impose an artificial 1p limit for Cooper Lake), then the Rome system won't be able to keep up in AVX512 performance but that perf/w would look better for the Rome system and that there's probably a couple of other very niche corner cases where the intel system will perform better as well. These systems will most likely be watercooled which isn't a foreign thing to compute cluster builders anymore. The people who need the best AVX512 performance are such a small group, it won't even be a blip on Intel's revenue radar, but for them, most likely they'll understand the compromises and won't care, they'll just want the best AVX512 performance available. Again, this is a very very small niche of customers, but they do exist. I think at least 95+% of customers would be better served with Rome if doing a new build out.

 

[nicalandia]

So far ALL of the benches done on Epyc vs Xeons have been without the SWAPGSAttack mitigations right? With new Side Channel attacks the Rome chips lead is about to get larger

 

[Jimzz]

So far ALL of the benches done on Epyc vs Xeons have been without the SWAPGSAttack mitigations right? With new Side Channel attacks the Rome chips lead is about to get larger

I thought the new issues would only take maybe 1-2% off the intel CPUs? Don;t get me wrong that still good for AMD but not end of world for intel difference.

 

[nicalandia]

I thought the new issues would only take maybe 1-2% off the intel CPUs? Don;t get me wrong that still good for AMD but not end of world for intel difference.

I am sure the 1-2% performance recession is nothing to be proud of, the Phoronix test was done on a modular 8 core desktop chip, how will that effect a MCM Xeon? Perhaps on the 5% range?

From phoronix

"At least as it stands now it seems to be largely a 1% or less performance hit (with select exceptions at ~5%), so not nearly as bad as the other Spectre/Meltdown/Foreshadow/Zombieload mitigations, but this is on top of all that we've seen since January 2018"

https://www.phoronix.com/scan.php?page=article&item=swapgs-spectre-impact&num=1

 

[Atari2600]

I'm confused, you're basically saying if you throw 2x the amount of Rome CPUs against Intel CPUs, their performance will be roughly on par in AVX512 heavy code. Didn't we already come to this conclusion?

Yeah, I kinda went around in circles a bit!

Basically, if you do the factoring on Tommy Cooper Lake clocks, you are looking at a ~1.6x advantage to Intel.

I dunno how much Zen2 optimisations will be worth, probably just a few percent off that. So figure a ~1.55x advantage to Intel on AVX512 heavy tasks.

****IF THE CLUSTER CAN COOL THE CHIPS****

These systems will most likely be watercooled which isn't a foreign thing to compute cluster builders anymore.

Is the cooling sink external to the room, or is this a conventional cooling block with radiator somewhere near the rack?

When it comes to big compute, you cannot double the amount of heat released into the room without consequences for how you are cooling the room. Which may then need a significant upgrade in the building's cooling capacity.

 

[Hitman928]

Yeah, I kinda went around in circles a bit!

Basically, if you do the factoring on Tommy Cooper Lake clocks, you are looking at a ~1.6x advantage to Intel.

I dunno how much Zen2 optimisations will be worth, probably just a few percent off that. So figure a ~1.55x advantage to Intel on AVX512 heavy tasks.

****IF THE CLUSTER CAN COOL THE CHIPS****





Is the cooling sink external to the room, or is this a conventional cooling block with radiator somewhere near the rack?

When it comes to big compute, you cannot double the amount of heat released into the room without consequences for how you are cooling the room. Which may then need a significant upgrade in the building's cooling capacity.

Typically radiator will sit near the rack though I've seen some more exotic mock-ups (not sure if ever actually implemented). Again, these systems will be built by people who understand the needs for cooling and will already have the HVAC needed to cool the room or will be prepared to upgrade. These aren't going to be purchased by a normal (even large enterprise) customer to put into their existing racks as a drop in replacement unless their IT/purchaser is an idiot and doesn't understand what they're buying. Think of it more as a semi-custom type of situation where the customer is prepared to do work on their part to make it happen.

 

[Atari2600]

These aren't going to be purchased by a normal (even large enterprise) customer to put into their existing racks as a drop in replacement unless their IT/purchaser is an idiot and doesn't understand what they're buying. Think of it more as a semi-custom type of situation where the customer is prepared to do work on their part to make it happen.

Tommy Cooper Lake arrives sometime in 2020 (having slipped from 2019).

Zen3 (Milan) is slated to arrive mid 2020 in much the same timeframe as Zen2.

If AMD upped their core counts further on Milan, then you might go to a lot of trouble rebuilding your infrastructure installing the equivalent of a 7740X.

If Ice-Lake looks strong, then they might get sales off the back of LGA4189 offering a quick upgrade - but if not its more a "don't forget about us" product.